diff --git "a/checkpoint-10000/trainer_state.json" "b/checkpoint-10000/trainer_state.json"
new file mode 100644--- /dev/null
+++ "b/checkpoint-10000/trainer_state.json"
@@ -0,0 +1,88488 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.7944073721004131,
+  "global_step": 10000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0,
+      "learning_rate": 7.936507936507936e-06,
+      "loss": 1.0932,
+      "theoretical_loss": 14.920781838632275,
+      "tokens_seen": 262144
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 1.5873015873015872e-05,
+      "loss": 1.0937,
+      "theoretical_loss": 12.718594708127029,
+      "tokens_seen": 524288
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 2.380952380952381e-05,
+      "loss": 1.0519,
+      "theoretical_loss": 11.615184291350435,
+      "tokens_seen": 786432
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 3.1746031746031745e-05,
+      "loss": 1.0058,
+      "theoretical_loss": 10.904893169100655,
+      "tokens_seen": 1048576
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 3.968253968253968e-05,
+      "loss": 0.976,
+      "theoretical_loss": 10.392029026407034,
+      "tokens_seen": 1310720
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 4.761904761904762e-05,
+      "loss": 0.9521,
+      "theoretical_loss": 9.996134261483984,
+      "tokens_seen": 1572864
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 5.555555555555555e-05,
+      "loss": 0.9364,
+      "theoretical_loss": 9.67682184172525,
+      "tokens_seen": 1835008
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 6.349206349206349e-05,
+      "loss": 0.9245,
+      "theoretical_loss": 9.41114487355416,
+      "tokens_seen": 2097152
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 7.142857142857142e-05,
+      "loss": 0.9103,
+      "theoretical_loss": 9.184905895151996,
+      "tokens_seen": 2359296
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 7.936507936507937e-05,
+      "loss": 0.8933,
+      "theoretical_loss": 8.988754572553061,
+      "tokens_seen": 2621440
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 8.73015873015873e-05,
+      "loss": 0.8721,
+      "theoretical_loss": 8.816230875422118,
+      "tokens_seen": 2883584
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 9.523809523809524e-05,
+      "loss": 0.8552,
+      "theoretical_loss": 8.66269920037918,
+      "tokens_seen": 3145728
+    },
+    {
+      "epoch": 0.0,
+      "objective/train/advantage_avg": 0.43383753299713135,
+      "objective/train/docs_used": 8371,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 8.528481483459473,
+      "objective/train/original_loss": 8.528482437133789,
+      "objective/train/theoretical_loss": 8.591947747254773,
+      "objective/train/tokens_used": 23736800,
+      "objective/train/value_avg": -0.4384765625,
+      "objective/train/value_loss": 0.1883670836687088,
+      "objective/train/value_max": -0.4375,
+      "objective/train/value_min": -0.441650390625,
+      "objective/train/value_reward_corr": -0.01855261992160691,
+      "objective/train/value_std": 0.0006289482116699219,
+      "objective/train/weight_avg": 1.5432828664779663,
+      "objective/train/weighted_lm_loss": 13.15995979309082,
+      "objective/train/weights_max": 1.555271863937378,
+      "objective/train/weights_min": 1.1650478839874268,
+      "theoretical_loss": 8.591947747254773,
+      "tokens_seen": 3276800
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00010317460317460317,
+      "loss": 0.8283,
+      "theoretical_loss": 8.524729102289708,
+      "tokens_seen": 3407872
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0001111111111111111,
+      "loss": 0.8015,
+      "theoretical_loss": 8.399716359763914,
+      "tokens_seen": 3670016
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00011904761904761905,
+      "loss": 0.7709,
+      "theoretical_loss": 8.285641004895568,
+      "tokens_seen": 3932160
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00012698412698412698,
+      "loss": 0.7402,
+      "theoretical_loss": 8.180907195283321,
+      "tokens_seen": 4194304
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0001349206349206349,
+      "loss": 0.7102,
+      "theoretical_loss": 8.084233979345122,
+      "tokens_seen": 4456448
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00014285714285714284,
+      "loss": 0.6751,
+      "theoretical_loss": 7.9945788049155055,
+      "tokens_seen": 4718592
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0001507936507936508,
+      "loss": 0.6503,
+      "theoretical_loss": 7.911082722632908,
+      "tokens_seen": 4980736
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00015873015873015873,
+      "loss": 0.6183,
+      "theoretical_loss": 7.83303033759787,
+      "tokens_seen": 5242880
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00016666666666666666,
+      "loss": 0.5919,
+      "theoretical_loss": 7.759820016443023,
+      "tokens_seen": 5505024
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0001746031746031746,
+      "loss": 0.5614,
+      "theoretical_loss": 7.690941370375033,
+      "tokens_seen": 5767168
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00018253968253968252,
+      "loss": 0.5314,
+      "theoretical_loss": 7.6259579939239845,
+      "tokens_seen": 6029312
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00019047619047619048,
+      "loss": 0.5048,
+      "theoretical_loss": 7.564494061943624,
+      "tokens_seen": 6291456
+    },
+    {
+      "epoch": 0.0,
+      "objective/train/advantage_avg": 0.19523106515407562,
+      "objective/train/docs_used": 9704,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 7.315228462219238,
+      "objective/train/original_loss": 7.315227508544922,
+      "objective/train/theoretical_loss": 7.5062238006917354,
+      "objective/train/tokens_used": 27013600,
+      "objective/train/value_avg": -0.2044677734375,
+      "objective/train/value_loss": 0.03861678019165993,
+      "objective/train/value_max": -0.1988525390625,
+      "objective/train/value_min": -0.251220703125,
+      "objective/train/value_reward_corr": -0.023237293515685216,
+      "objective/train/value_std": 0.01251983642578125,
+      "objective/train/weight_avg": 1.2158902883529663,
+      "objective/train/weighted_lm_loss": 8.87794303894043,
+      "objective/train/weights_max": 1.285593867301941,
+      "objective/train/weights_min": 0.46789029240608215,
+      "theoretical_loss": 7.5062238006917354,
+      "tokens_seen": 6553600
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0001984126984126984,
+      "loss": 0.4798,
+      "theoretical_loss": 7.5062238006917354,
+      "tokens_seen": 6553600
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00020634920634920634,
+      "loss": 0.4564,
+      "theoretical_loss": 7.45086312850561,
+      "tokens_seen": 6815744
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00021428571428571427,
+      "loss": 0.4397,
+      "theoretical_loss": 7.398162954262078,
+      "tokens_seen": 7077888
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0002222222222222222,
+      "loss": 0.4264,
+      "theoretical_loss": 7.347903756717382,
+      "tokens_seen": 7340032
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00023015873015873016,
+      "loss": 0.401,
+      "theoretical_loss": 7.299891163694537,
+      "tokens_seen": 7602176
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0002380952380952381,
+      "loss": 0.3912,
+      "theoretical_loss": 7.253952319156202,
+      "tokens_seen": 7864320
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.000246031746031746,
+      "loss": 0.3754,
+      "theoretical_loss": 7.2099328765932205,
+      "tokens_seen": 8126464
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00025396825396825396,
+      "loss": 0.366,
+      "theoretical_loss": 7.167694494355343,
+      "tokens_seen": 8388608
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0002619047619047619,
+      "loss": 0.3575,
+      "theoretical_loss": 7.127112736305475,
+      "tokens_seen": 8650752
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0002698412698412698,
+      "loss": 0.3477,
+      "theoretical_loss": 7.0880753020982725,
+      "tokens_seen": 8912896
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0002777777777777778,
+      "loss": 0.3462,
+      "theoretical_loss": 7.050480527300383,
+      "tokens_seen": 9175040
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0002857142857142857,
+      "loss": 0.3397,
+      "theoretical_loss": 7.014236105786485,
+      "tokens_seen": 9437184
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0002936507936507937,
+      "loss": 0.3387,
+      "theoretical_loss": 6.979257996300014,
+      "tokens_seen": 9699328
+    },
+    {
+      "epoch": 0.0,
+      "objective/train/advantage_avg": 0.035979945212602615,
+      "objective/train/docs_used": 11015,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 6.488903999328613,
+      "objective/train/original_loss": 6.488903522491455,
+      "objective/train/theoretical_loss": 6.962219571538605,
+      "objective/train/tokens_used": 30290400,
+      "objective/train/value_avg": -0.044891357421875,
+      "objective/train/value_loss": 0.002101501217111945,
+      "objective/train/value_max": -0.038330078125,
+      "objective/train/value_min": -0.092529296875,
+      "objective/train/value_reward_corr": -0.002001422374817204,
+      "objective/train/value_std": 0.0086822509765625,
+      "objective/train/weight_avg": 1.0370042324066162,
+      "objective/train/weighted_lm_loss": 6.7182722091674805,
+      "objective/train/weights_max": 1.0969452857971191,
+      "objective/train/weights_min": 0.3830130398273468,
+      "theoretical_loss": 6.962219571538605,
+      "tokens_seen": 9830400
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0003015873015873016,
+      "loss": 0.3357,
+      "theoretical_loss": 6.945469482441503,
+      "tokens_seen": 9961472
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00030952380952380956,
+      "loss": 0.332,
+      "theoretical_loss": 6.912800361140576,
+      "tokens_seen": 10223616
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00031746031746031746,
+      "loss": 0.3301,
+      "theoretical_loss": 6.881186239250335,
+      "tokens_seen": 10485760
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0003253968253968254,
+      "loss": 0.3283,
+      "theoretical_loss": 6.8505679215514235,
+      "tokens_seen": 10747904
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0003333333333333333,
+      "loss": 0.3248,
+      "theoretical_loss": 6.8208908763759295,
+      "tokens_seen": 11010048
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0003412698412698413,
+      "loss": 0.3232,
+      "theoretical_loss": 6.79210476741633,
+      "tokens_seen": 11272192
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0003492063492063492,
+      "loss": 0.3221,
+      "theoretical_loss": 6.76416304219278,
+      "tokens_seen": 11534336
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00035714285714285714,
+      "loss": 0.3173,
+      "theoretical_loss": 6.737022569206117,
+      "tokens_seen": 11796480
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00036507936507936505,
+      "loss": 0.3175,
+      "theoretical_loss": 6.710643317075979,
+      "tokens_seen": 12058624
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.000373015873015873,
+      "loss": 0.3144,
+      "theoretical_loss": 6.684988070009584,
+      "tokens_seen": 12320768
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00038095238095238096,
+      "loss": 0.3176,
+      "theoretical_loss": 6.660022174811009,
+      "tokens_seen": 12582912
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0003888888888888889,
+      "loss": 0.3123,
+      "theoretical_loss": 6.6357133153579175,
+      "tokens_seen": 12845056
+    },
+    {
+      "epoch": 0.0,
+      "objective/train/advantage_avg": 0.02608620934188366,
+      "objective/train/docs_used": 12170,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 6.028903007507324,
+      "objective/train/original_loss": 6.028903961181641,
+      "objective/train/theoretical_loss": 6.612031311070119,
+      "objective/train/tokens_used": 33567200,
+      "objective/train/value_avg": -0.03485107421875,
+      "objective/train/value_loss": 0.0012108575319871306,
+      "objective/train/value_max": -0.0136871337890625,
+      "objective/train/value_min": -0.05511474609375,
+      "objective/train/value_reward_corr": -0.03430143501280518,
+      "objective/train/value_std": 0.00576019287109375,
+      "objective/train/weight_avg": 1.0266841650009155,
+      "objective/train/weighted_lm_loss": 6.190183162689209,
+      "objective/train/weights_max": 1.05666184425354,
+      "objective/train/weights_min": 0.3770461976528168,
+      "theoretical_loss": 6.612031311070119,
+      "tokens_seen": 13107200
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0003968253968253968,
+      "loss": 0.3097,
+      "theoretical_loss": 6.612031311070119,
+      "tokens_seen": 13107200
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0004047619047619048,
+      "loss": 0.3104,
+      "theoretical_loss": 6.588947936394168,
+      "tokens_seen": 13369344
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0004126984126984127,
+      "loss": 0.3022,
+      "theoretical_loss": 6.566436758747731,
+      "tokens_seen": 13631488
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00042063492063492065,
+      "loss": 0.3027,
+      "theoretical_loss": 6.544472992721121,
+      "tokens_seen": 13893632
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00042857142857142855,
+      "loss": 0.3044,
+      "theoretical_loss": 6.523033368632323,
+      "tokens_seen": 14155776
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0004365079365079365,
+      "loss": 0.3036,
+      "theoretical_loss": 6.502096013785574,
+      "tokens_seen": 14417920
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0004444444444444444,
+      "loss": 0.3003,
+      "theoretical_loss": 6.481640344999435,
+      "tokens_seen": 14680064
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00045238095238095237,
+      "loss": 0.3021,
+      "theoretical_loss": 6.461646971154669,
+      "tokens_seen": 14942208
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00046031746031746033,
+      "loss": 0.3006,
+      "theoretical_loss": 6.442097604670096,
+      "tokens_seen": 15204352
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0004682539682539683,
+      "loss": 0.2922,
+      "theoretical_loss": 6.422974980950157,
+      "tokens_seen": 15466496
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0004761904761904762,
+      "loss": 0.2958,
+      "theoretical_loss": 6.404262784964672,
+      "tokens_seen": 15728640
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.00048412698412698415,
+      "loss": 0.2952,
+      "theoretical_loss": 6.3859455842220765,
+      "tokens_seen": 15990784
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.000492063492063492,
+      "loss": 0.2945,
+      "theoretical_loss": 6.368008767484675,
+      "tokens_seen": 16252928
+    },
+    {
+      "epoch": 0.0,
+      "objective/train/advantage_avg": 0.028485985472798347,
+      "objective/train/docs_used": 13382,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 5.559032440185547,
+      "objective/train/original_loss": 5.559032917022705,
+      "objective/train/theoretical_loss": 6.359178647759789,
+      "objective/train/tokens_used": 36844000,
+      "objective/train/value_avg": -0.03570556640625,
+      "objective/train/value_loss": 0.001410536002367735,
+      "objective/train/value_max": -0.01111602783203125,
+      "objective/train/value_min": -0.0867919921875,
+      "objective/train/value_reward_corr": 0.08617836470205797,
+      "objective/train/value_std": 0.01192474365234375,
+      "objective/train/weight_avg": 1.0291692018508911,
+      "objective/train/weighted_lm_loss": 5.7365899085998535,
+      "objective/train/weights_max": 1.0881431102752686,
+      "objective/train/weights_min": 0.378071665763855,
+      "theoretical_loss": 6.359178647759789,
+      "tokens_seen": 16384000
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005,
+      "loss": 0.293,
+      "theoretical_loss": 6.350438488650175,
+      "tokens_seen": 16515072
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005079365079365079,
+      "loss": 0.2876,
+      "theoretical_loss": 6.333221615289645,
+      "tokens_seen": 16777216
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005158730158730159,
+      "loss": 0.2909,
+      "theoretical_loss": 6.316345681389436,
+      "tokens_seen": 17039360
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005238095238095238,
+      "loss": 0.2851,
+      "theoretical_loss": 6.2997988438948465,
+      "tokens_seen": 17301504
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005317460317460317,
+      "loss": 0.2807,
+      "theoretical_loss": 6.283569842697203,
+      "tokens_seen": 17563648
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005396825396825396,
+      "loss": 0.2855,
+      "theoretical_loss": 6.26764796374462,
+      "tokens_seen": 17825792
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005476190476190477,
+      "loss": 0.2822,
+      "theoretical_loss": 6.25202300499066,
+      "tokens_seen": 18087936
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005555555555555556,
+      "loss": 0.2818,
+      "theoretical_loss": 6.236685244924882,
+      "tokens_seen": 18350080
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005634920634920635,
+      "loss": 0.2847,
+      "theoretical_loss": 6.2216254134558024,
+      "tokens_seen": 18612224
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005714285714285714,
+      "loss": 0.2838,
+      "theoretical_loss": 6.206834664939976,
+      "tokens_seen": 18874368
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005793650793650794,
+      "loss": 0.2875,
+      "theoretical_loss": 6.192304553171669,
+      "tokens_seen": 19136512
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005873015873015874,
+      "loss": 0.2852,
+      "theoretical_loss": 6.178027008165916,
+      "tokens_seen": 19398656
+    },
+    {
+      "epoch": 0.01,
+      "objective/train/advantage_avg": 0.01716114766895771,
+      "objective/train/docs_used": 14656,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 5.421322822570801,
+      "objective/train/original_loss": 5.421322822570801,
+      "objective/train/theoretical_loss": 6.163994314584031,
+      "objective/train/tokens_used": 40120800,
+      "objective/train/value_avg": -0.0253753662109375,
+      "objective/train/value_loss": 0.0010442916536703706,
+      "objective/train/value_max": -0.0090179443359375,
+      "objective/train/value_min": -0.04986572265625,
+      "objective/train/value_reward_corr": -0.02662585421402189,
+      "objective/train/value_std": 0.0084381103515625,
+      "objective/train/weight_avg": 1.0176551342010498,
+      "objective/train/weighted_lm_loss": 5.523168087005615,
+      "objective/train/weights_max": 1.051129937171936,
+      "objective/train/weights_min": 0.37920987606048584,
+      "theoretical_loss": 6.163994314584031,
+      "tokens_seen": 19660800
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0005952380952380953,
+      "loss": 0.283,
+      "theoretical_loss": 6.163994314584031,
+      "tokens_seen": 19660800
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006031746031746032,
+      "loss": 0.2769,
+      "theoretical_loss": 6.150199091665225,
+      "tokens_seen": 19922944
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006111111111111112,
+      "loss": 0.2773,
+      "theoretical_loss": 6.136634274540901,
+      "tokens_seen": 20185088
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006190476190476191,
+      "loss": 0.2813,
+      "theoretical_loss": 6.123293096819758,
+      "tokens_seen": 20447232
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000626984126984127,
+      "loss": 0.2788,
+      "theoretical_loss": 6.1101690743422505,
+      "tokens_seen": 20709376
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006349206349206349,
+      "loss": 0.2779,
+      "theoretical_loss": 6.097255990012153,
+      "tokens_seen": 20971520
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006428571428571429,
+      "loss": 0.2727,
+      "theoretical_loss": 6.084547879621354,
+      "tokens_seen": 21233664
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006507936507936508,
+      "loss": 0.2732,
+      "theoretical_loss": 6.072039018591484,
+      "tokens_seen": 21495808
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006587301587301587,
+      "loss": 0.2743,
+      "theoretical_loss": 6.059723909562683,
+      "tokens_seen": 21757952
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006666666666666666,
+      "loss": 0.2703,
+      "theoretical_loss": 6.047597270765904,
+      "tokens_seen": 22020096
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006746031746031747,
+      "loss": 0.2718,
+      "theoretical_loss": 6.035654025120612,
+      "tokens_seen": 22282240
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006825396825396826,
+      "loss": 0.269,
+      "theoretical_loss": 6.023889290004692,
+      "tokens_seen": 22544384
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006904761904761905,
+      "loss": 0.269,
+      "theoretical_loss": 6.012298367647816,
+      "tokens_seen": 22806528
+    },
+    {
+      "epoch": 0.01,
+      "objective/train/advantage_avg": 0.018673593178391457,
+      "objective/train/docs_used": 15907,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 5.2787933349609375,
+      "objective/train/original_loss": 5.2787933349609375,
+      "objective/train/theoretical_loss": 6.006566666513313,
+      "objective/train/tokens_used": 43397600,
+      "objective/train/value_avg": -0.0252227783203125,
+      "objective/train/value_loss": 0.0007268089102581143,
+      "objective/train/value_max": -0.00916290283203125,
+      "objective/train/value_min": -0.059967041015625,
+      "objective/train/value_reward_corr": -0.009804935861391459,
+      "objective/train/value_std": 0.0091552734375,
+      "objective/train/weight_avg": 1.0190335512161255,
+      "objective/train/weighted_lm_loss": 5.3884196281433105,
+      "objective/train/weights_max": 1.0590507984161377,
+      "objective/train/weights_min": 0.37753555178642273,
+      "theoretical_loss": 6.006566666513313,
+      "tokens_seen": 22937600
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0006984126984126984,
+      "loss": 0.266,
+      "theoretical_loss": 6.000876736103618,
+      "tokens_seen": 23068672
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0007063492063492064,
+      "loss": 0.2679,
+      "theoretical_loss": 5.989620040759641,
+      "tokens_seen": 23330816
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0007142857142857143,
+      "loss": 0.2638,
+      "theoretical_loss": 5.978524086347409,
+      "tokens_seen": 23592960
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0007222222222222222,
+      "loss": 0.2676,
+      "theoretical_loss": 5.967584829417934,
+      "tokens_seen": 23855104
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0007301587301587301,
+      "loss": 0.2632,
+      "theoretical_loss": 5.956798371250791,
+      "tokens_seen": 24117248
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0007380952380952381,
+      "loss": 0.2593,
+      "theoretical_loss": 5.9461609511673625,
+      "tokens_seen": 24379392
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000746031746031746,
+      "loss": 0.2628,
+      "theoretical_loss": 5.935668940221127,
+      "tokens_seen": 24641536
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000753968253968254,
+      "loss": 0.261,
+      "theoretical_loss": 5.92531883523999,
+      "tokens_seen": 24903680
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0007619047619047619,
+      "loss": 0.2601,
+      "theoretical_loss": 5.915107253197538,
+      "tokens_seen": 25165824
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0007698412698412699,
+      "loss": 0.2627,
+      "theoretical_loss": 5.905030925891829,
+      "tokens_seen": 25427968
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0007777777777777778,
+      "loss": 0.2632,
+      "theoretical_loss": 5.895086694911951,
+      "tokens_seen": 25690112
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0007857142857142857,
+      "loss": 0.261,
+      "theoretical_loss": 5.88527150687402,
+      "tokens_seen": 25952256
+    },
+    {
+      "epoch": 0.01,
+      "objective/train/advantage_avg": 0.014981591142714024,
+      "objective/train/docs_used": 17033,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 5.067455291748047,
+      "objective/train/original_loss": 5.067455291748047,
+      "objective/train/theoretical_loss": 5.8755824089096285,
+      "objective/train/tokens_used": 46674400,
+      "objective/train/value_avg": -0.0223236083984375,
+      "objective/train/value_loss": 0.001089372206479311,
+      "objective/train/value_max": -0.007785797119140625,
+      "objective/train/value_min": -0.08538818359375,
+      "objective/train/value_reward_corr": 0.11159853979639632,
+      "objective/train/value_std": 0.008026123046875,
+      "objective/train/weight_avg": 1.0154664516448975,
+      "objective/train/weighted_lm_loss": 5.153756618499756,
+      "objective/train/weights_max": 1.0678131580352783,
+      "objective/train/weights_min": 0.37597760558128357,
+      "theoretical_loss": 5.8755824089096285,
+      "tokens_seen": 26214400
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0007936507936507937,
+      "loss": 0.2609,
+      "theoretical_loss": 5.8755824089096285,
+      "tokens_seen": 26214400
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008015873015873017,
+      "loss": 0.2559,
+      "theoretical_loss": 5.866016544391016,
+      "tokens_seen": 26476544
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008095238095238096,
+      "loss": 0.2535,
+      "theoretical_loss": 5.856571148878293,
+      "tokens_seen": 26738688
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008174603174603175,
+      "loss": 0.2534,
+      "theoretical_loss": 5.847243546275179,
+      "tokens_seen": 27000832
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008253968253968254,
+      "loss": 0.2527,
+      "theoretical_loss": 5.838031145180573,
+      "tokens_seen": 27262976
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008333333333333334,
+      "loss": 0.2561,
+      "theoretical_loss": 5.82893143542425,
+      "tokens_seen": 27525120
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008412698412698413,
+      "loss": 0.2535,
+      "theoretical_loss": 5.81994198477569,
+      "tokens_seen": 27787264
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008492063492063492,
+      "loss": 0.2519,
+      "theoretical_loss": 5.811060435815881,
+      "tokens_seen": 28049408
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008571428571428571,
+      "loss": 0.2538,
+      "theoretical_loss": 5.802284502962563,
+      "tokens_seen": 28311552
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008650793650793651,
+      "loss": 0.2515,
+      "theoretical_loss": 5.793611969640068,
+      "tokens_seen": 28573696
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000873015873015873,
+      "loss": 0.2494,
+      "theoretical_loss": 5.785040685585437,
+      "tokens_seen": 28835840
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008809523809523809,
+      "loss": 0.2496,
+      "theoretical_loss": 5.7765685642831155,
+      "tokens_seen": 29097984
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008888888888888888,
+      "loss": 0.2526,
+      "theoretical_loss": 5.768193580520972,
+      "tokens_seen": 29360128
+    },
+    {
+      "epoch": 0.01,
+      "objective/train/advantage_avg": 0.015116632916033268,
+      "objective/train/docs_used": 18174,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 5.077436923980713,
+      "objective/train/original_loss": 5.0774359703063965,
+      "objective/train/theoretical_loss": 5.7640418985258295,
+      "objective/train/tokens_used": 49951200,
+      "objective/train/value_avg": -0.0243377685546875,
+      "objective/train/value_loss": 0.0008194705005735159,
+      "objective/train/value_max": -0.006877899169921875,
+      "objective/train/value_min": -0.139404296875,
+      "objective/train/value_reward_corr": 0.07501882751184434,
+      "objective/train/value_std": 0.01377105712890625,
+      "objective/train/weight_avg": 1.0155125856399536,
+      "objective/train/weighted_lm_loss": 5.159419536590576,
+      "objective/train/weights_max": 1.1385329961776733,
+      "objective/train/weights_min": 0.3745060861110687,
+      "theoretical_loss": 5.7640418985258295,
+      "tokens_seen": 29491200
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0008968253968253968,
+      "loss": 0.249,
+      "theoretical_loss": 5.759913768060882,
+      "tokens_seen": 29622272
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009047619047619047,
+      "loss": 0.2492,
+      "theoretical_loss": 5.7517272174175496,
+      "tokens_seen": 29884416
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009126984126984126,
+      "loss": 0.2506,
+      "theoretical_loss": 5.743632073739626,
+      "tokens_seen": 30146560
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009206349206349207,
+      "loss": 0.2489,
+      "theoretical_loss": 5.735626534787584,
+      "tokens_seen": 30408704
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009285714285714287,
+      "loss": 0.2431,
+      "theoretical_loss": 5.727708849003127,
+      "tokens_seen": 30670848
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009365079365079366,
+      "loss": 0.2469,
+      "theoretical_loss": 5.719877313665254,
+      "tokens_seen": 30932992
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009444444444444445,
+      "loss": 0.2464,
+      "theoretical_loss": 5.712130273128388,
+      "tokens_seen": 31195136
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009523809523809524,
+      "loss": 0.2492,
+      "theoretical_loss": 5.704466117138258,
+      "tokens_seen": 31457280
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009603174603174604,
+      "loss": 0.2453,
+      "theoretical_loss": 5.696883279221504,
+      "tokens_seen": 31719424
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009682539682539683,
+      "loss": 0.243,
+      "theoretical_loss": 5.689380235145171,
+      "tokens_seen": 31981568
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009761904761904762,
+      "loss": 0.2464,
+      "theoretical_loss": 5.6819555014425305,
+      "tokens_seen": 32243712
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000984126984126984,
+      "loss": 0.2427,
+      "theoretical_loss": 5.674607634001871,
+      "tokens_seen": 32505856
+    },
+    {
+      "epoch": 0.01,
+      "objective/train/advantage_avg": 0.014621232636272907,
+      "objective/train/docs_used": 19239,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 4.8382439613342285,
+      "objective/train/original_loss": 4.838243007659912,
+      "objective/train/theoretical_loss": 5.667335226715059,
+      "objective/train/tokens_used": 53228000,
+      "objective/train/value_avg": -0.02044677734375,
+      "objective/train/value_loss": 0.0005950998747721314,
+      "objective/train/value_max": -0.006191253662109375,
+      "objective/train/value_min": -0.11181640625,
+      "objective/train/value_reward_corr": 0.10875102217088843,
+      "objective/train/value_std": 0.00927734375,
+      "objective/train/weight_avg": 1.0149059295654297,
+      "objective/train/weighted_lm_loss": 4.914798736572266,
+      "objective/train/weights_max": 1.1008343696594238,
+      "objective/train/weights_min": 0.37406063079833984,
+      "theoretical_loss": 5.667335226715059,
+      "tokens_seen": 32768000
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000992063492063492,
+      "loss": 0.249,
+      "theoretical_loss": 5.667335226715059,
+      "tokens_seen": 32768000
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.001,
+      "loss": 0.2444,
+      "theoretical_loss": 5.6601369101828904,
+      "tokens_seen": 33030144
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009999197560584176,
+      "loss": 0.244,
+      "theoretical_loss": 5.6530113504744435,
+      "tokens_seen": 33292288
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009998395121168352,
+      "loss": 0.2429,
+      "theoretical_loss": 5.645957247937725,
+      "tokens_seen": 33554432
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009997592681752529,
+      "loss": 0.2414,
+      "theoretical_loss": 5.638973336059157,
+      "tokens_seen": 33816576
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009996790242336705,
+      "loss": 0.2359,
+      "theoretical_loss": 5.632058380369512,
+      "tokens_seen": 34078720
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009995987802920879,
+      "loss": 0.2366,
+      "theoretical_loss": 5.625211177394046,
+      "tokens_seen": 34340864
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009995185363505057,
+      "loss": 0.2361,
+      "theoretical_loss": 5.618430553644782,
+      "tokens_seen": 34603008
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000999438292408923,
+      "loss": 0.2379,
+      "theoretical_loss": 5.611715364652864,
+      "tokens_seen": 34865152
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009993580484673407,
+      "loss": 0.2324,
+      "theoretical_loss": 5.605064494039176,
+      "tokens_seen": 35127296
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009992778045257583,
+      "loss": 0.2315,
+      "theoretical_loss": 5.598476852621397,
+      "tokens_seen": 35389440
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000999197560584176,
+      "loss": 0.2364,
+      "theoretical_loss": 5.591951377555809,
+      "tokens_seen": 35651584
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009991173166425935,
+      "loss": 0.2376,
+      "theoretical_loss": 5.585487031512276,
+      "tokens_seen": 35913728
+    },
+    {
+      "epoch": 0.01,
+      "objective/train/advantage_avg": 0.008301016874611378,
+      "objective/train/docs_used": 20470,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 4.471249103546143,
+      "objective/train/original_loss": 4.471248626708984,
+      "objective/train/theoretical_loss": 5.582277464488612,
+      "objective/train/tokens_used": 56504800,
+      "objective/train/value_avg": -0.017913818359375,
+      "objective/train/value_loss": 0.00041489681461825967,
+      "objective/train/value_max": -0.00595855712890625,
+      "objective/train/value_min": -0.11065673828125,
+      "objective/train/value_reward_corr": 0.3373516244538544,
+      "objective/train/value_std": 0.0095672607421875,
+      "objective/train/weight_avg": 1.0085070133209229,
+      "objective/train/weighted_lm_loss": 4.517763614654541,
+      "objective/train/weights_max": 1.0755970478057861,
+      "objective/train/weights_min": 0.7941007018089294,
+      "theoretical_loss": 5.582277464488612,
+      "tokens_seen": 36044800
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009990370727010112,
+      "loss": 0.2363,
+      "theoretical_loss": 5.579082801880871,
+      "tokens_seen": 36175872
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009989568287594288,
+      "loss": 0.234,
+      "theoretical_loss": 5.572737700008718,
+      "tokens_seen": 36438016
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009988765848178462,
+      "loss": 0.2357,
+      "theoretical_loss": 5.56645076046569,
+      "tokens_seen": 36700160
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009987963408762638,
+      "loss": 0.2324,
+      "theoretical_loss": 5.5602210403376775,
+      "tokens_seen": 36962304
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009987160969346814,
+      "loss": 0.2362,
+      "theoretical_loss": 5.554047618546193,
+      "tokens_seen": 37224448
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000998635852993099,
+      "loss": 0.2376,
+      "theoretical_loss": 5.547929595193182,
+      "tokens_seen": 37486592
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009985556090515166,
+      "loss": 0.2321,
+      "theoretical_loss": 5.5418660909298945,
+      "tokens_seen": 37748736
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009984753651099342,
+      "loss": 0.2291,
+      "theoretical_loss": 5.535856246348814,
+      "tokens_seen": 38010880
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009983951211683518,
+      "loss": 0.2304,
+      "theoretical_loss": 5.529899221397624,
+      "tokens_seen": 38273024
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009983148772267695,
+      "loss": 0.2279,
+      "theoretical_loss": 5.523994194814273,
+      "tokens_seen": 38535168
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000998234633285187,
+      "loss": 0.229,
+      "theoretical_loss": 5.518140363582252,
+      "tokens_seen": 38797312
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009981543893436047,
+      "loss": 0.2318,
+      "theoretical_loss": 5.512336942405216,
+      "tokens_seen": 39059456
+    },
+    {
+      "epoch": 0.01,
+      "objective/train/advantage_avg": 0.0074028936214745045,
+      "objective/train/docs_used": 21760,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 4.182587146759033,
+      "objective/train/original_loss": 4.182587623596191,
+      "objective/train/theoretical_loss": 5.506583163200142,
+      "objective/train/tokens_used": 59781600,
+      "objective/train/value_avg": -0.0196990966796875,
+      "objective/train/value_loss": 0.0009806023444980383,
+      "objective/train/value_max": -0.00634002685546875,
+      "objective/train/value_min": -0.1546630859375,
+      "objective/train/value_reward_corr": 0.24312695151689936,
+      "objective/train/value_std": 0.0124359130859375,
+      "objective/train/weight_avg": 1.0078480243682861,
+      "objective/train/weighted_lm_loss": 4.2220258712768555,
+      "objective/train/weights_max": 1.1438496112823486,
+      "objective/train/weights_min": 0.3742261826992035,
+      "theoretical_loss": 5.506583163200142,
+      "tokens_seen": 39321600
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000998074145402022,
+      "loss": 0.2256,
+      "theoretical_loss": 5.506583163200142,
+      "tokens_seen": 39321600
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009979939014604397,
+      "loss": 0.2304,
+      "theoretical_loss": 5.5008782746082625,
+      "tokens_seen": 39583744
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009979136575188573,
+      "loss": 0.2272,
+      "theoretical_loss": 5.495221541523011,
+      "tokens_seen": 39845888
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000997833413577275,
+      "loss": 0.2288,
+      "theoretical_loss": 5.489612244634316,
+      "tokens_seen": 40108032
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009977531696356925,
+      "loss": 0.2282,
+      "theoretical_loss": 5.48404967998854,
+      "tokens_seen": 40370176
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009976729256941101,
+      "loss": 0.2278,
+      "theoretical_loss": 5.478533158563456,
+      "tokens_seen": 40632320
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009975926817525277,
+      "loss": 0.2312,
+      "theoretical_loss": 5.473062005857637,
+      "tokens_seen": 40894464
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009975124378109451,
+      "loss": 0.2301,
+      "theoretical_loss": 5.467635561493681,
+      "tokens_seen": 41156608
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000997432193869363,
+      "loss": 0.2279,
+      "theoretical_loss": 5.462253178834744,
+      "tokens_seen": 41418752
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009973519499277804,
+      "loss": 0.2244,
+      "theoretical_loss": 5.456914224613812,
+      "tokens_seen": 41680896
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009972717059861982,
+      "loss": 0.2276,
+      "theoretical_loss": 5.451618078575256,
+      "tokens_seen": 41943040
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009971914620446156,
+      "loss": 0.2234,
+      "theoretical_loss": 5.446364133128155,
+      "tokens_seen": 42205184
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009971112181030332,
+      "loss": 0.2237,
+      "theoretical_loss": 5.44115179301095,
+      "tokens_seen": 42467328
+    },
+    {
+      "epoch": 0.01,
+      "objective/train/advantage_avg": 0.012641006149351597,
+      "objective/train/docs_used": 22938,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 4.515160083770752,
+      "objective/train/original_loss": 4.51516056060791,
+      "objective/train/theoretical_loss": 5.438561042223066,
+      "objective/train/tokens_used": 63058400,
+      "objective/train/value_avg": -0.01983642578125,
+      "objective/train/value_loss": 0.0004896495374850929,
+      "objective/train/value_max": -0.00661468505859375,
+      "objective/train/value_min": -0.0869140625,
+      "objective/train/value_reward_corr": 0.07471170255551907,
+      "objective/train/value_std": 0.0099029541015625,
+      "objective/train/weight_avg": 1.012885570526123,
+      "objective/train/weighted_lm_loss": 4.5774455070495605,
+      "objective/train/weights_max": 1.0841727256774902,
+      "objective/train/weights_min": 0.7922973036766052,
+      "theoretical_loss": 5.438561042223066,
+      "tokens_seen": 42598400
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009970309741614508,
+      "loss": 0.2268,
+      "theoretical_loss": 5.435980474966981,
+      "tokens_seen": 42729472
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009969507302198684,
+      "loss": 0.2231,
+      "theoretical_loss": 5.430849607430501,
+      "tokens_seen": 42991616
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000996870486278286,
+      "loss": 0.2219,
+      "theoretical_loss": 5.425758630222747,
+      "tokens_seen": 43253760
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009967902423367037,
+      "loss": 0.2263,
+      "theoretical_loss": 5.42070699425771,
+      "tokens_seen": 43515904
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009967099983951213,
+      "loss": 0.2268,
+      "theoretical_loss": 5.415694161257225,
+      "tokens_seen": 43778048
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009966297544535387,
+      "loss": 0.2216,
+      "theoretical_loss": 5.410719603475034,
+      "tokens_seen": 44040192
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009965495105119565,
+      "loss": 0.224,
+      "theoretical_loss": 5.405782803429483,
+      "tokens_seen": 44302336
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009964692665703739,
+      "loss": 0.2221,
+      "theoretical_loss": 5.400883253644551,
+      "tokens_seen": 44564480
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009963890226287915,
+      "loss": 0.226,
+      "theoretical_loss": 5.396020456398885,
+      "tokens_seen": 44826624
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009963087786872091,
+      "loss": 0.2213,
+      "theoretical_loss": 5.391193923482547,
+      "tokens_seen": 45088768
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009962285347456267,
+      "loss": 0.2216,
+      "theoretical_loss": 5.386403175961223,
+      "tokens_seen": 45350912
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009961482908040443,
+      "loss": 0.2215,
+      "theoretical_loss": 5.381647743947578,
+      "tokens_seen": 45613056
+    },
+    {
+      "epoch": 0.01,
+      "objective/train/advantage_avg": 0.011346699669957161,
+      "objective/train/docs_used": 24173,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 4.3351922035217285,
+      "objective/train/original_loss": 4.335192680358887,
+      "objective/train/theoretical_loss": 5.37692716637954,
+      "objective/train/tokens_used": 66335200,
+      "objective/train/value_avg": -0.02081298828125,
+      "objective/train/value_loss": 0.0009375278605148196,
+      "objective/train/value_max": -0.006389617919921875,
+      "objective/train/value_min": -0.09466552734375,
+      "objective/train/value_reward_corr": 0.22103111537465234,
+      "objective/train/value_std": 0.01107025146484375,
+      "objective/train/weight_avg": 1.0117790699005127,
+      "objective/train/weighted_lm_loss": 4.3885650634765625,
+      "objective/train/weights_max": 1.0966105461120605,
+      "objective/train/weights_min": 0.37649428844451904,
+      "theoretical_loss": 5.37692716637954,
+      "tokens_seen": 45875200
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000996068046862462,
+      "loss": 0.2257,
+      "theoretical_loss": 5.37692716637954,
+      "tokens_seen": 45875200
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009959878029208796,
+      "loss": 0.2205,
+      "theoretical_loss": 5.372240990805237,
+      "tokens_seen": 46137344
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009959075589792972,
+      "loss": 0.2197,
+      "theoretical_loss": 5.367588773174377,
+      "tokens_seen": 46399488
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009958273150377146,
+      "loss": 0.225,
+      "theoretical_loss": 5.36297007763582,
+      "tokens_seen": 46661632
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009957470710961322,
+      "loss": 0.2232,
+      "theoretical_loss": 5.358384476341126,
+      "tokens_seen": 46923776
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009956668271545498,
+      "loss": 0.2187,
+      "theoretical_loss": 5.353831549253895,
+      "tokens_seen": 47185920
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009955865832129674,
+      "loss": 0.2213,
+      "theoretical_loss": 5.349310883964664,
+      "tokens_seen": 47448064
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.000995506339271385,
+      "loss": 0.2195,
+      "theoretical_loss": 5.344822075511196,
+      "tokens_seen": 47710208
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009954260953298026,
+      "loss": 0.2177,
+      "theoretical_loss": 5.340364726203955,
+      "tokens_seen": 47972352
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009953458513882203,
+      "loss": 0.2156,
+      "theoretical_loss": 5.3359384454566055,
+      "tokens_seen": 48234496
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009952656074466376,
+      "loss": 0.2146,
+      "theoretical_loss": 5.331542849621357,
+      "tokens_seen": 48496640
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009951853635050555,
+      "loss": 0.2172,
+      "theoretical_loss": 5.327177561828993,
+      "tokens_seen": 48758784
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009951051195634729,
+      "loss": 0.219,
+      "theoretical_loss": 5.32284221183342,
+      "tokens_seen": 49020928
+    },
+    {
+      "epoch": 0.01,
+      "objective/train/advantage_avg": 0.0004111192829441279,
+      "objective/train/docs_used": 25394,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 4.30299186706543,
+      "objective/train/original_loss": 4.30299186706543,
+      "objective/train/theoretical_loss": 5.3206856495812715,
+      "objective/train/tokens_used": 69612000,
+      "objective/train/value_avg": -0.018402099609375,
+      "objective/train/value_loss": 0.001803424907848239,
+      "objective/train/value_max": -0.00555419921875,
+      "objective/train/value_min": -0.13427734375,
+      "objective/train/value_reward_corr": 0.352072119180548,
+      "objective/train/value_std": 0.01030731201171875,
+      "objective/train/weight_avg": 1.0012028217315674,
+      "objective/train/weighted_lm_loss": 4.307867527008057,
+      "objective/train/weights_max": 1.0850714445114136,
+      "objective/train/weights_min": 0.37296921014785767,
+      "theoretical_loss": 5.3206856495812715,
+      "tokens_seen": 49152000
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 0.0009950248756218905,
+      "loss": 0.2198,
+      "theoretical_loss": 5.318536435860599,
+      "tokens_seen": 49283072
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000994944631680308,
+      "loss": 0.2143,
+      "theoretical_loss": 5.314259876461705,
+      "tokens_seen": 49545216
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009948643877387257,
+      "loss": 0.2153,
+      "theoretical_loss": 5.310012182370359,
+      "tokens_seen": 49807360
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009947841437971433,
+      "loss": 0.2137,
+      "theoretical_loss": 5.305793008363841,
+      "tokens_seen": 50069504
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000994703899855561,
+      "loss": 0.2132,
+      "theoretical_loss": 5.301602015128104,
+      "tokens_seen": 50331648
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009946236559139785,
+      "loss": 0.2117,
+      "theoretical_loss": 5.297438869126498,
+      "tokens_seen": 50593792
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009945434119723962,
+      "loss": 0.2191,
+      "theoretical_loss": 5.293303242472074,
+      "tokens_seen": 50855936
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009944631680308138,
+      "loss": 0.2133,
+      "theoretical_loss": 5.289194812803347,
+      "tokens_seen": 51118080
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009943829240892312,
+      "loss": 0.2112,
+      "theoretical_loss": 5.285113263163414,
+      "tokens_seen": 51380224
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000994302680147649,
+      "loss": 0.2129,
+      "theoretical_loss": 5.2810582818823235,
+      "tokens_seen": 51642368
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009942224362060664,
+      "loss": 0.215,
+      "theoretical_loss": 5.27702956246258,
+      "tokens_seen": 51904512
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000994142192264484,
+      "loss": 0.2124,
+      "theoretical_loss": 5.273026803467695,
+      "tokens_seen": 52166656
+    },
+    {
+      "epoch": 0.02,
+      "objective/train/advantage_avg": 0.01028979942202568,
+      "objective/train/docs_used": 26541,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 4.072467803955078,
+      "objective/train/original_loss": 4.072467803955078,
+      "objective/train/theoretical_loss": 5.269049708413682,
+      "objective/train/tokens_used": 72888800,
+      "objective/train/value_avg": -0.016021728515625,
+      "objective/train/value_loss": 0.00028009802917949855,
+      "objective/train/value_max": -0.004608154296875,
+      "objective/train/value_min": -0.0889892578125,
+      "objective/train/value_reward_corr": 0.4423854028983263,
+      "objective/train/value_std": 0.008819580078125,
+      "objective/train/weight_avg": 1.0104297399520874,
+      "objective/train/weighted_lm_loss": 4.120135307312012,
+      "objective/train/weights_max": 1.0731053352355957,
+      "objective/train/weights_min": 0.8264122009277344,
+      "theoretical_loss": 5.269049708413682,
+      "tokens_seen": 52428800
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009940619483229016,
+      "loss": 0.2105,
+      "theoretical_loss": 5.269049708413682,
+      "tokens_seen": 52428800
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009939817043813192,
+      "loss": 0.2123,
+      "theoretical_loss": 5.265097985663418,
+      "tokens_seen": 52690944
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009939014604397368,
+      "loss": 0.2115,
+      "theoretical_loss": 5.261171348323755,
+      "tokens_seen": 52953088
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009938212164981545,
+      "loss": 0.2084,
+      "theoretical_loss": 5.257269514145324,
+      "tokens_seen": 53215232
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000993740972556572,
+      "loss": 0.2062,
+      "theoretical_loss": 5.2533922054249365,
+      "tokens_seen": 53477376
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009936607286149895,
+      "loss": 0.2101,
+      "theoretical_loss": 5.2495391489104986,
+      "tokens_seen": 53739520
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009935804846734073,
+      "loss": 0.2142,
+      "theoretical_loss": 5.24571007570837,
+      "tokens_seen": 54001664
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009935002407318247,
+      "loss": 0.2097,
+      "theoretical_loss": 5.2419047211930865,
+      "tokens_seen": 54263808
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009934199967902423,
+      "loss": 0.206,
+      "theoretical_loss": 5.238122824919387,
+      "tokens_seen": 54525952
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.00099333975284866,
+      "loss": 0.2103,
+      "theoretical_loss": 5.234364130536457,
+      "tokens_seen": 54788096
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009932595089070775,
+      "loss": 0.21,
+      "theoretical_loss": 5.230628385704337,
+      "tokens_seen": 55050240
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009931792649654951,
+      "loss": 0.2083,
+      "theoretical_loss": 5.2269153420124255,
+      "tokens_seen": 55312384
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009930990210239128,
+      "loss": 0.2085,
+      "theoretical_loss": 5.223224754900014,
+      "tokens_seen": 55574528
+    },
+    {
+      "epoch": 0.02,
+      "objective/train/advantage_avg": 0.01192543189972639,
+      "objective/train/docs_used": 27758,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.997652769088745,
+      "objective/train/original_loss": 3.997653007507324,
+      "objective/train/theoretical_loss": 5.221387807169494,
+      "objective/train/tokens_used": 76165600,
+      "objective/train/value_avg": -0.0164794921875,
+      "objective/train/value_loss": 0.0003872321976814419,
+      "objective/train/value_max": -0.004451751708984375,
+      "objective/train/value_min": -0.07354736328125,
+      "objective/train/value_reward_corr": 0.07308995924059121,
+      "objective/train/value_std": 0.006809234619140625,
+      "objective/train/weight_avg": 1.0121105909347534,
+      "objective/train/weighted_lm_loss": 4.049195289611816,
+      "objective/train/weights_max": 1.0763195753097534,
+      "objective/train/weights_min": 0.37184974551200867,
+      "theoretical_loss": 5.221387807169494,
+      "tokens_seen": 55705600
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009930187770823304,
+      "loss": 0.2098,
+      "theoretical_loss": 5.219556383578795,
+      "tokens_seen": 55836672
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000992938533140748,
+      "loss": 0.2069,
+      "theoretical_loss": 5.215909990957291,
+      "tokens_seen": 56098816
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009928582891991654,
+      "loss": 0.2075,
+      "theoretical_loss": 5.212285343567135,
+      "tokens_seen": 56360960
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000992778045257583,
+      "loss": 0.2067,
+      "theoretical_loss": 5.208682211491157,
+      "tokens_seen": 56623104
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009926978013160006,
+      "loss": 0.2043,
+      "theoretical_loss": 5.205100368293225,
+      "tokens_seen": 56885248
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009926175573744182,
+      "loss": 0.2042,
+      "theoretical_loss": 5.201539590949796,
+      "tokens_seen": 57147392
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009925373134328358,
+      "loss": 0.206,
+      "theoretical_loss": 5.1979996597831,
+      "tokens_seen": 57409536
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009924570694912534,
+      "loss": 0.2041,
+      "theoretical_loss": 5.1944803583959525,
+      "tokens_seen": 57671680
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000992376825549671,
+      "loss": 0.2063,
+      "theoretical_loss": 5.190981473608112,
+      "tokens_seen": 57933824
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009922965816080887,
+      "loss": 0.2063,
+      "theoretical_loss": 5.18750279539416,
+      "tokens_seen": 58195968
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009922163376665063,
+      "loss": 0.2021,
+      "theoretical_loss": 5.184044116822849,
+      "tokens_seen": 58458112
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009921360937249237,
+      "loss": 0.2067,
+      "theoretical_loss": 5.1806052339978965,
+      "tokens_seen": 58720256
+    },
+    {
+      "epoch": 0.02,
+      "objective/train/advantage_avg": 0.008248094469308853,
+      "objective/train/docs_used": 29071,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.7817437648773193,
+      "objective/train/original_loss": 3.7817440032958984,
+      "objective/train/theoretical_loss": 5.1771859460001455,
+      "objective/train/tokens_used": 79442400,
+      "objective/train/value_avg": -0.0163726806640625,
+      "objective/train/value_loss": 0.00034792307997122407,
+      "objective/train/value_max": -0.003795623779296875,
+      "objective/train/value_min": -0.0994873046875,
+      "objective/train/value_reward_corr": 0.447085684972137,
+      "objective/train/value_std": 0.00957489013671875,
+      "objective/train/weight_avg": 1.0084155797958374,
+      "objective/train/weighted_lm_loss": 3.8198161125183105,
+      "objective/train/weights_max": 1.0903370380401611,
+      "objective/train/weights_min": 0.37122613191604614,
+      "theoretical_loss": 5.1771859460001455,
+      "tokens_seen": 58982400
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009920558497833415,
+      "loss": 0.1986,
+      "theoretical_loss": 5.1771859460001455,
+      "tokens_seen": 58982400
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000991975605841759,
+      "loss": 0.2047,
+      "theoretical_loss": 5.1737860548311065,
+      "tokens_seen": 59244544
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009918953619001765,
+      "loss": 0.2018,
+      "theoretical_loss": 5.170405365357794,
+      "tokens_seen": 59506688
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009918151179585941,
+      "loss": 0.2003,
+      "theoretical_loss": 5.167043685258852,
+      "tokens_seen": 59768832
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009917348740170117,
+      "loss": 0.2018,
+      "theoretical_loss": 5.163700824971922,
+      "tokens_seen": 60030976
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009916546300754293,
+      "loss": 0.2067,
+      "theoretical_loss": 5.160376597642223,
+      "tokens_seen": 60293120
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000991574386133847,
+      "loss": 0.2066,
+      "theoretical_loss": 5.157070819072301,
+      "tokens_seen": 60555264
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009914941421922646,
+      "loss": 0.2065,
+      "theoretical_loss": 5.153783307672935,
+      "tokens_seen": 60817408
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000991413898250682,
+      "loss": 0.2048,
+      "theoretical_loss": 5.150513884415149,
+      "tokens_seen": 61079552
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009913336543090998,
+      "loss": 0.2018,
+      "theoretical_loss": 5.14726237278331,
+      "tokens_seen": 61341696
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009912534103675172,
+      "loss": 0.2058,
+      "theoretical_loss": 5.144028598729285,
+      "tokens_seen": 61603840
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009911731664259348,
+      "loss": 0.2015,
+      "theoretical_loss": 5.140812390627624,
+      "tokens_seen": 61865984
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009910929224843524,
+      "loss": 0.2039,
+      "theoretical_loss": 5.137613579231737,
+      "tokens_seen": 62128128
+    },
+    {
+      "epoch": 0.02,
+      "objective/train/advantage_avg": 0.010975364595651627,
+      "objective/train/docs_used": 30276,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 4.172832489013672,
+      "objective/train/original_loss": 4.17283296585083,
+      "objective/train/theoretical_loss": 5.136020645048806,
+      "objective/train/tokens_used": 82719200,
+      "objective/train/value_avg": -0.016754150390625,
+      "objective/train/value_loss": 0.0010546700796112418,
+      "objective/train/value_max": -0.004718780517578125,
+      "objective/train/value_min": -0.08447265625,
+      "objective/train/value_reward_corr": 0.2364809296876953,
+      "objective/train/value_std": 0.007312774658203125,
+      "objective/train/weight_avg": 1.0114343166351318,
+      "objective/train/weighted_lm_loss": 4.22389554977417,
+      "objective/train/weights_max": 1.077765703201294,
+      "objective/train/weights_min": 0.372718870639801,
+      "theoretical_loss": 5.136020645048806,
+      "tokens_seen": 62259200
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.00099101267854277,
+      "loss": 0.2036,
+      "theoretical_loss": 5.134431997631053,
+      "tokens_seen": 62390272
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009909324346011876,
+      "loss": 0.1996,
+      "theoretical_loss": 5.1312674812091235,
+      "tokens_seen": 62652416
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009908521906596053,
+      "loss": 0.1987,
+      "theoretical_loss": 5.128119867602646,
+      "tokens_seen": 62914560
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009907719467180229,
+      "loss": 0.2017,
+      "theoretical_loss": 5.124988996661393,
+      "tokens_seen": 63176704
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009906917027764405,
+      "loss": 0.1973,
+      "theoretical_loss": 5.121874710409012,
+      "tokens_seen": 63438848
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000990611458834858,
+      "loss": 0.1999,
+      "theoretical_loss": 5.118776853004677,
+      "tokens_seen": 63700992
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009905312148932755,
+      "loss": 0.1977,
+      "theoretical_loss": 5.115695270705579,
+      "tokens_seen": 63963136
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000990450970951693,
+      "loss": 0.1956,
+      "theoretical_loss": 5.112629811830217,
+      "tokens_seen": 64225280
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009903707270101107,
+      "loss": 0.1989,
+      "theoretical_loss": 5.10958032672248,
+      "tokens_seen": 64487424
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009902904830685283,
+      "loss": 0.1946,
+      "theoretical_loss": 5.106546667716508,
+      "tokens_seen": 64749568
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000990210239126946,
+      "loss": 0.2015,
+      "theoretical_loss": 5.103528689102281,
+      "tokens_seen": 65011712
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009901299951853636,
+      "loss": 0.1959,
+      "theoretical_loss": 5.100526247091967,
+      "tokens_seen": 65273856
+    },
+    {
+      "epoch": 0.02,
+      "objective/train/advantage_avg": 0.007889259606599808,
+      "objective/train/docs_used": 31526,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 4.06732177734375,
+      "objective/train/original_loss": 4.06732177734375,
+      "objective/train/theoretical_loss": 5.097539199786951,
+      "objective/train/tokens_used": 85996000,
+      "objective/train/value_avg": -0.0160980224609375,
+      "objective/train/value_loss": 0.0005863794358447194,
+      "objective/train/value_max": -0.00464630126953125,
+      "objective/train/value_min": -0.1026611328125,
+      "objective/train/value_reward_corr": 0.33268385277114015,
+      "objective/train/value_std": 0.00913238525390625,
+      "objective/train/weight_avg": 1.0081636905670166,
+      "objective/train/weighted_lm_loss": 4.102910041809082,
+      "objective/train/weights_max": 1.1006338596343994,
+      "objective/train/weights_min": 0.49862930178642273,
+      "theoretical_loss": 5.097539199786951,
+      "tokens_seen": 65536000
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009900497512437812,
+      "loss": 0.199,
+      "theoretical_loss": 5.097539199786951,
+      "tokens_seen": 65536000
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009899695073021988,
+      "loss": 0.2,
+      "theoretical_loss": 5.094567407145588,
+      "tokens_seen": 65798144
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009898892633606162,
+      "loss": 0.1976,
+      "theoretical_loss": 5.09161073095161,
+      "tokens_seen": 66060288
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000989809019419034,
+      "loss": 0.1955,
+      "theoretical_loss": 5.0886690347832015,
+      "tokens_seen": 66322432
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009897287754774514,
+      "loss": 0.1975,
+      "theoretical_loss": 5.0857421839827275,
+      "tokens_seen": 66584576
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000989648531535869,
+      "loss": 0.193,
+      "theoretical_loss": 5.082830045627072,
+      "tokens_seen": 66846720
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009895682875942866,
+      "loss": 0.1985,
+      "theoretical_loss": 5.079932488498602,
+      "tokens_seen": 67108864
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009894880436527042,
+      "loss": 0.1952,
+      "theoretical_loss": 5.077049383056725,
+      "tokens_seen": 67371008
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009894077997111218,
+      "loss": 0.1976,
+      "theoretical_loss": 5.074180601410026,
+      "tokens_seen": 67633152
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009893275557695395,
+      "loss": 0.1896,
+      "theoretical_loss": 5.0713260172889845,
+      "tokens_seen": 67895296
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000989247311827957,
+      "loss": 0.1925,
+      "theoretical_loss": 5.068485506019231,
+      "tokens_seen": 68157440
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009891670678863745,
+      "loss": 0.1949,
+      "theoretical_loss": 5.06565894449535,
+      "tokens_seen": 68419584
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009890868239447923,
+      "loss": 0.1939,
+      "theoretical_loss": 5.06284621115523,
+      "tokens_seen": 68681728
+    },
+    {
+      "epoch": 0.02,
+      "objective/train/advantage_avg": 0.008360575884580612,
+      "objective/train/docs_used": 32629,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.77909517288208,
+      "objective/train/original_loss": 3.77909517288208,
+      "objective/train/theoretical_loss": 5.061444992481711,
+      "objective/train/tokens_used": 89272800,
+      "objective/train/value_avg": -0.016082763671875,
+      "objective/train/value_loss": 0.0004964357358403504,
+      "objective/train/value_max": -0.00473785400390625,
+      "objective/train/value_min": -0.08770751953125,
+      "objective/train/value_reward_corr": 0.21250019484264865,
+      "objective/train/value_std": 0.00782012939453125,
+      "objective/train/weight_avg": 1.0085920095443726,
+      "objective/train/weighted_lm_loss": 3.811274290084839,
+      "objective/train/weights_max": 1.0717307329177856,
+      "objective/train/weights_min": 0.37172776460647583,
+      "theoretical_loss": 5.061444992481711,
+      "tokens_seen": 68812800
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009890065800032097,
+      "loss": 0.1985,
+      "theoretical_loss": 5.060047185954893,
+      "tokens_seen": 68943872
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009889263360616273,
+      "loss": 0.1918,
+      "theoretical_loss": 5.057261750343864,
+      "tokens_seen": 69206016
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000988846092120045,
+      "loss": 0.1924,
+      "theoretical_loss": 5.0544897872410095,
+      "tokens_seen": 69468160
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009887658481784625,
+      "loss": 0.191,
+      "theoretical_loss": 5.051731181010866,
+      "tokens_seen": 69730304
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009886856042368801,
+      "loss": 0.1899,
+      "theoretical_loss": 5.048985817440432,
+      "tokens_seen": 69992448
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009886053602952978,
+      "loss": 0.1937,
+      "theoretical_loss": 5.046253583716425,
+      "tokens_seen": 70254592
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009885251163537154,
+      "loss": 0.1905,
+      "theoretical_loss": 5.043534368402973,
+      "tokens_seen": 70516736
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000988444872412133,
+      "loss": 0.1894,
+      "theoretical_loss": 5.040828061419762,
+      "tokens_seen": 70778880
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009883646284705506,
+      "loss": 0.1918,
+      "theoretical_loss": 5.038134554020587,
+      "tokens_seen": 71041024
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000988284384528968,
+      "loss": 0.1892,
+      "theoretical_loss": 5.03545373877234,
+      "tokens_seen": 71303168
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009882041405873858,
+      "loss": 0.191,
+      "theoretical_loss": 5.032785509534391,
+      "tokens_seen": 71565312
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009881238966458032,
+      "loss": 0.1886,
+      "theoretical_loss": 5.030129761438376,
+      "tokens_seen": 71827456
+    },
+    {
+      "epoch": 0.02,
+      "objective/train/advantage_avg": 0.008393102325499058,
+      "objective/train/docs_used": 33987,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.8430488109588623,
+      "objective/train/original_loss": 3.8430490493774414,
+      "objective/train/theoretical_loss": 5.0274863908683685,
+      "objective/train/tokens_used": 92549600,
+      "objective/train/value_avg": -0.014801025390625,
+      "objective/train/value_loss": 0.0004412951529957354,
+      "objective/train/value_max": -0.004314422607421875,
+      "objective/train/value_min": -0.091064453125,
+      "objective/train/value_reward_corr": 0.1426956661504004,
+      "objective/train/value_std": 0.006847381591796875,
+      "objective/train/weight_avg": 1.0085957050323486,
+      "objective/train/weighted_lm_loss": 3.876188039779663,
+      "objective/train/weights_max": 1.0953395366668701,
+      "objective/train/weights_min": 0.3716682195663452,
+      "theoretical_loss": 5.0274863908683685,
+      "tokens_seen": 72089600
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009880436527042208,
+      "loss": 0.1877,
+      "theoretical_loss": 5.0274863908683685,
+      "tokens_seen": 72089600
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009879634087626384,
+      "loss": 0.1913,
+      "theoretical_loss": 5.024855295441432,
+      "tokens_seen": 72351744
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000987883164821056,
+      "loss": 0.1894,
+      "theoretical_loss": 5.022236373988544,
+      "tokens_seen": 72613888
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009878029208794737,
+      "loss": 0.1873,
+      "theoretical_loss": 5.01962952653588,
+      "tokens_seen": 72876032
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009877226769378913,
+      "loss": 0.1886,
+      "theoretical_loss": 5.017034654286462,
+      "tokens_seen": 73138176
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000987642432996309,
+      "loss": 0.1903,
+      "theoretical_loss": 5.0144516596021385,
+      "tokens_seen": 73400320
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009875621890547263,
+      "loss": 0.183,
+      "theoretical_loss": 5.011880445985916,
+      "tokens_seen": 73662464
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000987481945113144,
+      "loss": 0.1888,
+      "theoretical_loss": 5.009320918064615,
+      "tokens_seen": 73924608
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009874017011715615,
+      "loss": 0.1886,
+      "theoretical_loss": 5.006772981571855,
+      "tokens_seen": 74186752
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009873214572299791,
+      "loss": 0.1868,
+      "theoretical_loss": 5.004236543331345,
+      "tokens_seen": 74448896
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009872412132883967,
+      "loss": 0.1858,
+      "theoretical_loss": 5.001711511240506,
+      "tokens_seen": 74711040
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009871609693468144,
+      "loss": 0.1867,
+      "theoretical_loss": 4.999197794254371,
+      "tokens_seen": 74973184
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000987080725405232,
+      "loss": 0.1865,
+      "theoretical_loss": 4.9966953023697975,
+      "tokens_seen": 75235328
+    },
+    {
+      "epoch": 0.02,
+      "objective/train/advantage_avg": 0.007199964951723814,
+      "objective/train/docs_used": 35047,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.54990816116333,
+      "objective/train/original_loss": 3.54990816116333,
+      "objective/train/theoretical_loss": 4.995448237995667,
+      "objective/train/tokens_used": 95826400,
+      "objective/train/value_avg": -0.01419830322265625,
+      "objective/train/value_loss": 0.0003746067523024976,
+      "objective/train/value_max": -0.003337860107421875,
+      "objective/train/value_min": -0.08599853515625,
+      "objective/train/value_reward_corr": 0.27188860037013274,
+      "objective/train/value_std": 0.007312774658203125,
+      "objective/train/weight_avg": 1.0073779821395874,
+      "objective/train/weighted_lm_loss": 3.5777623653411865,
+      "objective/train/weights_max": 1.0898046493530273,
+      "objective/train/weights_min": 0.3704579770565033,
+      "theoretical_loss": 4.995448237995667,
+      "tokens_seen": 75366400
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009870004814636496,
+      "loss": 0.1861,
+      "theoretical_loss": 4.994203946609964,
+      "tokens_seen": 75497472
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000986920237522067,
+      "loss": 0.1833,
+      "theoretical_loss": 4.991723639009154,
+      "tokens_seen": 75759616
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009868399935804848,
+      "loss": 0.1844,
+      "theoretical_loss": 4.989254292597813,
+      "tokens_seen": 76021760
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009867597496389022,
+      "loss": 0.1873,
+      "theoretical_loss": 4.986795821387878,
+      "tokens_seen": 76283904
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009866795056973198,
+      "loss": 0.1853,
+      "theoretical_loss": 4.984348140358374,
+      "tokens_seen": 76546048
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009865992617557374,
+      "loss": 0.1862,
+      "theoretical_loss": 4.981911165441273,
+      "tokens_seen": 76808192
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000986519017814155,
+      "loss": 0.1847,
+      "theoretical_loss": 4.979484813507599,
+      "tokens_seen": 77070336
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009864387738725726,
+      "loss": 0.1831,
+      "theoretical_loss": 4.977069002353792,
+      "tokens_seen": 77332480
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009863585299309903,
+      "loss": 0.1819,
+      "theoretical_loss": 4.974663650688306,
+      "tokens_seen": 77594624
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009862782859894079,
+      "loss": 0.1851,
+      "theoretical_loss": 4.972268678118454,
+      "tokens_seen": 77856768
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009861980420478253,
+      "loss": 0.1813,
+      "theoretical_loss": 4.969884005137479,
+      "tokens_seen": 78118912
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000986117798106243,
+      "loss": 0.1859,
+      "theoretical_loss": 4.967509553111862,
+      "tokens_seen": 78381056
+    },
+    {
+      "epoch": 0.02,
+      "objective/train/advantage_avg": 0.005118685308843851,
+      "objective/train/docs_used": 36308,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.8223462104797363,
+      "objective/train/original_loss": 3.8223462104797363,
+      "objective/train/theoretical_loss": 4.96514524426884,
+      "objective/train/tokens_used": 99103200,
+      "objective/train/value_avg": -0.0140838623046875,
+      "objective/train/value_loss": 0.00042888522148132324,
+      "objective/train/value_max": -0.00323486328125,
+      "objective/train/value_min": -0.122314453125,
+      "objective/train/value_reward_corr": 0.37372443614015616,
+      "objective/train/value_std": 0.00824737548828125,
+      "objective/train/weight_avg": 1.0053234100341797,
+      "objective/train/weighted_lm_loss": 3.844878673553467,
+      "objective/train/weights_max": 1.107279658317566,
+      "objective/train/weights_min": 0.3930663764476776,
+      "theoretical_loss": 4.96514524426884,
+      "tokens_seen": 78643200
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009860375541646605,
+      "loss": 0.1816,
+      "theoretical_loss": 4.96514524426884,
+      "tokens_seen": 78643200
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009859573102230783,
+      "loss": 0.184,
+      "theoretical_loss": 4.962791001684167,
+      "tokens_seen": 78905344
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009858770662814957,
+      "loss": 0.1822,
+      "theoretical_loss": 4.960446749270055,
+      "tokens_seen": 79167488
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009857968223399133,
+      "loss": 0.184,
+      "theoretical_loss": 4.958112411763365,
+      "tokens_seen": 79429632
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000985716578398331,
+      "loss": 0.1829,
+      "theoretical_loss": 4.955787914713962,
+      "tokens_seen": 79691776
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009856363344567486,
+      "loss": 0.1834,
+      "theoretical_loss": 4.953473184473312,
+      "tokens_seen": 79953920
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009855560905151662,
+      "loss": 0.183,
+      "theoretical_loss": 4.951168148183246,
+      "tokens_seen": 80216064
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009854758465735838,
+      "loss": 0.1818,
+      "theoretical_loss": 4.948872733764926,
+      "tokens_seen": 80478208
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009853956026320014,
+      "loss": 0.1799,
+      "theoretical_loss": 4.946586869908014,
+      "tokens_seen": 80740352
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009853153586904188,
+      "loss": 0.1828,
+      "theoretical_loss": 4.944310486060004,
+      "tokens_seen": 81002496
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009852351147488366,
+      "loss": 0.1825,
+      "theoretical_loss": 4.942043512415751,
+      "tokens_seen": 81264640
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.000985154870807254,
+      "loss": 0.1838,
+      "theoretical_loss": 4.939785879907176,
+      "tokens_seen": 81526784
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009850746268656716,
+      "loss": 0.1813,
+      "theoretical_loss": 4.937537520193139,
+      "tokens_seen": 81788928
+    },
+    {
+      "epoch": 0.02,
+      "objective/train/advantage_avg": 0.004108819644898176,
+      "objective/train/docs_used": 37569,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.581354856491089,
+      "objective/train/original_loss": 3.5813543796539307,
+      "objective/train/theoretical_loss": 4.936416796473786,
+      "objective/train/tokens_used": 102380000,
+      "objective/train/value_avg": -0.0154571533203125,
+      "objective/train/value_loss": 0.001122858258895576,
+      "objective/train/value_max": -0.00327301025390625,
+      "objective/train/value_min": -0.1788330078125,
+      "objective/train/value_reward_corr": 0.38812370857659007,
+      "objective/train/value_std": 0.0128631591796875,
+      "objective/train/weight_avg": 1.004586100578308,
+      "objective/train/weighted_lm_loss": 3.600618600845337,
+      "objective/train/weights_max": 1.1683320999145508,
+      "objective/train/weights_min": 0.37106189131736755,
+      "theoretical_loss": 4.936416796473786,
+      "tokens_seen": 81920000
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009849943829240892,
+      "loss": 0.1815,
+      "theoretical_loss": 4.93529836564949,
+      "tokens_seen": 82051072
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 0.0009849141389825069,
+      "loss": 0.1813,
+      "theoretical_loss": 4.933068349359283,
+      "tokens_seen": 82313216
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009848338950409245,
+      "loss": 0.1783,
+      "theoretical_loss": 4.93084740510316,
+      "tokens_seen": 82575360
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000984753651099342,
+      "loss": 0.1779,
+      "theoretical_loss": 4.928635467349885,
+      "tokens_seen": 82837504
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009846734071577597,
+      "loss": 0.1768,
+      "theoretical_loss": 4.92643247124705,
+      "tokens_seen": 83099648
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009845931632161773,
+      "loss": 0.1805,
+      "theoretical_loss": 4.924238352611924,
+      "tokens_seen": 83361792
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009845129192745947,
+      "loss": 0.1793,
+      "theoretical_loss": 4.922053047922455,
+      "tokens_seen": 83623936
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009844326753330123,
+      "loss": 0.1758,
+      "theoretical_loss": 4.919876494308432,
+      "tokens_seen": 83886080
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.00098435243139143,
+      "loss": 0.1795,
+      "theoretical_loss": 4.917708629542775,
+      "tokens_seen": 84148224
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009842721874498475,
+      "loss": 0.1775,
+      "theoretical_loss": 4.915549392032985,
+      "tokens_seen": 84410368
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009841919435082652,
+      "loss": 0.1797,
+      "theoretical_loss": 4.913398720812719,
+      "tokens_seen": 84672512
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009841116995666828,
+      "loss": 0.176,
+      "theoretical_loss": 4.9112565555335115,
+      "tokens_seen": 84934656
+    },
+    {
+      "epoch": 0.03,
+      "objective/train/advantage_avg": 0.009054622612893581,
+      "objective/train/docs_used": 38678,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.4969615936279297,
+      "objective/train/original_loss": 3.496961832046509,
+      "objective/train/theoretical_loss": 4.909122836456632,
+      "objective/train/tokens_used": 105656800,
+      "objective/train/value_avg": -0.013458251953125,
+      "objective/train/value_loss": 0.00028301688143983483,
+      "objective/train/value_max": -0.0036067962646484375,
+      "objective/train/value_min": -0.09588623046875,
+      "objective/train/value_reward_corr": 0.19258322737032904,
+      "objective/train/value_std": 0.00646209716796875,
+      "objective/train/weight_avg": 1.0091947317123413,
+      "objective/train/weighted_lm_loss": 3.5327301025390625,
+      "objective/train/weights_max": 1.0849108695983887,
+      "objective/train/weights_min": 0.8251521587371826,
+      "theoretical_loss": 4.909122836456632,
+      "tokens_seen": 85196800
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009840314556251004,
+      "loss": 0.1783,
+      "theoretical_loss": 4.909122836456632,
+      "tokens_seen": 85196800
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009839512116835178,
+      "loss": 0.1773,
+      "theoretical_loss": 4.906997504445066,
+      "tokens_seen": 85458944
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009838709677419356,
+      "loss": 0.1789,
+      "theoretical_loss": 4.904880500955633,
+      "tokens_seen": 85721088
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000983790723800353,
+      "loss": 0.1752,
+      "theoretical_loss": 4.90277176803123,
+      "tokens_seen": 85983232
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009837104798587706,
+      "loss": 0.1746,
+      "theoretical_loss": 4.9006712482931984,
+      "tokens_seen": 86245376
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009836302359171882,
+      "loss": 0.1762,
+      "theoretical_loss": 4.89857888493381,
+      "tokens_seen": 86507520
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009835499919756058,
+      "loss": 0.1782,
+      "theoretical_loss": 4.896494621708882,
+      "tokens_seen": 86769664
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009834697480340234,
+      "loss": 0.1751,
+      "theoretical_loss": 4.8944184029305,
+      "tokens_seen": 87031808
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000983389504092441,
+      "loss": 0.1761,
+      "theoretical_loss": 4.892350173459863,
+      "tokens_seen": 87293952
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009833092601508587,
+      "loss": 0.1734,
+      "theoretical_loss": 4.890289878700239,
+      "tokens_seen": 87556096
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009832290162092763,
+      "loss": 0.1735,
+      "theoretical_loss": 4.888237464590028,
+      "tokens_seen": 87818240
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000983148772267694,
+      "loss": 0.1766,
+      "theoretical_loss": 4.8861928775959464,
+      "tokens_seen": 88080384
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009830685283261113,
+      "loss": 0.1752,
+      "theoretical_loss": 4.884156064706302,
+      "tokens_seen": 88342528
+    },
+    {
+      "epoch": 0.03,
+      "objective/train/advantage_avg": 0.008355207741260529,
+      "objective/train/docs_used": 39491,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.4316039085388184,
+      "objective/train/original_loss": 3.4316043853759766,
+      "objective/train/theoretical_loss": 4.8831405571254844,
+      "objective/train/tokens_used": 108933600,
+      "objective/train/value_avg": -0.01488494873046875,
+      "objective/train/value_loss": 0.0006291866302490234,
+      "objective/train/value_max": -0.0031604766845703125,
+      "objective/train/value_min": -0.15185546875,
+      "objective/train/value_reward_corr": 0.39128727304023586,
+      "objective/train/value_std": 0.00885772705078125,
+      "objective/train/weight_avg": 1.0086404085159302,
+      "objective/train/weighted_lm_loss": 3.462775230407715,
+      "objective/train/weights_max": 1.1140156984329224,
+      "objective/train/weights_min": 0.3899836838245392,
+      "theoretical_loss": 4.8831405571254844,
+      "tokens_seen": 88473600
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009829882843845291,
+      "loss": 0.1745,
+      "theoretical_loss": 4.882126973424384,
+      "tokens_seen": 88604672
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009829080404429465,
+      "loss": 0.1729,
+      "theoretical_loss": 4.880105551761961,
+      "tokens_seen": 88866816
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009828277965013641,
+      "loss": 0.1784,
+      "theoretical_loss": 4.87809174823286,
+      "tokens_seen": 89128960
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009827475525597817,
+      "loss": 0.1722,
+      "theoretical_loss": 4.876085511846673,
+      "tokens_seen": 89391104
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009826673086181994,
+      "loss": 0.1772,
+      "theoretical_loss": 4.874086792102535,
+      "tokens_seen": 89653248
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000982587064676617,
+      "loss": 0.1766,
+      "theoretical_loss": 4.872095538983015,
+      "tokens_seen": 89915392
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009825068207350346,
+      "loss": 0.1784,
+      "theoretical_loss": 4.870111702948094,
+      "tokens_seen": 90177536
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009824265767934522,
+      "loss": 0.1736,
+      "theoretical_loss": 4.868135234929232,
+      "tokens_seen": 90439680
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009823463328518696,
+      "loss": 0.1743,
+      "theoretical_loss": 4.866166086323535,
+      "tokens_seen": 90701824
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009822660889102874,
+      "loss": 0.1779,
+      "theoretical_loss": 4.864204208988003,
+      "tokens_seen": 90963968
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009821858449687048,
+      "loss": 0.1733,
+      "theoretical_loss": 4.86224955523387,
+      "tokens_seen": 91226112
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009821056010271224,
+      "loss": 0.1766,
+      "theoretical_loss": 4.860302077821023,
+      "tokens_seen": 91488256
+    },
+    {
+      "epoch": 0.03,
+      "objective/train/advantage_avg": 0.006558452267199755,
+      "objective/train/docs_used": 40748,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.2776377201080322,
+      "objective/train/original_loss": 3.2776379585266113,
+      "objective/train/theoretical_loss": 4.858361729952518,
+      "objective/train/tokens_used": 112210400,
+      "objective/train/value_avg": -0.0146636962890625,
+      "objective/train/value_loss": 0.0004739796568173915,
+      "objective/train/value_max": -0.00359344482421875,
+      "objective/train/value_min": -0.0782470703125,
+      "objective/train/value_reward_corr": 0.21044601071028804,
+      "objective/train/value_std": 0.007022857666015625,
+      "objective/train/weight_avg": 1.0067769289016724,
+      "objective/train/weighted_lm_loss": 3.300762176513672,
+      "objective/train/weights_max": 1.0740199089050293,
+      "objective/train/weights_min": 0.3733791708946228,
+      "theoretical_loss": 4.858361729952518,
+      "tokens_seen": 91750400
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.00098202535708554,
+      "loss": 0.1726,
+      "theoretical_loss": 4.858361729952518,
+      "tokens_seen": 91750400
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009819451131439577,
+      "loss": 0.1719,
+      "theoretical_loss": 4.856428465269159,
+      "tokens_seen": 92012544
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009818648692023753,
+      "loss": 0.1728,
+      "theoretical_loss": 4.8545022378441836,
+      "tokens_seen": 92274688
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009817846252607929,
+      "loss": 0.174,
+      "theoretical_loss": 4.852583002178001,
+      "tokens_seen": 92536832
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009817043813192105,
+      "loss": 0.1715,
+      "theoretical_loss": 4.8506707131930344,
+      "tokens_seen": 92798976
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000981624137377628,
+      "loss": 0.1724,
+      "theoretical_loss": 4.848765326228618,
+      "tokens_seen": 93061120
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009815438934360455,
+      "loss": 0.1693,
+      "theoretical_loss": 4.846866797035984,
+      "tokens_seen": 93323264
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009814636494944631,
+      "loss": 0.1727,
+      "theoretical_loss": 4.844975081773322,
+      "tokens_seen": 93585408
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009813834055528807,
+      "loss": 0.1724,
+      "theoretical_loss": 4.843090137000904,
+      "tokens_seen": 93847552
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009813031616112983,
+      "loss": 0.1726,
+      "theoretical_loss": 4.841211919676287,
+      "tokens_seen": 94109696
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000981222917669716,
+      "loss": 0.1679,
+      "theoretical_loss": 4.839340387149586,
+      "tokens_seen": 94371840
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009811426737281336,
+      "loss": 0.1715,
+      "theoretical_loss": 4.837475497158817,
+      "tokens_seen": 94633984
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009810624297865512,
+      "loss": 0.1691,
+      "theoretical_loss": 4.835617207825303,
+      "tokens_seen": 94896128
+    },
+    {
+      "epoch": 0.03,
+      "objective/train/advantage_avg": -0.0007320477161556482,
+      "objective/train/docs_used": 42039,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.1431312561035156,
+      "objective/train/original_loss": 3.1431312561035156,
+      "objective/train/theoretical_loss": 4.8346905254216255,
+      "objective/train/tokens_used": 115487200,
+      "objective/train/value_avg": -0.01483154296875,
+      "objective/train/value_loss": 0.0013606772990897298,
+      "objective/train/value_max": -0.0031108856201171875,
+      "objective/train/value_min": -0.1416015625,
+      "objective/train/value_reward_corr": 0.5555549113783049,
+      "objective/train/value_std": 0.0105743408203125,
+      "objective/train/weight_avg": 0.9998704791069031,
+      "objective/train/weighted_lm_loss": 3.1514244079589844,
+      "objective/train/weights_max": 1.0741158723831177,
+      "objective/train/weights_min": 0.3726819157600403,
+      "theoretical_loss": 4.8346905254216255,
+      "tokens_seen": 95027200
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009809821858449686,
+      "loss": 0.1675,
+      "theoretical_loss": 4.83376547764915,
+      "tokens_seen": 95158272
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009809019419033864,
+      "loss": 0.1731,
+      "theoretical_loss": 4.831920265504792,
+      "tokens_seen": 95420416
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009808216979618038,
+      "loss": 0.1728,
+      "theoretical_loss": 4.830081530636594,
+      "tokens_seen": 95682560
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009807414540202216,
+      "loss": 0.1722,
+      "theoretical_loss": 4.8282492326545245,
+      "tokens_seen": 95944704
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000980661210078639,
+      "loss": 0.1718,
+      "theoretical_loss": 4.826423331529884,
+      "tokens_seen": 96206848
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009805809661370566,
+      "loss": 0.1696,
+      "theoretical_loss": 4.824603787591102,
+      "tokens_seen": 96468992
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009805007221954742,
+      "loss": 0.1729,
+      "theoretical_loss": 4.822790561519591,
+      "tokens_seen": 96731136
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009804204782538919,
+      "loss": 0.1719,
+      "theoretical_loss": 4.8209836143456535,
+      "tokens_seen": 96993280
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009803402343123095,
+      "loss": 0.1678,
+      "theoretical_loss": 4.81918290744446,
+      "tokens_seen": 97255424
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000980259990370727,
+      "loss": 0.1712,
+      "theoretical_loss": 4.817388402532074,
+      "tokens_seen": 97517568
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009801797464291447,
+      "loss": 0.1768,
+      "theoretical_loss": 4.815600061661536,
+      "tokens_seen": 97779712
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000980099502487562,
+      "loss": 0.1706,
+      "theoretical_loss": 4.813817847219008,
+      "tokens_seen": 98041856
+    },
+    {
+      "epoch": 0.03,
+      "objective/train/advantage_avg": 0.006740586832165718,
+      "objective/train/docs_used": 43181,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.3436176776885986,
+      "objective/train/original_loss": 3.3436179161071777,
+      "objective/train/theoretical_loss": 4.812041721919962,
+      "objective/train/tokens_used": 118764000,
+      "objective/train/value_avg": -0.013214111328125,
+      "objective/train/value_loss": 0.0005143888993188739,
+      "objective/train/value_max": -0.0032100677490234375,
+      "objective/train/value_min": -0.17919921875,
+      "objective/train/value_reward_corr": 0.2184640658155713,
+      "objective/train/value_std": 0.008026123046875,
+      "objective/train/weight_avg": 1.0069705247879028,
+      "objective/train/weighted_lm_loss": 3.366760492324829,
+      "objective/train/weights_max": 1.1044021844863892,
+      "objective/train/weights_min": 0.3713989555835724,
+      "theoretical_loss": 4.812041721919962,
+      "tokens_seen": 98304000
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.00098001925854598,
+      "loss": 0.1654,
+      "theoretical_loss": 4.812041721919962,
+      "tokens_seen": 98304000
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009799390146043973,
+      "loss": 0.1666,
+      "theoretical_loss": 4.810271648805427,
+      "tokens_seen": 98566144
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000979858770662815,
+      "loss": 0.1702,
+      "theoretical_loss": 4.8085075912383015,
+      "tokens_seen": 98828288
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009797785267212325,
+      "loss": 0.1683,
+      "theoretical_loss": 4.806749512899687,
+      "tokens_seen": 99090432
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009796982827796502,
+      "loss": 0.1686,
+      "theoretical_loss": 4.804997377785307,
+      "tokens_seen": 99352576
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009796180388380678,
+      "loss": 0.1674,
+      "theoretical_loss": 4.8032511502019535,
+      "tokens_seen": 99614720
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009795377948964854,
+      "loss": 0.1677,
+      "theoretical_loss": 4.801510794763988,
+      "tokens_seen": 99876864
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000979457550954903,
+      "loss": 0.1666,
+      "theoretical_loss": 4.799776276389897,
+      "tokens_seen": 100139008
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009793773070133206,
+      "loss": 0.1693,
+      "theoretical_loss": 4.798047560298882,
+      "tokens_seen": 100401152
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009792970630717382,
+      "loss": 0.1695,
+      "theoretical_loss": 4.796324612007515,
+      "tokens_seen": 100663296
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009792168191301556,
+      "loss": 0.1668,
+      "theoretical_loss": 4.794607397326421,
+      "tokens_seen": 100925440
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009791365751885732,
+      "loss": 0.1685,
+      "theoretical_loss": 4.792895882357019,
+      "tokens_seen": 101187584
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009790563312469908,
+      "loss": 0.1679,
+      "theoretical_loss": 4.791190033488302,
+      "tokens_seen": 101449728
+    },
+    {
+      "epoch": 0.03,
+      "objective/train/advantage_avg": 0.005793612916022539,
+      "objective/train/docs_used": 44310,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.2112574577331543,
+      "objective/train/original_loss": 3.2112574577331543,
+      "objective/train/theoretical_loss": 4.790339223416113,
+      "objective/train/tokens_used": 122040800,
+      "objective/train/value_avg": -0.01383209228515625,
+      "objective/train/value_loss": 0.0006092512048780918,
+      "objective/train/value_max": -0.0031719207763671875,
+      "objective/train/value_min": -0.1739501953125,
+      "objective/train/value_reward_corr": 0.3657779485356697,
+      "objective/train/value_std": 0.0101318359375,
+      "objective/train/weight_avg": 1.0060677528381348,
+      "objective/train/weighted_lm_loss": 3.2284328937530518,
+      "objective/train/weights_max": 1.1621463298797607,
+      "objective/train/weights_min": 0.3706558644771576,
+      "theoretical_loss": 4.790339223416113,
+      "tokens_seen": 101580800
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009789760873054085,
+      "loss": 0.1651,
+      "theoretical_loss": 4.7894898173936635,
+      "tokens_seen": 101711872
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000978895843363826,
+      "loss": 0.1639,
+      "theoretical_loss": 4.787795201027757,
+      "tokens_seen": 101974016
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009788155994222437,
+      "loss": 0.1675,
+      "theoretical_loss": 4.786106151623423,
+      "tokens_seen": 102236160
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009787353554806613,
+      "loss": 0.1616,
+      "theoretical_loss": 4.784422636688622,
+      "tokens_seen": 102498304
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000978655111539079,
+      "loss": 0.1641,
+      "theoretical_loss": 4.782744624003442,
+      "tokens_seen": 102760448
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009785748675974963,
+      "loss": 0.167,
+      "theoretical_loss": 4.781072081617127,
+      "tokens_seen": 103022592
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000978494623655914,
+      "loss": 0.166,
+      "theoretical_loss": 4.779404977845148,
+      "tokens_seen": 103284736
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009784143797143315,
+      "loss": 0.1642,
+      "theoretical_loss": 4.777743281266321,
+      "tokens_seen": 103546880
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009783341357727491,
+      "loss": 0.1611,
+      "theoretical_loss": 4.776086960719956,
+      "tokens_seen": 103809024
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009782538918311668,
+      "loss": 0.1675,
+      "theoretical_loss": 4.774435985303043,
+      "tokens_seen": 104071168
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009781736478895844,
+      "loss": 0.1605,
+      "theoretical_loss": 4.772790324367482,
+      "tokens_seen": 104333312
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000978093403948002,
+      "loss": 0.1659,
+      "theoretical_loss": 4.771149947517346,
+      "tokens_seen": 104595456
+    },
+    {
+      "epoch": 0.03,
+      "objective/train/advantage_avg": 0.006030434276908636,
+      "objective/train/docs_used": 45444,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.323787212371826,
+      "objective/train/original_loss": 3.323786973953247,
+      "objective/train/theoretical_loss": 4.769514824606178,
+      "objective/train/tokens_used": 125317600,
+      "objective/train/value_avg": -0.01244354248046875,
+      "objective/train/value_loss": 0.0006740050157532096,
+      "objective/train/value_max": -0.00269317626953125,
+      "objective/train/value_min": -0.1654052734375,
+      "objective/train/value_reward_corr": 0.42685833190639455,
+      "objective/train/value_std": 0.008056640625,
+      "objective/train/weight_avg": 1.0063230991363525,
+      "objective/train/weighted_lm_loss": 3.346615791320801,
+      "objective/train/weights_max": 1.1140156984329224,
+      "objective/train/weights_min": 0.3717136085033417,
+      "theoretical_loss": 4.769514824606178,
+      "tokens_seen": 104857600
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009780131600064196,
+      "loss": 0.1615,
+      "theoretical_loss": 4.769514824606178,
+      "tokens_seen": 104857600
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009779329160648372,
+      "loss": 0.1595,
+      "theoretical_loss": 4.7678849257343305,
+      "tokens_seen": 105119744
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009778526721232546,
+      "loss": 0.1621,
+      "theoretical_loss": 4.766260221246329,
+      "tokens_seen": 105381888
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009777724281816724,
+      "loss": 0.1613,
+      "theoretical_loss": 4.764640681728281,
+      "tokens_seen": 105644032
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009776921842400898,
+      "loss": 0.1644,
+      "theoretical_loss": 4.763026278005314,
+      "tokens_seen": 105906176
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009776119402985074,
+      "loss": 0.1608,
+      "theoretical_loss": 4.761416981139046,
+      "tokens_seen": 106168320
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000977531696356925,
+      "loss": 0.1626,
+      "theoretical_loss": 4.75981276242509,
+      "tokens_seen": 106430464
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009774514524153427,
+      "loss": 0.162,
+      "theoretical_loss": 4.758213593390595,
+      "tokens_seen": 106692608
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009773712084737603,
+      "loss": 0.1606,
+      "theoretical_loss": 4.756619445791808,
+      "tokens_seen": 106954752
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009772909645321779,
+      "loss": 0.1626,
+      "theoretical_loss": 4.755030291611678,
+      "tokens_seen": 107216896
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009772107205905955,
+      "loss": 0.1627,
+      "theoretical_loss": 4.753446103057492,
+      "tokens_seen": 107479040
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000977130476649013,
+      "loss": 0.1607,
+      "theoretical_loss": 4.751866852558529,
+      "tokens_seen": 107741184
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009770502327074307,
+      "loss": 0.1627,
+      "theoretical_loss": 4.7502925127637585,
+      "tokens_seen": 108003328
+    },
+    {
+      "epoch": 0.03,
+      "objective/train/advantage_avg": 0.0075781201012432575,
+      "objective/train/docs_used": 46682,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.372586250305176,
+      "objective/train/original_loss": 3.372586727142334,
+      "objective/train/theoretical_loss": 4.749507175892756,
+      "objective/train/tokens_used": 128594400,
+      "objective/train/value_avg": -0.013763427734375,
+      "objective/train/value_loss": 0.0003684050461743027,
+      "objective/train/value_max": -0.0030040740966796875,
+      "objective/train/value_min": -0.1995849609375,
+      "objective/train/value_reward_corr": 0.3359668813555781,
+      "objective/train/value_std": 0.0087127685546875,
+      "objective/train/weight_avg": 1.007756233215332,
+      "objective/train/weighted_lm_loss": 3.40148663520813,
+      "objective/train/weights_max": 1.1865150928497314,
+      "objective/train/weights_min": 0.3717136085033417,
+      "theoretical_loss": 4.749507175892756,
+      "tokens_seen": 108134400
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009769699887658481,
+      "loss": 0.1621,
+      "theoretical_loss": 4.7487230565395535,
+      "tokens_seen": 108265472
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009768897448242657,
+      "loss": 0.1611,
+      "theoretical_loss": 4.747158456967452,
+      "tokens_seen": 108527616
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009768095008826833,
+      "loss": 0.162,
+      "theoretical_loss": 4.745598687341927,
+      "tokens_seen": 108789760
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000976729256941101,
+      "loss": 0.1631,
+      "theoretical_loss": 4.744043721168196,
+      "tokens_seen": 109051904
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009766490129995186,
+      "loss": 0.1598,
+      "theoretical_loss": 4.74249353216006,
+      "tokens_seen": 109314048
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009765687690579362,
+      "loss": 0.1589,
+      "theoretical_loss": 4.740948094237761,
+      "tokens_seen": 109576192
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009764885251163538,
+      "loss": 0.1612,
+      "theoretical_loss": 4.739407381525874,
+      "tokens_seen": 109838336
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009764082811747713,
+      "loss": 0.1607,
+      "theoretical_loss": 4.7378713683512235,
+      "tokens_seen": 110100480
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009763280372331889,
+      "loss": 0.1598,
+      "theoretical_loss": 4.7363400292408215,
+      "tokens_seen": 110362624
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009762477932916065,
+      "loss": 0.1606,
+      "theoretical_loss": 4.734813338919842,
+      "tokens_seen": 110624768
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000976167549350024,
+      "loss": 0.1582,
+      "theoretical_loss": 4.733291272309609,
+      "tokens_seen": 110886912
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009760873054084417,
+      "loss": 0.1579,
+      "theoretical_loss": 4.731773804525616,
+      "tokens_seen": 111149056
+    },
+    {
+      "epoch": 0.03,
+      "objective/train/advantage_avg": 0.004441777244210243,
+      "objective/train/docs_used": 47788,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.1115875244140625,
+      "objective/train/original_loss": 3.1115875244140625,
+      "objective/train/theoretical_loss": 4.73026091087557,
+      "objective/train/tokens_used": 131871200,
+      "objective/train/value_avg": -0.01221466064453125,
+      "objective/train/value_loss": 0.0004770817467942834,
+      "objective/train/value_max": -0.00266265869140625,
+      "objective/train/value_min": -0.118408203125,
+      "objective/train/value_reward_corr": 0.48949773977616423,
+      "objective/train/value_std": 0.007781982421875,
+      "objective/train/weight_avg": 1.0046648979187012,
+      "objective/train/weighted_lm_loss": 3.124300003051758,
+      "objective/train/weights_max": 1.0862188339233398,
+      "objective/train/weights_min": 0.3779332637786865,
+      "theoretical_loss": 4.73026091087557,
+      "tokens_seen": 111411200
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009760070614668593,
+      "loss": 0.1564,
+      "theoretical_loss": 4.73026091087557,
+      "tokens_seen": 111411200
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009759268175252769,
+      "loss": 0.1583,
+      "theoretical_loss": 4.728752566857459,
+      "tokens_seen": 111673344
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009758465735836945,
+      "loss": 0.1586,
+      "theoretical_loss": 4.72724874815764,
+      "tokens_seen": 111935488
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009757663296421121,
+      "loss": 0.1579,
+      "theoretical_loss": 4.725749430648958,
+      "tokens_seen": 112197632
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009756860857005296,
+      "loss": 0.1593,
+      "theoretical_loss": 4.724254590388881,
+      "tokens_seen": 112459776
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009756058417589472,
+      "loss": 0.1575,
+      "theoretical_loss": 4.722764203617663,
+      "tokens_seen": 112721920
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009755255978173648,
+      "loss": 0.159,
+      "theoretical_loss": 4.7212782467565235,
+      "tokens_seen": 112984064
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009754453538757823,
+      "loss": 0.1581,
+      "theoretical_loss": 4.719796696405858,
+      "tokens_seen": 113246208
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009753651099342,
+      "loss": 0.1561,
+      "theoretical_loss": 4.718319529343462,
+      "tokens_seen": 113508352
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009752848659926175,
+      "loss": 0.1577,
+      "theoretical_loss": 4.716846722522781,
+      "tokens_seen": 113770496
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009752046220510353,
+      "loss": 0.155,
+      "theoretical_loss": 4.715378253071181,
+      "tokens_seen": 114032640
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009751243781094528,
+      "loss": 0.1538,
+      "theoretical_loss": 4.713914098288242,
+      "tokens_seen": 114294784
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009750441341678703,
+      "loss": 0.1555,
+      "theoretical_loss": 4.712454235644064,
+      "tokens_seen": 114556928
+    },
+    {
+      "epoch": 0.03,
+      "objective/train/advantage_avg": 0.008879084140062332,
+      "objective/train/docs_used": 48994,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.0145037174224854,
+      "objective/train/original_loss": 3.0145039558410645,
+      "objective/train/theoretical_loss": 4.711725906879634,
+      "objective/train/tokens_used": 135148000,
+      "objective/train/value_avg": -0.0135955810546875,
+      "objective/train/value_loss": 0.0002984044549521059,
+      "objective/train/value_max": -0.0026531219482421875,
+      "objective/train/value_min": -0.254150390625,
+      "objective/train/value_reward_corr": 0.3232125163801465,
+      "objective/train/value_std": 0.0081634521484375,
+      "objective/train/weight_avg": 1.0090208053588867,
+      "objective/train/weighted_lm_loss": 3.04550838470459,
+      "objective/train/weights_max": 1.097188949584961,
+      "objective/train/weights_min": 0.3715803325176239,
+      "theoretical_loss": 4.711725906879634,
+      "tokens_seen": 114688000
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.000974963890226288,
+      "loss": 0.1568,
+      "theoretical_loss": 4.710998642777606,
+      "tokens_seen": 114819072
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009748836462847055,
+      "loss": 0.1594,
+      "theoretical_loss": 4.709547297495034,
+      "tokens_seen": 115081216
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 0.0009748034023431231,
+      "loss": 0.1531,
+      "theoretical_loss": 4.708100177768094,
+      "tokens_seen": 115343360
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009747231584015407,
+      "loss": 0.156,
+      "theoretical_loss": 4.7066572617325075,
+      "tokens_seen": 115605504
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009746429144599583,
+      "loss": 0.159,
+      "theoretical_loss": 4.705218527686375,
+      "tokens_seen": 115867648
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009745626705183758,
+      "loss": 0.1583,
+      "theoretical_loss": 4.703783954088612,
+      "tokens_seen": 116129792
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009744824265767935,
+      "loss": 0.159,
+      "theoretical_loss": 4.702353519557398,
+      "tokens_seen": 116391936
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009744021826352111,
+      "loss": 0.1551,
+      "theoretical_loss": 4.700927202868639,
+      "tokens_seen": 116654080
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009743219386936286,
+      "loss": 0.1544,
+      "theoretical_loss": 4.699504982954452,
+      "tokens_seen": 116916224
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009742416947520463,
+      "loss": 0.1559,
+      "theoretical_loss": 4.698086838901676,
+      "tokens_seen": 117178368
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009741614508104638,
+      "loss": 0.154,
+      "theoretical_loss": 4.696672749950385,
+      "tokens_seen": 117440512
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009740812068688814,
+      "loss": 0.1536,
+      "theoretical_loss": 4.695262695492428,
+      "tokens_seen": 117702656
+    },
+    {
+      "epoch": 0.04,
+      "objective/train/advantage_avg": 0.005216541700065136,
+      "objective/train/docs_used": 50256,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 3.161985158920288,
+      "objective/train/original_loss": 3.161985397338867,
+      "objective/train/theoretical_loss": 4.693856655069986,
+      "objective/train/tokens_used": 138424800,
+      "objective/train/value_avg": -0.013153076171875,
+      "objective/train/value_loss": 0.0008696132572367787,
+      "objective/train/value_max": -0.00238800048828125,
+      "objective/train/value_min": -0.26220703125,
+      "objective/train/value_reward_corr": 0.28899940519313116,
+      "objective/train/value_std": 0.008697509765625,
+      "objective/train/weight_avg": 1.0055670738220215,
+      "objective/train/weighted_lm_loss": 3.1820406913757324,
+      "objective/train/weights_max": 1.1238499879837036,
+      "objective/train/weights_min": 0.22635191679000854,
+      "theoretical_loss": 4.693856655069986,
+      "tokens_seen": 117964800
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.000974000962927299,
+      "loss": 0.1545,
+      "theoretical_loss": 4.693856655069986,
+      "tokens_seen": 117964800
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009739207189857165,
+      "loss": 0.1559,
+      "theoretical_loss": 4.692454608374145,
+      "tokens_seen": 118226944
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009738404750441343,
+      "loss": 0.1526,
+      "theoretical_loss": 4.6910565352434785,
+      "tokens_seen": 118489088
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009737602311025518,
+      "loss": 0.1531,
+      "theoretical_loss": 4.689662415662662,
+      "tokens_seen": 118751232
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009736799871609694,
+      "loss": 0.1525,
+      "theoretical_loss": 4.688272229761087,
+      "tokens_seen": 119013376
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.000973599743219387,
+      "loss": 0.1542,
+      "theoretical_loss": 4.686885957811503,
+      "tokens_seen": 119275520
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009735194992778046,
+      "loss": 0.1505,
+      "theoretical_loss": 4.685503580228671,
+      "tokens_seen": 119537664
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009734392553362221,
+      "loss": 0.1508,
+      "theoretical_loss": 4.684125077568028,
+      "tokens_seen": 119799808
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009733590113946397,
+      "loss": 0.1548,
+      "theoretical_loss": 4.682750430524376,
+      "tokens_seen": 120061952
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009732787674530573,
+      "loss": 0.1532,
+      "theoretical_loss": 4.6813796199305795,
+      "tokens_seen": 120324096
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009731985235114748,
+      "loss": 0.1521,
+      "theoretical_loss": 4.68001262675628,
+      "tokens_seen": 120586240
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009731182795698925,
+      "loss": 0.1479,
+      "theoretical_loss": 4.678649432106627,
+      "tokens_seen": 120848384
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.00097303803562831,
+      "loss": 0.1518,
+      "theoretical_loss": 4.677290017221017,
+      "tokens_seen": 121110528
+    },
+    {
+      "epoch": 0.04,
+      "objective/train/advantage_avg": 0.006032022647559643,
+      "objective/train/docs_used": 51395,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.9272947311401367,
+      "objective/train/original_loss": 2.9272947311401367,
+      "objective/train/theoretical_loss": 4.676611721363443,
+      "objective/train/tokens_used": 141701600,
+      "objective/train/value_avg": -0.01284027099609375,
+      "objective/train/value_loss": 0.000364342937245965,
+      "objective/train/value_max": -0.001987457275390625,
+      "objective/train/value_min": -0.30517578125,
+      "objective/train/value_reward_corr": 0.3808824724269897,
+      "objective/train/value_std": 0.010894775390625,
+      "objective/train/weight_avg": 1.0062044858932495,
+      "objective/train/weighted_lm_loss": 2.945836067199707,
+      "objective/train/weights_max": 1.2145042419433594,
+      "objective/train/weights_min": 0.3716398775577545,
+      "theoretical_loss": 4.676611721363443,
+      "tokens_seen": 121241600
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009729577916867277,
+      "loss": 0.1482,
+      "theoretical_loss": 4.675934363471857,
+      "tokens_seen": 121372672
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009728775477451453,
+      "loss": 0.1491,
+      "theoretical_loss": 4.674582452363334,
+      "tokens_seen": 121634816
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009727973038035629,
+      "loss": 0.1481,
+      "theoretical_loss": 4.673234265530201,
+      "tokens_seen": 121896960
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009727170598619804,
+      "loss": 0.1474,
+      "theoretical_loss": 4.671889784736576,
+      "tokens_seen": 122159104
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.000972636815920398,
+      "loss": 0.148,
+      "theoretical_loss": 4.670548991874758,
+      "tokens_seen": 122421248
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009725565719788156,
+      "loss": 0.1472,
+      "theoretical_loss": 4.669211868964052,
+      "tokens_seen": 122683392
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009724763280372332,
+      "loss": 0.1496,
+      "theoretical_loss": 4.66787839814961,
+      "tokens_seen": 122945536
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009723960840956508,
+      "loss": 0.1483,
+      "theoretical_loss": 4.666548561701285,
+      "tokens_seen": 123207680
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009723158401540683,
+      "loss": 0.1466,
+      "theoretical_loss": 4.665222342012491,
+      "tokens_seen": 123469824
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009722355962124861,
+      "loss": 0.1473,
+      "theoretical_loss": 4.663899721599093,
+      "tokens_seen": 123731968
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009721553522709036,
+      "loss": 0.1437,
+      "theoretical_loss": 4.6625806830982865,
+      "tokens_seen": 123994112
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009720751083293211,
+      "loss": 0.1438,
+      "theoretical_loss": 4.661265209267507,
+      "tokens_seen": 124256256
+    },
+    {
+      "epoch": 0.04,
+      "objective/train/advantage_avg": 0.005084542091935873,
+      "objective/train/docs_used": 52572,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.7900288105010986,
+      "objective/train/original_loss": 2.7900285720825195,
+      "objective/train/theoretical_loss": 4.659953282983348,
+      "objective/train/tokens_used": 144978400,
+      "objective/train/value_avg": -0.01073455810546875,
+      "objective/train/value_loss": 0.00028898261371068656,
+      "objective/train/value_max": -0.001483917236328125,
+      "objective/train/value_min": -0.2239990234375,
+      "objective/train/value_reward_corr": 0.39944171540418233,
+      "objective/train/value_std": 0.008880615234375,
+      "objective/train/weight_avg": 1.00521981716156,
+      "objective/train/weighted_lm_loss": 2.8061976432800293,
+      "objective/train/weights_max": 1.1481295824050903,
+      "objective/train/weights_min": 0.3976996839046478,
+      "theoretical_loss": 4.659953282983348,
+      "tokens_seen": 124518400
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009719948643877388,
+      "loss": 0.1466,
+      "theoretical_loss": 4.659953282983348,
+      "tokens_seen": 124518400
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009719146204461563,
+      "loss": 0.1488,
+      "theoretical_loss": 4.658644887240481,
+      "tokens_seen": 124780544
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009718343765045739,
+      "loss": 0.1485,
+      "theoretical_loss": 4.657340005150602,
+      "tokens_seen": 125042688
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009717541325629915,
+      "loss": 0.1492,
+      "theoretical_loss": 4.656038619941382,
+      "tokens_seen": 125304832
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009716738886214091,
+      "loss": 0.1457,
+      "theoretical_loss": 4.654740714955429,
+      "tokens_seen": 125566976
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009715936446798266,
+      "loss": 0.145,
+      "theoretical_loss": 4.653446273649259,
+      "tokens_seen": 125829120
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009715134007382443,
+      "loss": 0.1478,
+      "theoretical_loss": 4.652155279592286,
+      "tokens_seen": 126091264
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009714331567966619,
+      "loss": 0.1463,
+      "theoretical_loss": 4.650867716465819,
+      "tokens_seen": 126353408
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009713529128550794,
+      "loss": 0.1463,
+      "theoretical_loss": 4.6495835680620665,
+      "tokens_seen": 126615552
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009712726689134971,
+      "loss": 0.1432,
+      "theoretical_loss": 4.648302818283158,
+      "tokens_seen": 126877696
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009711924249719146,
+      "loss": 0.1477,
+      "theoretical_loss": 4.64702545114017,
+      "tokens_seen": 127139840
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009711121810303323,
+      "loss": 0.1452,
+      "theoretical_loss": 4.645751450752172,
+      "tokens_seen": 127401984
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009710319370887498,
+      "loss": 0.1419,
+      "theoretical_loss": 4.644480801345268,
+      "tokens_seen": 127664128
+    },
+    {
+      "epoch": 0.04,
+      "objective/train/advantage_avg": 0.0065161967650055885,
+      "objective/train/docs_used": 53741,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.8748562335968018,
+      "objective/train/original_loss": 2.8748559951782227,
+      "objective/train/theoretical_loss": 4.6438467283594225,
+      "objective/train/tokens_used": 148255200,
+      "objective/train/value_avg": -0.0144805908203125,
+      "objective/train/value_loss": 0.0003499473095871508,
+      "objective/train/value_max": -0.00269317626953125,
+      "objective/train/value_min": -0.36376953125,
+      "objective/train/value_reward_corr": 0.4529447418859527,
+      "objective/train/value_std": 0.01155853271484375,
+      "objective/train/weight_avg": 1.0066853761672974,
+      "objective/train/weighted_lm_loss": 2.8947949409484863,
+      "objective/train/weights_max": 1.2178692817687988,
+      "objective/train/weights_min": 0.37203988432884216,
+      "theoretical_loss": 4.6438467283594225,
+      "tokens_seen": 127795200
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009709516931471673,
+      "loss": 0.1469,
+      "theoretical_loss": 4.643213487251664,
+      "tokens_seen": 127926272
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.000970871449205585,
+      "loss": 0.1431,
+      "theoretical_loss": 4.641949492908737,
+      "tokens_seen": 128188416
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009707912052640026,
+      "loss": 0.1445,
+      "theoretical_loss": 4.640688802858113,
+      "tokens_seen": 128450560
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009707109613224202,
+      "loss": 0.1478,
+      "theoretical_loss": 4.63943140174476,
+      "tokens_seen": 128712704
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009706307173808378,
+      "loss": 0.1449,
+      "theoretical_loss": 4.6381772743160905,
+      "tokens_seen": 128974848
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009705504734392554,
+      "loss": 0.1467,
+      "theoretical_loss": 4.636926405421065,
+      "tokens_seen": 129236992
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009704702294976729,
+      "loss": 0.1425,
+      "theoretical_loss": 4.635678780009318,
+      "tokens_seen": 129499136
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009703899855560905,
+      "loss": 0.1433,
+      "theoretical_loss": 4.634434383130284,
+      "tokens_seen": 129761280
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009703097416145081,
+      "loss": 0.1452,
+      "theoretical_loss": 4.633193199932336,
+      "tokens_seen": 130023424
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009702294976729256,
+      "loss": 0.145,
+      "theoretical_loss": 4.631955215661932,
+      "tokens_seen": 130285568
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009701492537313433,
+      "loss": 0.1436,
+      "theoretical_loss": 4.630720415662774,
+      "tokens_seen": 130547712
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009700690097897609,
+      "loss": 0.1374,
+      "theoretical_loss": 4.629488785374969,
+      "tokens_seen": 130809856
+    },
+    {
+      "epoch": 0.04,
+      "objective/train/advantage_avg": 0.005413413047790527,
+      "objective/train/docs_used": 54897,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.7717833518981934,
+      "objective/train/original_loss": 2.7717831134796143,
+      "objective/train/theoretical_loss": 4.628260310334209,
+      "objective/train/tokens_used": 151532000,
+      "objective/train/value_avg": -0.012939453125,
+      "objective/train/value_loss": 0.0007990457233972847,
+      "objective/train/value_max": -0.002124786376953125,
+      "objective/train/value_min": -0.2451171875,
+      "objective/train/value_reward_corr": 0.2380011948215656,
+      "objective/train/value_std": 0.01189422607421875,
+      "objective/train/weight_avg": 1.0057445764541626,
+      "objective/train/weighted_lm_loss": 2.7883291244506836,
+      "objective/train/weights_max": 1.2066720724105835,
+      "objective/train/weights_min": 0.22692078351974487,
+      "theoretical_loss": 4.628260310334209,
+      "tokens_seen": 131072000
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009699887658481786,
+      "loss": 0.1404,
+      "theoretical_loss": 4.628260310334209,
+      "tokens_seen": 131072000
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009699085219065961,
+      "loss": 0.1406,
+      "theoretical_loss": 4.6270349761709495,
+      "tokens_seen": 131334144
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009698282779650137,
+      "loss": 0.1373,
+      "theoretical_loss": 4.625812768609601,
+      "tokens_seen": 131596288
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009697480340234313,
+      "loss": 0.1405,
+      "theoretical_loss": 4.62459367346773,
+      "tokens_seen": 131858432
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009696677900818488,
+      "loss": 0.1395,
+      "theoretical_loss": 4.623377676655271,
+      "tokens_seen": 132120576
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009695875461402664,
+      "loss": 0.1398,
+      "theoretical_loss": 4.622164764173735,
+      "tokens_seen": 132382720
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.000969507302198684,
+      "loss": 0.1402,
+      "theoretical_loss": 4.62095492211544,
+      "tokens_seen": 132644864
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009694270582571016,
+      "loss": 0.1464,
+      "theoretical_loss": 4.6197481366627455,
+      "tokens_seen": 132907008
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009693468143155191,
+      "loss": 0.1408,
+      "theoretical_loss": 4.618544394087287,
+      "tokens_seen": 133169152
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009692665703739369,
+      "loss": 0.1341,
+      "theoretical_loss": 4.617343680749233,
+      "tokens_seen": 133431296
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009691863264323544,
+      "loss": 0.1365,
+      "theoretical_loss": 4.61614598309653,
+      "tokens_seen": 133693440
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009691060824907719,
+      "loss": 0.1375,
+      "theoretical_loss": 4.614951287664179,
+      "tokens_seen": 133955584
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009690258385491896,
+      "loss": 0.139,
+      "theoretical_loss": 4.613759581073502,
+      "tokens_seen": 134217728
+    },
+    {
+      "epoch": 0.04,
+      "objective/train/advantage_avg": 0.005763855762779713,
+      "objective/train/docs_used": 56110,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.6633012294769287,
+      "objective/train/original_loss": 2.663301467895508,
+      "objective/train/theoretical_loss": 4.613164844436379,
+      "objective/train/tokens_used": 154808800,
+      "objective/train/value_avg": -0.01210784912109375,
+      "objective/train/value_loss": 0.00031003085314296186,
+      "objective/train/value_max": -0.0025310516357421875,
+      "objective/train/value_min": -0.236572265625,
+      "objective/train/value_reward_corr": 0.1698134460465261,
+      "objective/train/value_std": 0.0075225830078125,
+      "objective/train/weight_avg": 1.0059170722961426,
+      "objective/train/weighted_lm_loss": 2.677839756011963,
+      "objective/train/weights_max": 1.1739808320999146,
+      "objective/train/weights_min": 0.7241759300231934,
+      "theoretical_loss": 4.613164844436379,
+      "tokens_seen": 134348800
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009689455946076071,
+      "loss": 0.1382,
+      "theoretical_loss": 4.612570850031418,
+      "tokens_seen": 134479872
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009688653506660247,
+      "loss": 0.1358,
+      "theoretical_loss": 4.611385081329736,
+      "tokens_seen": 134742016
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009687851067244423,
+      "loss": 0.1363,
+      "theoretical_loss": 4.610202261844444,
+      "tokens_seen": 135004160
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009687048627828599,
+      "loss": 0.1375,
+      "theoretical_loss": 4.6090223785350135,
+      "tokens_seen": 135266304
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009686246188412776,
+      "loss": 0.1364,
+      "theoretical_loss": 4.607845418443706,
+      "tokens_seen": 135528448
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009685443748996951,
+      "loss": 0.1384,
+      "theoretical_loss": 4.606671368694888,
+      "tokens_seen": 135790592
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009684641309581127,
+      "loss": 0.1369,
+      "theoretical_loss": 4.6055002164943595,
+      "tokens_seen": 136052736
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009683838870165303,
+      "loss": 0.1321,
+      "theoretical_loss": 4.604331949128672,
+      "tokens_seen": 136314880
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009683036430749479,
+      "loss": 0.1365,
+      "theoretical_loss": 4.603166553964474,
+      "tokens_seen": 136577024
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009682233991333654,
+      "loss": 0.1396,
+      "theoretical_loss": 4.60200401844785,
+      "tokens_seen": 136839168
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009681431551917831,
+      "loss": 0.133,
+      "theoretical_loss": 4.6008443301036746,
+      "tokens_seen": 137101312
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009680629112502006,
+      "loss": 0.1364,
+      "theoretical_loss": 4.5996874765349585,
+      "tokens_seen": 137363456
+    },
+    {
+      "epoch": 0.04,
+      "objective/train/advantage_avg": 0.0023926738649606705,
+      "objective/train/docs_used": 57180,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.8781347274780273,
+      "objective/train/original_loss": 2.8781347274780273,
+      "objective/train/theoretical_loss": 4.598533445422221,
+      "objective/train/tokens_used": 158085600,
+      "objective/train/value_avg": -0.0099029541015625,
+      "objective/train/value_loss": 0.00033420659019611776,
+      "objective/train/value_max": -0.0016613006591796875,
+      "objective/train/value_min": -0.1622314453125,
+      "objective/train/value_reward_corr": 0.23598254350018186,
+      "objective/train/value_std": 0.00702667236328125,
+      "objective/train/weight_avg": 1.0025557279586792,
+      "objective/train/weighted_lm_loss": 2.8859775066375732,
+      "objective/train/weights_max": 1.1505653858184814,
+      "objective/train/weights_min": 0.7221318483352661,
+      "theoretical_loss": 4.598533445422221,
+      "tokens_seen": 137625600
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009679826673086181,
+      "loss": 0.1361,
+      "theoretical_loss": 4.598533445422221,
+      "tokens_seen": 137625600
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009679024233670359,
+      "loss": 0.1347,
+      "theoretical_loss": 4.597382224522855,
+      "tokens_seen": 137887744
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009678221794254534,
+      "loss": 0.1362,
+      "theoretical_loss": 4.596233801670502,
+      "tokens_seen": 138149888
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.000967741935483871,
+      "loss": 0.1346,
+      "theoretical_loss": 4.595088164774435,
+      "tokens_seen": 138412032
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009676616915422886,
+      "loss": 0.1367,
+      "theoretical_loss": 4.593945301818941,
+      "tokens_seen": 138674176
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009675814476007062,
+      "loss": 0.1307,
+      "theoretical_loss": 4.592805200862726,
+      "tokens_seen": 138936320
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009675012036591238,
+      "loss": 0.133,
+      "theoretical_loss": 4.591667850038302,
+      "tokens_seen": 139198464
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009674209597175413,
+      "loss": 0.1317,
+      "theoretical_loss": 4.590533237551401,
+      "tokens_seen": 139460608
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009673407157759589,
+      "loss": 0.1315,
+      "theoretical_loss": 4.589401351680385,
+      "tokens_seen": 139722752
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009672604718343765,
+      "loss": 0.1337,
+      "theoretical_loss": 4.588272180775659,
+      "tokens_seen": 139984896
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009671802278927941,
+      "loss": 0.1296,
+      "theoretical_loss": 4.587145713259102,
+      "tokens_seen": 140247040
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009670999839512117,
+      "loss": 0.1301,
+      "theoretical_loss": 4.5860219376234905,
+      "tokens_seen": 140509184
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009670197400096294,
+      "loss": 0.1319,
+      "theoretical_loss": 4.584900842431934,
+      "tokens_seen": 140771328
+    },
+    {
+      "epoch": 0.04,
+      "objective/train/advantage_avg": 0.006780576426535845,
+      "objective/train/docs_used": 58312,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.43525767326355,
+      "objective/train/original_loss": 2.43525767326355,
+      "objective/train/theoretical_loss": 4.584341296447742,
+      "objective/train/tokens_used": 161362400,
+      "objective/train/value_avg": -0.010894775390625,
+      "objective/train/value_loss": 0.00027257209876552224,
+      "objective/train/value_max": -0.0015916824340820312,
+      "objective/train/value_min": -0.367919921875,
+      "objective/train/value_reward_corr": 0.26859711054421476,
+      "objective/train/value_std": 0.00720977783203125,
+      "objective/train/weight_avg": 1.0069042444229126,
+      "objective/train/weighted_lm_loss": 2.4517366886138916,
+      "objective/train/weights_max": 1.1828418970108032,
+      "objective/train/weights_min": 0.3722556531429291,
+      "theoretical_loss": 4.584341296447742,
+      "tokens_seen": 140902400
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009669394960680469,
+      "loss": 0.1284,
+      "theoretical_loss": 4.583782416317316,
+      "tokens_seen": 141033472
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009668592521264644,
+      "loss": 0.1302,
+      "theoretical_loss": 4.582666647981739,
+      "tokens_seen": 141295616
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009667790081848821,
+      "loss": 0.128,
+      "theoretical_loss": 4.581553526195974,
+      "tokens_seen": 141557760
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009666987642432996,
+      "loss": 0.1304,
+      "theoretical_loss": 4.580443039798922,
+      "tokens_seen": 141819904
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009666185203017172,
+      "loss": 0.1305,
+      "theoretical_loss": 4.57933517769707,
+      "tokens_seen": 142082048
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009665382763601348,
+      "loss": 0.1314,
+      "theoretical_loss": 4.578229928863959,
+      "tokens_seen": 142344192
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009664580324185524,
+      "loss": 0.1286,
+      "theoretical_loss": 4.57712728233966,
+      "tokens_seen": 142606336
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.00096637778847697,
+      "loss": 0.127,
+      "theoretical_loss": 4.576027227230245,
+      "tokens_seen": 142868480
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009662975445353877,
+      "loss": 0.1315,
+      "theoretical_loss": 4.574929752707274,
+      "tokens_seen": 143130624
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009662173005938052,
+      "loss": 0.128,
+      "theoretical_loss": 4.573834848007284,
+      "tokens_seen": 143392768
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009661370566522228,
+      "loss": 0.1269,
+      "theoretical_loss": 4.572742502431272,
+      "tokens_seen": 143654912
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009660568127106404,
+      "loss": 0.1264,
+      "theoretical_loss": 4.571652705344202,
+      "tokens_seen": 143917056
+    },
+    {
+      "epoch": 0.04,
+      "objective/train/advantage_avg": 0.003318200586363673,
+      "objective/train/docs_used": 59583,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.476109027862549,
+      "objective/train/original_loss": 2.476109027862549,
+      "objective/train/theoretical_loss": 4.570565446174504,
+      "objective/train/tokens_used": 164639200,
+      "objective/train/value_avg": -0.0129547119140625,
+      "objective/train/value_loss": 0.0007415720028802752,
+      "objective/train/value_max": -0.0013942718505859375,
+      "objective/train/value_min": -0.254150390625,
+      "objective/train/value_reward_corr": 0.4452921464992741,
+      "objective/train/value_std": 0.01247406005859375,
+      "objective/train/weight_avg": 1.0036523342132568,
+      "objective/train/weighted_lm_loss": 2.485182285308838,
+      "objective/train/weights_max": 1.1576626300811768,
+      "objective/train/weights_min": 0.37203988432884216,
+      "theoretical_loss": 4.570565446174504,
+      "tokens_seen": 144179200
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009659765687690579,
+      "loss": 0.1281,
+      "theoretical_loss": 4.570565446174504,
+      "tokens_seen": 144179200
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009658963248274756,
+      "loss": 0.1276,
+      "theoretical_loss": 4.569480714413578,
+      "tokens_seen": 144441344
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009658160808858931,
+      "loss": 0.1301,
+      "theoretical_loss": 4.568398499615305,
+      "tokens_seen": 144703488
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009657358369443107,
+      "loss": 0.1291,
+      "theoretical_loss": 4.56731879139557,
+      "tokens_seen": 144965632
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009656555930027284,
+      "loss": 0.1264,
+      "theoretical_loss": 4.566241579431776,
+      "tokens_seen": 145227776
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009655753490611459,
+      "loss": 0.1285,
+      "theoretical_loss": 4.565166853462371,
+      "tokens_seen": 145489920
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009654951051195635,
+      "loss": 0.13,
+      "theoretical_loss": 4.564094603286375,
+      "tokens_seen": 145752064
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009654148611779811,
+      "loss": 0.1248,
+      "theoretical_loss": 4.5630248187629245,
+      "tokens_seen": 146014208
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009653346172363987,
+      "loss": 0.1261,
+      "theoretical_loss": 4.561957489810798,
+      "tokens_seen": 146276352
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009652543732948162,
+      "loss": 0.1243,
+      "theoretical_loss": 4.5608926064079665,
+      "tokens_seen": 146538496
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009651741293532339,
+      "loss": 0.1281,
+      "theoretical_loss": 4.559830158591139,
+      "tokens_seen": 146800640
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009650938854116514,
+      "loss": 0.1229,
+      "theoretical_loss": 4.558770136455316,
+      "tokens_seen": 147062784
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009650136414700689,
+      "loss": 0.1253,
+      "theoretical_loss": 4.557712530153342,
+      "tokens_seen": 147324928
+    },
+    {
+      "epoch": 0.04,
+      "objective/train/advantage_avg": 0.002189334249123931,
+      "objective/train/docs_used": 60747,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.354844570159912,
+      "objective/train/original_loss": 2.354844570159912,
+      "objective/train/theoretical_loss": 4.557184629878522,
+      "objective/train/tokens_used": 167916000,
+      "objective/train/value_avg": -0.00820159912109375,
+      "objective/train/value_loss": 0.0003651907609310001,
+      "objective/train/value_max": -0.001239776611328125,
+      "objective/train/value_min": -0.23779296875,
+      "objective/train/value_reward_corr": 0.19971720203272725,
+      "objective/train/value_std": 0.00504302978515625,
+      "objective/train/weight_avg": 1.0023622512817383,
+      "objective/train/weighted_lm_loss": 2.3594002723693848,
+      "objective/train/weights_max": 1.057178020477295,
+      "objective/train/weights_min": 0.6187294125556946,
+      "theoretical_loss": 4.557184629878522,
+      "tokens_seen": 147456000
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009649333975284866,
+      "loss": 0.1229,
+      "theoretical_loss": 4.556657329895469,
+      "tokens_seen": 147587072
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009648531535869042,
+      "loss": 0.1265,
+      "theoretical_loss": 4.5556045259489135,
+      "tokens_seen": 147849216
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009647729096453219,
+      "loss": 0.1269,
+      "theoretical_loss": 4.554554108637437,
+      "tokens_seen": 148111360
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 0.0009646926657037394,
+      "loss": 0.1242,
+      "theoretical_loss": 4.553506068340907,
+      "tokens_seen": 148373504
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000964612421762157,
+      "loss": 0.1229,
+      "theoretical_loss": 4.552460395494878,
+      "tokens_seen": 148635648
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009645321778205746,
+      "loss": 0.1247,
+      "theoretical_loss": 4.55141708059017,
+      "tokens_seen": 148897792
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009644519338789921,
+      "loss": 0.1208,
+      "theoretical_loss": 4.5503761141724555,
+      "tokens_seen": 149159936
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009643716899374097,
+      "loss": 0.1224,
+      "theoretical_loss": 4.549337486841843,
+      "tokens_seen": 149422080
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009642914459958273,
+      "loss": 0.1231,
+      "theoretical_loss": 4.548301189252473,
+      "tokens_seen": 149684224
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000964211202054245,
+      "loss": 0.124,
+      "theoretical_loss": 4.54726721211211,
+      "tokens_seen": 149946368
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009641309581126625,
+      "loss": 0.1215,
+      "theoretical_loss": 4.546235546181743,
+      "tokens_seen": 150208512
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009640507141710802,
+      "loss": 0.1179,
+      "theoretical_loss": 4.545206182275189,
+      "tokens_seen": 150470656
+    },
+    {
+      "epoch": 0.05,
+      "objective/train/advantage_avg": 0.004590987227857113,
+      "objective/train/docs_used": 61752,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.4665567874908447,
+      "objective/train/original_loss": 2.466557025909424,
+      "objective/train/theoretical_loss": 4.5441791112587016,
+      "objective/train/tokens_used": 171192800,
+      "objective/train/value_avg": -0.01053619384765625,
+      "objective/train/value_loss": 0.0002645233762450516,
+      "objective/train/value_max": -0.00133514404296875,
+      "objective/train/value_min": -0.203369140625,
+      "objective/train/value_reward_corr": 0.45074871885232737,
+      "objective/train/value_std": 0.0080718994140625,
+      "objective/train/weight_avg": 1.004720687866211,
+      "objective/train/weighted_lm_loss": 2.478607177734375,
+      "objective/train/weights_max": 1.1166024208068848,
+      "objective/train/weights_min": 0.6107982993125916,
+      "theoretical_loss": 4.5441791112587016,
+      "tokens_seen": 150732800
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009639704702294977,
+      "loss": 0.1199,
+      "theoretical_loss": 4.5441791112587016,
+      "tokens_seen": 150732800
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009638902262879152,
+      "loss": 0.1258,
+      "theoretical_loss": 4.5431543240505725,
+      "tokens_seen": 150994944
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009638099823463329,
+      "loss": 0.123,
+      "theoretical_loss": 4.5421318116207585,
+      "tokens_seen": 151257088
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009637297384047504,
+      "loss": 0.1184,
+      "theoretical_loss": 4.541111564990485,
+      "tokens_seen": 151519232
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009636494944631681,
+      "loss": 0.1201,
+      "theoretical_loss": 4.540093575231879,
+      "tokens_seen": 151781376
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009635692505215856,
+      "loss": 0.1185,
+      "theoretical_loss": 4.539077833467582,
+      "tokens_seen": 152043520
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009634890065800032,
+      "loss": 0.12,
+      "theoretical_loss": 4.538064330870389,
+      "tokens_seen": 152305664
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009634087626384209,
+      "loss": 0.1218,
+      "theoretical_loss": 4.537053058662869,
+      "tokens_seen": 152567808
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009633285186968385,
+      "loss": 0.1205,
+      "theoretical_loss": 4.536044008117005,
+      "tokens_seen": 152829952
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000963248274755256,
+      "loss": 0.124,
+      "theoretical_loss": 4.535037170553833,
+      "tokens_seen": 153092096
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009631680308136736,
+      "loss": 0.1195,
+      "theoretical_loss": 4.534032537343078,
+      "tokens_seen": 153354240
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009630877868720912,
+      "loss": 0.1228,
+      "theoretical_loss": 4.533030099902803,
+      "tokens_seen": 153616384
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009630075429305087,
+      "loss": 0.1237,
+      "theoretical_loss": 4.53202984969905,
+      "tokens_seen": 153878528
+    },
+    {
+      "epoch": 0.05,
+      "objective/train/advantage_avg": 0.004563726019114256,
+      "objective/train/docs_used": 63027,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.6050822734832764,
+      "objective/train/original_loss": 2.6050822734832764,
+      "objective/train/theoretical_loss": 4.531530542157043,
+      "objective/train/tokens_used": 174469600,
+      "objective/train/value_avg": -0.00949859619140625,
+      "objective/train/value_loss": 0.0006905548507347703,
+      "objective/train/value_max": -0.0010986328125,
+      "objective/train/value_min": -0.29541015625,
+      "objective/train/value_reward_corr": 0.3721653551526731,
+      "objective/train/value_std": 0.00933074951171875,
+      "objective/train/weight_avg": 1.004849910736084,
+      "objective/train/weighted_lm_loss": 2.618398904800415,
+      "objective/train/weights_max": 1.2023754119873047,
+      "objective/train/weights_min": 0.36947569251060486,
+      "theoretical_loss": 4.531530542157043,
+      "tokens_seen": 154009600
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009629272989889264,
+      "loss": 0.1251,
+      "theoretical_loss": 4.531031778245499,
+      "tokens_seen": 154140672
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009628470550473439,
+      "loss": 0.1225,
+      "theoretical_loss": 4.530035877103115,
+      "tokens_seen": 154402816
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009627668111057615,
+      "loss": 0.1218,
+      "theoretical_loss": 4.529042137879809,
+      "tokens_seen": 154664960
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009626865671641792,
+      "loss": 0.1192,
+      "theoretical_loss": 4.528050552230092,
+      "tokens_seen": 154927104
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009626063232225967,
+      "loss": 0.1212,
+      "theoretical_loss": 4.527061111854746,
+      "tokens_seen": 155189248
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009625260792810143,
+      "loss": 0.1221,
+      "theoretical_loss": 4.526073808500481,
+      "tokens_seen": 155451392
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009624458353394319,
+      "loss": 0.1189,
+      "theoretical_loss": 4.525088633959613,
+      "tokens_seen": 155713536
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009623655913978495,
+      "loss": 0.1184,
+      "theoretical_loss": 4.524105580069728,
+      "tokens_seen": 155975680
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009622853474562671,
+      "loss": 0.1185,
+      "theoretical_loss": 4.523124638713361,
+      "tokens_seen": 156237824
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009622051035146847,
+      "loss": 0.1208,
+      "theoretical_loss": 4.522145801817673,
+      "tokens_seen": 156499968
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009621248595731022,
+      "loss": 0.1158,
+      "theoretical_loss": 4.521169061354129,
+      "tokens_seen": 156762112
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009620446156315198,
+      "loss": 0.1177,
+      "theoretical_loss": 4.520194409338185,
+      "tokens_seen": 157024256
+    },
+    {
+      "epoch": 0.05,
+      "objective/train/advantage_avg": 0.004033830948174,
+      "objective/train/docs_used": 64097,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.5532474517822266,
+      "objective/train/original_loss": 2.5532474517822266,
+      "objective/train/theoretical_loss": 4.519221837828971,
+      "objective/train/tokens_used": 177746400,
+      "objective/train/value_avg": -0.01045989990234375,
+      "objective/train/value_loss": 0.00029761483892798424,
+      "objective/train/value_max": -0.0013408660888671875,
+      "objective/train/value_min": -0.1632080078125,
+      "objective/train/value_reward_corr": 0.47638880122386174,
+      "objective/train/value_std": 0.006755828857421875,
+      "objective/train/weight_avg": 1.0041792392730713,
+      "objective/train/weighted_lm_loss": 2.564786434173584,
+      "objective/train/weights_max": 1.0706846714019775,
+      "objective/train/weights_min": 0.7207310795783997,
+      "theoretical_loss": 4.519221837828971,
+      "tokens_seen": 157286400
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009619643716899374,
+      "loss": 0.1201,
+      "theoretical_loss": 4.519221837828971,
+      "tokens_seen": 157286400
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000961884127748355,
+      "loss": 0.1197,
+      "theoretical_loss": 4.51825133892898,
+      "tokens_seen": 157548544
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009618038838067727,
+      "loss": 0.1175,
+      "theoretical_loss": 4.517282904783764,
+      "tokens_seen": 157810688
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009617236398651902,
+      "loss": 0.1204,
+      "theoretical_loss": 4.516316527581621,
+      "tokens_seen": 158072832
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009616433959236078,
+      "loss": 0.1225,
+      "theoretical_loss": 4.515352199553295,
+      "tokens_seen": 158334976
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009615631519820254,
+      "loss": 0.1203,
+      "theoretical_loss": 4.514389912971679,
+      "tokens_seen": 158597120
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009614829080404429,
+      "loss": 0.1203,
+      "theoretical_loss": 4.513429660151513,
+      "tokens_seen": 158859264
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009614026640988605,
+      "loss": 0.1171,
+      "theoretical_loss": 4.51247143344909,
+      "tokens_seen": 159121408
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009613224201572781,
+      "loss": 0.12,
+      "theoretical_loss": 4.511515225261961,
+      "tokens_seen": 159383552
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009612421762156957,
+      "loss": 0.1183,
+      "theoretical_loss": 4.5105610280286506,
+      "tokens_seen": 159645696
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009611619322741134,
+      "loss": 0.1196,
+      "theoretical_loss": 4.509608834228365,
+      "tokens_seen": 159907840
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000961081688332531,
+      "loss": 0.1155,
+      "theoretical_loss": 4.508658636380705,
+      "tokens_seen": 160169984
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009610014443909485,
+      "loss": 0.1206,
+      "theoretical_loss": 4.507710427045389,
+      "tokens_seen": 160432128
+    },
+    {
+      "epoch": 0.05,
+      "objective/train/advantage_avg": 0.0053785815834999084,
+      "objective/train/docs_used": 65316,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.3940212726593018,
+      "objective/train/original_loss": 2.3940210342407227,
+      "objective/train/theoretical_loss": 4.507237065755672,
+      "objective/train/tokens_used": 181023200,
+      "objective/train/value_avg": -0.00850677490234375,
+      "objective/train/value_loss": 0.00017240087618120015,
+      "objective/train/value_max": -0.0016107559204101562,
+      "objective/train/value_min": -0.1729736328125,
+      "objective/train/value_reward_corr": 0.16862680553399217,
+      "objective/train/value_std": 0.005374908447265625,
+      "objective/train/weight_avg": 1.005456805229187,
+      "objective/train/weighted_lm_loss": 2.4071598052978516,
+      "objective/train/weights_max": 1.0916688442230225,
+      "objective/train/weights_min": 0.370330810546875,
+      "theoretical_loss": 4.507237065755672,
+      "tokens_seen": 160563200
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009609212004493661,
+      "loss": 0.1181,
+      "theoretical_loss": 4.50676419882197,
+      "tokens_seen": 160694272
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009608409565077837,
+      "loss": 0.1183,
+      "theoretical_loss": 4.505819944349556,
+      "tokens_seen": 160956416
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009607607125662012,
+      "loss": 0.1155,
+      "theoretical_loss": 4.504877656306535,
+      "tokens_seen": 161218560
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009606804686246189,
+      "loss": 0.1189,
+      "theoretical_loss": 4.503937327410306,
+      "tokens_seen": 161480704
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009606002246830364,
+      "loss": 0.1185,
+      "theoretical_loss": 4.502998950417004,
+      "tokens_seen": 161742848
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000960519980741454,
+      "loss": 0.1188,
+      "theoretical_loss": 4.502062518121232,
+      "tokens_seen": 162004992
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009604397367998717,
+      "loss": 0.1163,
+      "theoretical_loss": 4.501128023355796,
+      "tokens_seen": 162267136
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009603594928582893,
+      "loss": 0.1164,
+      "theoretical_loss": 4.500195458991443,
+      "tokens_seen": 162529280
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009602792489167068,
+      "loss": 0.1192,
+      "theoretical_loss": 4.499264817936593,
+      "tokens_seen": 162791424
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009601990049751244,
+      "loss": 0.1209,
+      "theoretical_loss": 4.498336093137089,
+      "tokens_seen": 163053568
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000960118761033542,
+      "loss": 0.1193,
+      "theoretical_loss": 4.49740927757593,
+      "tokens_seen": 163315712
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009600385170919595,
+      "loss": 0.1164,
+      "theoretical_loss": 4.496484364273021,
+      "tokens_seen": 163577856
+    },
+    {
+      "epoch": 0.05,
+      "objective/train/advantage_avg": 0.00311545399017632,
+      "objective/train/docs_used": 66569,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.244389295578003,
+      "objective/train/original_loss": 2.244389533996582,
+      "objective/train/theoretical_loss": 4.495561346284925,
+      "objective/train/tokens_used": 184300000,
+      "objective/train/value_avg": -0.00873565673828125,
+      "objective/train/value_loss": 0.00021333472977858037,
+      "objective/train/value_max": -0.0010862350463867188,
+      "objective/train/value_min": -0.260009765625,
+      "objective/train/value_reward_corr": 0.3512147373222607,
+      "objective/train/value_std": 0.00701904296875,
+      "objective/train/weight_avg": 1.0032168626785278,
+      "objective/train/weighted_lm_loss": 2.2516353130340576,
+      "objective/train/weights_max": 1.1541095972061157,
+      "objective/train/weights_min": 0.3705597221851349,
+      "theoretical_loss": 4.495561346284925,
+      "tokens_seen": 163840000
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009599582731503772,
+      "loss": 0.1161,
+      "theoretical_loss": 4.495561346284925,
+      "tokens_seen": 163840000
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009598780292087947,
+      "loss": 0.1176,
+      "theoretical_loss": 4.494640216704598,
+      "tokens_seen": 164102144
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009597977852672124,
+      "loss": 0.1193,
+      "theoretical_loss": 4.493720968661158,
+      "tokens_seen": 164364288
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.00095971754132563,
+      "loss": 0.116,
+      "theoretical_loss": 4.492803595319623,
+      "tokens_seen": 164626432
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009596372973840475,
+      "loss": 0.1149,
+      "theoretical_loss": 4.49188808988068,
+      "tokens_seen": 164888576
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009595570534424652,
+      "loss": 0.1118,
+      "theoretical_loss": 4.490974445580429,
+      "tokens_seen": 165150720
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009594768095008827,
+      "loss": 0.113,
+      "theoretical_loss": 4.490062655690153,
+      "tokens_seen": 165412864
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009593965655593003,
+      "loss": 0.113,
+      "theoretical_loss": 4.489152713516077,
+      "tokens_seen": 165675008
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009593163216177179,
+      "loss": 0.1102,
+      "theoretical_loss": 4.488244612399129,
+      "tokens_seen": 165937152
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009592360776761355,
+      "loss": 0.1119,
+      "theoretical_loss": 4.487338345714707,
+      "tokens_seen": 166199296
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000959155833734553,
+      "loss": 0.1138,
+      "theoretical_loss": 4.486433906872448,
+      "tokens_seen": 166461440
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009590755897929706,
+      "loss": 0.1152,
+      "theoretical_loss": 4.485531289315997,
+      "tokens_seen": 166723584
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009589953458513882,
+      "loss": 0.1099,
+      "theoretical_loss": 4.484630486522775,
+      "tokens_seen": 166985728
+    },
+    {
+      "epoch": 0.05,
+      "objective/train/advantage_avg": 0.002676489297300577,
+      "objective/train/docs_used": 67770,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.380342483520508,
+      "objective/train/original_loss": 2.380342483520508,
+      "objective/train/theoretical_loss": 4.48418076363325,
+      "objective/train/tokens_used": 187576800,
+      "objective/train/value_avg": -0.01169586181640625,
+      "objective/train/value_loss": 0.0005142788286320865,
+      "objective/train/value_max": -0.0011920928955078125,
+      "objective/train/value_min": -0.332763671875,
+      "objective/train/value_reward_corr": 0.49718080147400245,
+      "objective/train/value_std": 0.01336669921875,
+      "objective/train/weight_avg": 1.0029011964797974,
+      "objective/train/weighted_lm_loss": 2.385775327682495,
+      "objective/train/weights_max": 1.2041622400283813,
+      "objective/train/weights_min": 0.39096081256866455,
+      "theoretical_loss": 4.48418076363325,
+      "tokens_seen": 167116800
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009589151019098058,
+      "loss": 0.117,
+      "theoretical_loss": 4.483731492003757,
+      "tokens_seen": 167247872
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009588348579682235,
+      "loss": 0.113,
+      "theoretical_loss": 4.482834299303246,
+      "tokens_seen": 167510016
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000958754614026641,
+      "loss": 0.1132,
+      "theoretical_loss": 4.481938901998647,
+      "tokens_seen": 167772160
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009586743700850587,
+      "loss": 0.1124,
+      "theoretical_loss": 4.481045293700248,
+      "tokens_seen": 168034304
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009585941261434762,
+      "loss": 0.1135,
+      "theoretical_loss": 4.480153468051001,
+      "tokens_seen": 168296448
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009585138822018937,
+      "loss": 0.1152,
+      "theoretical_loss": 4.4792634187263065,
+      "tokens_seen": 168558592
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009584336382603114,
+      "loss": 0.1142,
+      "theoretical_loss": 4.4783751394337905,
+      "tokens_seen": 168820736
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009583533943187289,
+      "loss": 0.1093,
+      "theoretical_loss": 4.4774886239131,
+      "tokens_seen": 169082880
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009582731503771465,
+      "loss": 0.116,
+      "theoretical_loss": 4.476603865935683,
+      "tokens_seen": 169345024
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009581929064355642,
+      "loss": 0.1161,
+      "theoretical_loss": 4.475720859304583,
+      "tokens_seen": 169607168
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009581126624939818,
+      "loss": 0.1115,
+      "theoretical_loss": 4.474839597854226,
+      "tokens_seen": 169869312
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009580324185523993,
+      "loss": 0.1149,
+      "theoretical_loss": 4.473960075450218,
+      "tokens_seen": 170131456
+    },
+    {
+      "epoch": 0.05,
+      "objective/train/advantage_avg": 0.005700128152966499,
+      "objective/train/docs_used": 69022,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.486732006072998,
+      "objective/train/original_loss": 2.486732006072998,
+      "objective/train/theoretical_loss": 4.473082285989134,
+      "objective/train/tokens_used": 190853600,
+      "objective/train/value_avg": -0.0110015869140625,
+      "objective/train/value_loss": 0.0002670148969627917,
+      "objective/train/value_max": -0.0014438629150390625,
+      "objective/train/value_min": -0.30224609375,
+      "objective/train/value_reward_corr": 0.4285667679812899,
+      "objective/train/value_std": 0.01032257080078125,
+      "objective/train/weight_avg": 1.0058281421661377,
+      "objective/train/weighted_lm_loss": 2.500220775604248,
+      "objective/train/weights_max": 1.1895605325698853,
+      "objective/train/weights_min": 0.3700695335865021,
+      "theoretical_loss": 4.473082285989134,
+      "tokens_seen": 170393600
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009579521746108169,
+      "loss": 0.118,
+      "theoretical_loss": 4.473082285989134,
+      "tokens_seen": 170393600
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009578719306692345,
+      "loss": 0.1144,
+      "theoretical_loss": 4.472206223398325,
+      "tokens_seen": 170655744
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000957791686727652,
+      "loss": 0.1134,
+      "theoretical_loss": 4.471331881635698,
+      "tokens_seen": 170917888
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009577114427860697,
+      "loss": 0.113,
+      "theoretical_loss": 4.470459254689533,
+      "tokens_seen": 171180032
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009576311988444872,
+      "loss": 0.1129,
+      "theoretical_loss": 4.469588336578277,
+      "tokens_seen": 171442176
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009575509549029048,
+      "loss": 0.1111,
+      "theoretical_loss": 4.468719121350343,
+      "tokens_seen": 171704320
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009574707109613225,
+      "loss": 0.1102,
+      "theoretical_loss": 4.467851603083923,
+      "tokens_seen": 171966464
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.00095739046701974,
+      "loss": 0.1133,
+      "theoretical_loss": 4.466985775886784,
+      "tokens_seen": 172228608
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009573102230781577,
+      "loss": 0.1121,
+      "theoretical_loss": 4.466121633896087,
+      "tokens_seen": 172490752
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009572299791365752,
+      "loss": 0.1152,
+      "theoretical_loss": 4.465259171278182,
+      "tokens_seen": 172752896
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009571497351949928,
+      "loss": 0.1114,
+      "theoretical_loss": 4.464398382228435,
+      "tokens_seen": 173015040
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009570694912534104,
+      "loss": 0.1146,
+      "theoretical_loss": 4.463539260971023,
+      "tokens_seen": 173277184
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000956989247311828,
+      "loss": 0.1145,
+      "theoretical_loss": 4.462681801758762,
+      "tokens_seen": 173539328
+    },
+    {
+      "epoch": 0.05,
+      "objective/train/advantage_avg": 0.0046385833993554115,
+      "objective/train/docs_used": 70257,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.40885853767395,
+      "objective/train/original_loss": 2.408858299255371,
+      "objective/train/theoretical_loss": 4.462253693631352,
+      "objective/train/tokens_used": 194130400,
+      "objective/train/value_avg": -0.0102081298828125,
+      "objective/train/value_loss": 0.00034067677916027606,
+      "objective/train/value_max": -0.0015306472778320312,
+      "objective/train/value_min": -0.171630859375,
+      "objective/train/value_reward_corr": 0.4651631801087514,
+      "objective/train/value_std": 0.008270263671875,
+      "objective/train/weight_avg": 1.0047951936721802,
+      "objective/train/weighted_lm_loss": 2.420976161956787,
+      "objective/train/weights_max": 1.1363425254821777,
+      "objective/train/weights_min": 0.5437347888946533,
+      "theoretical_loss": 4.462253693631352,
+      "tokens_seen": 173670400
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009569090033702455,
+      "loss": 0.1156,
+      "theoretical_loss": 4.461825998872914,
+      "tokens_seen": 173801472
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009568287594286632,
+      "loss": 0.1096,
+      "theoretical_loss": 4.460971846623005,
+      "tokens_seen": 174063616
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009567485154870808,
+      "loss": 0.1121,
+      "theoretical_loss": 4.460119339346643,
+      "tokens_seen": 174325760
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009566682715454983,
+      "loss": 0.1099,
+      "theoretical_loss": 4.45926847140934,
+      "tokens_seen": 174587904
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000956588027603916,
+      "loss": 0.1113,
+      "theoretical_loss": 4.45841923720433,
+      "tokens_seen": 174850048
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009565077836623335,
+      "loss": 0.1171,
+      "theoretical_loss": 4.4575716311523905,
+      "tokens_seen": 175112192
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009564275397207511,
+      "loss": 0.1131,
+      "theoretical_loss": 4.456725647701669,
+      "tokens_seen": 175374336
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009563472957791687,
+      "loss": 0.1123,
+      "theoretical_loss": 4.455881281327508,
+      "tokens_seen": 175636480
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009562670518375863,
+      "loss": 0.1149,
+      "theoretical_loss": 4.4550385265322685,
+      "tokens_seen": 175898624
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009561868078960038,
+      "loss": 0.1116,
+      "theoretical_loss": 4.45419737784516,
+      "tokens_seen": 176160768
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009561065639544214,
+      "loss": 0.1141,
+      "theoretical_loss": 4.45335782982207,
+      "tokens_seen": 176422912
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000956026320012839,
+      "loss": 0.1142,
+      "theoretical_loss": 4.452519877045393,
+      "tokens_seen": 176685056
+    },
+    {
+      "epoch": 0.05,
+      "objective/train/advantage_avg": 0.004688139073550701,
+      "objective/train/docs_used": 71338,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.032062292098999,
+      "objective/train/original_loss": 2.032062292098999,
+      "objective/train/theoretical_loss": 4.451683514123864,
+      "objective/train/tokens_used": 197407200,
+      "objective/train/value_avg": -0.01114654541015625,
+      "objective/train/value_loss": 0.00021585206559393555,
+      "objective/train/value_max": -0.00131988525390625,
+      "objective/train/value_min": -0.267822265625,
+      "objective/train/value_reward_corr": 0.5101379035115723,
+      "objective/train/value_std": 0.008941650390625,
+      "objective/train/weight_avg": 1.004790186882019,
+      "objective/train/weighted_lm_loss": 2.041867971420288,
+      "objective/train/weights_max": 1.1229585409164429,
+      "objective/train/weights_min": 0.36935028433799744,
+      "theoretical_loss": 4.451683514123864,
+      "tokens_seen": 176947200
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009559460760712567,
+      "loss": 0.1141,
+      "theoretical_loss": 4.451683514123864,
+      "tokens_seen": 176947200
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009558658321296743,
+      "loss": 0.1138,
+      "theoretical_loss": 4.450848735692391,
+      "tokens_seen": 177209344
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009557855881880918,
+      "loss": 0.1133,
+      "theoretical_loss": 4.450015536411886,
+      "tokens_seen": 177471488
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009557053442465095,
+      "loss": 0.1106,
+      "theoretical_loss": 4.449183910969108,
+      "tokens_seen": 177733632
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000955625100304927,
+      "loss": 0.1089,
+      "theoretical_loss": 4.448353854076494,
+      "tokens_seen": 177995776
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009555448563633445,
+      "loss": 0.1118,
+      "theoretical_loss": 4.4475253604719995,
+      "tokens_seen": 178257920
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009554646124217622,
+      "loss": 0.1077,
+      "theoretical_loss": 4.446698424918937,
+      "tokens_seen": 178520064
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009553843684801797,
+      "loss": 0.1118,
+      "theoretical_loss": 4.44587304220582,
+      "tokens_seen": 178782208
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009553041245385973,
+      "loss": 0.1074,
+      "theoretical_loss": 4.4450492071462,
+      "tokens_seen": 179044352
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000955223880597015,
+      "loss": 0.1145,
+      "theoretical_loss": 4.444226914578513,
+      "tokens_seen": 179306496
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009551436366554326,
+      "loss": 0.1149,
+      "theoretical_loss": 4.4434061593659235,
+      "tokens_seen": 179568640
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009550633927138501,
+      "loss": 0.116,
+      "theoretical_loss": 4.442586936396171,
+      "tokens_seen": 179830784
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009549831487722677,
+      "loss": 0.1145,
+      "theoretical_loss": 4.441769240581412,
+      "tokens_seen": 180092928
+    },
+    {
+      "epoch": 0.05,
+      "objective/train/advantage_avg": 0.0043923812918365,
+      "objective/train/docs_used": 72550,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.181901216506958,
+      "objective/train/original_loss": 2.181901454925537,
+      "objective/train/theoretical_loss": 4.441360963773846,
+      "objective/train/tokens_used": 200684000,
+      "objective/train/value_avg": -0.010162353515625,
+      "objective/train/value_loss": 0.00025610229931771755,
+      "objective/train/value_max": -0.0012645721435546875,
+      "objective/train/value_min": -0.2271728515625,
+      "objective/train/value_reward_corr": 0.5822503996716603,
+      "objective/train/value_std": 0.00917816162109375,
+      "objective/train/weight_avg": 1.0045078992843628,
+      "objective/train/weighted_lm_loss": 2.191445827484131,
+      "objective/train/weights_max": 1.0695090293884277,
+      "objective/train/weights_min": 0.3694573640823364,
+      "theoretical_loss": 4.441360963773846,
+      "tokens_seen": 180224000
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009549029048306853,
+      "loss": 0.1193,
+      "theoretical_loss": 4.440953066858077,
+      "tokens_seen": 180355072
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009548226608891028,
+      "loss": 0.1133,
+      "theoretical_loss": 4.4401384101867105,
+      "tokens_seen": 180617216
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009547424169475205,
+      "loss": 0.1127,
+      "theoretical_loss": 4.439325265551826,
+      "tokens_seen": 180879360
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.000954662173005938,
+      "loss": 0.1135,
+      "theoretical_loss": 4.438513627961757,
+      "tokens_seen": 181141504
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 0.0009545819290643557,
+      "loss": 0.1131,
+      "theoretical_loss": 4.437703492448509,
+      "tokens_seen": 181403648
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009545016851227733,
+      "loss": 0.1149,
+      "theoretical_loss": 4.436894854067614,
+      "tokens_seen": 181665792
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009544214411811908,
+      "loss": 0.1132,
+      "theoretical_loss": 4.436087707897984,
+      "tokens_seen": 181927936
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009543411972396085,
+      "loss": 0.1131,
+      "theoretical_loss": 4.435282049041769,
+      "tokens_seen": 182190080
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000954260953298026,
+      "loss": 0.1115,
+      "theoretical_loss": 4.434477872624212,
+      "tokens_seen": 182452224
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009541807093564436,
+      "loss": 0.1139,
+      "theoretical_loss": 4.433675173793507,
+      "tokens_seen": 182714368
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009541004654148612,
+      "loss": 0.1125,
+      "theoretical_loss": 4.43287394772066,
+      "tokens_seen": 182976512
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009540202214732788,
+      "loss": 0.1156,
+      "theoretical_loss": 4.43207418959935,
+      "tokens_seen": 183238656
+    },
+    {
+      "epoch": 0.06,
+      "objective/train/advantage_avg": -0.0008087606402114034,
+      "objective/train/docs_used": 73723,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.2917051315307617,
+      "objective/train/original_loss": 2.2917051315307617,
+      "objective/train/theoretical_loss": 4.431275894645784,
+      "objective/train/tokens_used": 203960800,
+      "objective/train/value_avg": -0.01041412353515625,
+      "objective/train/value_loss": 0.0009768769377842546,
+      "objective/train/value_max": -0.0010118484497070312,
+      "objective/train/value_min": -0.2391357421875,
+      "objective/train/value_reward_corr": 0.8343898472090797,
+      "objective/train/value_std": 0.0133514404296875,
+      "objective/train/weight_avg": 0.9996550679206848,
+      "objective/train/weighted_lm_loss": 2.28629732131958,
+      "objective/train/weights_max": 1.0813897848129272,
+      "objective/train/weights_min": 0.3707435429096222,
+      "theoretical_loss": 4.431275894645784,
+      "tokens_seen": 183500800
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009539399775316963,
+      "loss": 0.1081,
+      "theoretical_loss": 4.431275894645784,
+      "tokens_seen": 183500800
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000953859733590114,
+      "loss": 0.1118,
+      "theoretical_loss": 4.43047905809857,
+      "tokens_seen": 183762944
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009537794896485315,
+      "loss": 0.1102,
+      "theoretical_loss": 4.42968367521857,
+      "tokens_seen": 184025088
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000953699245706949,
+      "loss": 0.1121,
+      "theoretical_loss": 4.428889741288771,
+      "tokens_seen": 184287232
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009536190017653668,
+      "loss": 0.1135,
+      "theoretical_loss": 4.428097251614145,
+      "tokens_seen": 184549376
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009535387578237843,
+      "loss": 0.1115,
+      "theoretical_loss": 4.427306201521524,
+      "tokens_seen": 184811520
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000953458513882202,
+      "loss": 0.1107,
+      "theoretical_loss": 4.426516586359458,
+      "tokens_seen": 185073664
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009533782699406195,
+      "loss": 0.1126,
+      "theoretical_loss": 4.425728401498089,
+      "tokens_seen": 185335808
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009532980259990371,
+      "loss": 0.1136,
+      "theoretical_loss": 4.424941642329019,
+      "tokens_seen": 185597952
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009532177820574547,
+      "loss": 0.1118,
+      "theoretical_loss": 4.42415630426518,
+      "tokens_seen": 185860096
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009531375381158722,
+      "loss": 0.1099,
+      "theoretical_loss": 4.423372382740707,
+      "tokens_seen": 186122240
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009530572941742898,
+      "loss": 0.1075,
+      "theoretical_loss": 4.422589873210806,
+      "tokens_seen": 186384384
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009529770502327075,
+      "loss": 0.1079,
+      "theoretical_loss": 4.4218087711516345,
+      "tokens_seen": 186646528
+    },
+    {
+      "epoch": 0.06,
+      "objective/train/advantage_avg": 0.0038929590955376625,
+      "objective/train/docs_used": 74904,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.182293653488159,
+      "objective/train/original_loss": 2.18229341506958,
+      "objective/train/theoretical_loss": 4.421418746515567,
+      "objective/train/tokens_used": 207237600,
+      "objective/train/value_avg": -0.01084136962890625,
+      "objective/train/value_loss": 0.0002643898769747466,
+      "objective/train/value_max": -0.0011692047119140625,
+      "objective/train/value_min": -0.2230224609375,
+      "objective/train/value_reward_corr": 0.4553354681669406,
+      "objective/train/value_std": 0.0082855224609375,
+      "objective/train/weight_avg": 1.0040152072906494,
+      "objective/train/weighted_lm_loss": 2.1902849674224854,
+      "objective/train/weights_max": 1.0839669704437256,
+      "objective/train/weights_min": 0.3749292194843292,
+      "theoretical_loss": 4.421418746515567,
+      "tokens_seen": 186777600
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009528968062911251,
+      "loss": 0.1131,
+      "theoretical_loss": 4.421029072060167,
+      "tokens_seen": 186908672
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009528165623495426,
+      "loss": 0.1097,
+      "theoretical_loss": 4.420250771454078,
+      "tokens_seen": 187170816
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009527363184079603,
+      "loss": 0.1131,
+      "theoretical_loss": 4.419473864871613,
+      "tokens_seen": 187432960
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009526560744663778,
+      "loss": 0.1132,
+      "theoretical_loss": 4.4186983478714685,
+      "tokens_seen": 187695104
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009525758305247953,
+      "loss": 0.1127,
+      "theoretical_loss": 4.417924216032667,
+      "tokens_seen": 187957248
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000952495586583213,
+      "loss": 0.111,
+      "theoretical_loss": 4.417151464954437,
+      "tokens_seen": 188219392
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009524153426416305,
+      "loss": 0.1102,
+      "theoretical_loss": 4.416380090256095,
+      "tokens_seen": 188481536
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009523350987000481,
+      "loss": 0.1144,
+      "theoretical_loss": 4.415610087576923,
+      "tokens_seen": 188743680
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009522548547584658,
+      "loss": 0.1116,
+      "theoretical_loss": 4.414841452576049,
+      "tokens_seen": 189005824
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009521746108168834,
+      "loss": 0.1083,
+      "theoretical_loss": 4.414074180932333,
+      "tokens_seen": 189267968
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000952094366875301,
+      "loss": 0.1086,
+      "theoretical_loss": 4.413308268344249,
+      "tokens_seen": 189530112
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009520141229337185,
+      "loss": 0.1115,
+      "theoretical_loss": 4.412543710529766,
+      "tokens_seen": 189792256
+    },
+    {
+      "epoch": 0.06,
+      "objective/train/advantage_avg": 0.0025555319152772427,
+      "objective/train/docs_used": 76077,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.207807779312134,
+      "objective/train/original_loss": 2.207808017730713,
+      "objective/train/theoretical_loss": 4.411780503226238,
+      "objective/train/tokens_used": 210514400,
+      "objective/train/value_avg": -0.0090789794921875,
+      "objective/train/value_loss": 0.0002816997584886849,
+      "objective/train/value_max": -0.0007886886596679688,
+      "objective/train/value_min": -0.25732421875,
+      "objective/train/value_reward_corr": 0.4408398455844697,
+      "objective/train/value_std": 0.0070037841796875,
+      "objective/train/weight_avg": 1.0026865005493164,
+      "objective/train/weighted_lm_loss": 2.213961362838745,
+      "objective/train/weights_max": 1.1967458724975586,
+      "objective/train/weights_min": 0.38571086525917053,
+      "theoretical_loss": 4.411780503226238,
+      "tokens_seen": 190054400
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009519338789921361,
+      "loss": 0.111,
+      "theoretical_loss": 4.411780503226238,
+      "tokens_seen": 190054400
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009518536350505537,
+      "loss": 0.1131,
+      "theoretical_loss": 4.4110186421902835,
+      "tokens_seen": 190316544
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009517733911089713,
+      "loss": 0.1077,
+      "theoretical_loss": 4.4102581231976785,
+      "tokens_seen": 190578688
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009516931471673888,
+      "loss": 0.1121,
+      "theoretical_loss": 4.409498942043237,
+      "tokens_seen": 190840832
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009516129032258065,
+      "loss": 0.1111,
+      "theoretical_loss": 4.408741094540707,
+      "tokens_seen": 191102976
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000951532659284224,
+      "loss": 0.1087,
+      "theoretical_loss": 4.407984576522653,
+      "tokens_seen": 191365120
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009514524153426416,
+      "loss": 0.111,
+      "theoretical_loss": 4.407229383840347,
+      "tokens_seen": 191627264
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009513721714010593,
+      "loss": 0.1098,
+      "theoretical_loss": 4.406475512363663,
+      "tokens_seen": 191889408
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009512919274594768,
+      "loss": 0.1062,
+      "theoretical_loss": 4.405722957980962,
+      "tokens_seen": 192151552
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009512116835178944,
+      "loss": 0.1097,
+      "theoretical_loss": 4.40497171659899,
+      "tokens_seen": 192413696
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000951131439576312,
+      "loss": 0.1064,
+      "theoretical_loss": 4.404221784142768,
+      "tokens_seen": 192675840
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009510511956347296,
+      "loss": 0.1104,
+      "theoretical_loss": 4.403473156555487,
+      "tokens_seen": 192937984
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009509709516931472,
+      "loss": 0.1077,
+      "theoretical_loss": 4.402725829798397,
+      "tokens_seen": 193200128
+    },
+    {
+      "epoch": 0.06,
+      "objective/train/advantage_avg": 0.0021290904842317104,
+      "objective/train/docs_used": 77266,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.261606454849243,
+      "objective/train/original_loss": 2.261606454849243,
+      "objective/train/theoretical_loss": 4.40235265297398,
+      "objective/train/tokens_used": 213791200,
+      "objective/train/value_avg": -0.00881195068359375,
+      "objective/train/value_loss": 0.00027175722061656415,
+      "objective/train/value_max": -0.0007319450378417969,
+      "objective/train/value_min": -0.278076171875,
+      "objective/train/value_reward_corr": 0.45116222074040196,
+      "objective/train/value_std": 0.00730133056640625,
+      "objective/train/weight_avg": 1.0022554397583008,
+      "objective/train/weighted_lm_loss": 2.266122341156006,
+      "objective/train/weights_max": 1.106561303138733,
+      "objective/train/weights_min": 0.36955323815345764,
+      "theoretical_loss": 4.40235265297398,
+      "tokens_seen": 193331200
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009508907077515648,
+      "loss": 0.1074,
+      "theoretical_loss": 4.4019797998507135,
+      "tokens_seen": 193462272
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009508104638099823,
+      "loss": 0.1048,
+      "theoretical_loss": 4.401235062709502,
+      "tokens_seen": 193724416
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009507302198684,
+      "loss": 0.1085,
+      "theoretical_loss": 4.400491614389582,
+      "tokens_seen": 193986560
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009506499759268176,
+      "loss": 0.1047,
+      "theoretical_loss": 4.3997494509234185,
+      "tokens_seen": 194248704
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009505697319852351,
+      "loss": 0.108,
+      "theoretical_loss": 4.399008568361027,
+      "tokens_seen": 194510848
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009504894880436528,
+      "loss": 0.11,
+      "theoretical_loss": 4.398268962769867,
+      "tokens_seen": 194772992
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009504092441020703,
+      "loss": 0.1121,
+      "theoretical_loss": 4.397530630234744,
+      "tokens_seen": 195035136
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009503290001604879,
+      "loss": 0.1068,
+      "theoretical_loss": 4.396793566857708,
+      "tokens_seen": 195297280
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009502487562189055,
+      "loss": 0.1059,
+      "theoretical_loss": 4.396057768757957,
+      "tokens_seen": 195559424
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000950168512277323,
+      "loss": 0.1079,
+      "theoretical_loss": 4.395323232071737,
+      "tokens_seen": 195821568
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009500882683357406,
+      "loss": 0.1078,
+      "theoretical_loss": 4.394589952952247,
+      "tokens_seen": 196083712
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009500080243941583,
+      "loss": 0.1036,
+      "theoretical_loss": 4.393857927569534,
+      "tokens_seen": 196345856
+    },
+    {
+      "epoch": 0.06,
+      "objective/train/advantage_avg": 0.005338339600712061,
+      "objective/train/docs_used": 78313,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9282069206237793,
+      "objective/train/original_loss": 1.9282069206237793,
+      "objective/train/theoretical_loss": 4.393127152110409,
+      "objective/train/tokens_used": 217068000,
+      "objective/train/value_avg": -0.01195526123046875,
+      "objective/train/value_loss": 0.0004946094704791903,
+      "objective/train/value_max": -0.0013303756713867188,
+      "objective/train/value_min": -0.52392578125,
+      "objective/train/value_reward_corr": 0.5750984043308986,
+      "objective/train/value_std": 0.014373779296875,
+      "objective/train/weight_avg": 1.0055615901947021,
+      "objective/train/weighted_lm_loss": 1.9375295639038086,
+      "objective/train/weights_max": 1.6414918899536133,
+      "objective/train/weights_min": 0.4095984101295471,
+      "theoretical_loss": 4.393127152110409,
+      "tokens_seen": 196608000
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009499277804525759,
+      "loss": 0.1065,
+      "theoretical_loss": 4.393127152110409,
+      "tokens_seen": 196608000
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009498475365109934,
+      "loss": 0.1074,
+      "theoretical_loss": 4.392397622778343,
+      "tokens_seen": 196870144
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009497672925694111,
+      "loss": 0.107,
+      "theoretical_loss": 4.391669335793372,
+      "tokens_seen": 197132288
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009496870486278286,
+      "loss": 0.1047,
+      "theoretical_loss": 4.39094228739201,
+      "tokens_seen": 197394432
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009496068046862462,
+      "loss": 0.1091,
+      "theoretical_loss": 4.390216473827143,
+      "tokens_seen": 197656576
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009495265607446638,
+      "loss": 0.103,
+      "theoretical_loss": 4.389491891367953,
+      "tokens_seen": 197918720
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009494463168030813,
+      "loss": 0.108,
+      "theoretical_loss": 4.388768536299808,
+      "tokens_seen": 198180864
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000949366072861499,
+      "loss": 0.1101,
+      "theoretical_loss": 4.388046404924184,
+      "tokens_seen": 198443008
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009492858289199166,
+      "loss": 0.1064,
+      "theoretical_loss": 4.387325493558566,
+      "tokens_seen": 198705152
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009492055849783342,
+      "loss": 0.1078,
+      "theoretical_loss": 4.386605798536362,
+      "tokens_seen": 198967296
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009491253410367518,
+      "loss": 0.1102,
+      "theoretical_loss": 4.385887316206812,
+      "tokens_seen": 199229440
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009490450970951693,
+      "loss": 0.1037,
+      "theoretical_loss": 4.385170042934896,
+      "tokens_seen": 199491584
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009489648531535869,
+      "loss": 0.1051,
+      "theoretical_loss": 4.384453975101251,
+      "tokens_seen": 199753728
+    },
+    {
+      "epoch": 0.06,
+      "objective/train/advantage_avg": 0.0038920550141483545,
+      "objective/train/docs_used": 79488,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9931830167770386,
+      "objective/train/original_loss": 1.993182897567749,
+      "objective/train/theoretical_loss": 4.384096392096988,
+      "objective/train/tokens_used": 220344800,
+      "objective/train/value_avg": -0.00962066650390625,
+      "objective/train/value_loss": 0.00014405451656784862,
+      "objective/train/value_max": -0.0009326934814453125,
+      "objective/train/value_min": -0.26123046875,
+      "objective/train/value_reward_corr": 0.6021631165702985,
+      "objective/train/value_std": 0.00872039794921875,
+      "objective/train/weight_avg": 1.0039596557617188,
+      "objective/train/weighted_lm_loss": 2.001870632171631,
+      "objective/train/weights_max": 1.085601806640625,
+      "objective/train/weights_min": 0.3811125159263611,
+      "theoretical_loss": 4.384096392096988,
+      "tokens_seen": 199884800
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009488846092120045,
+      "loss": 0.1077,
+      "theoretical_loss": 4.38373910910208,
+      "tokens_seen": 200015872
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009488043652704221,
+      "loss": 0.1058,
+      "theoretical_loss": 4.383025441349063,
+      "tokens_seen": 200278016
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009487241213288396,
+      "loss": 0.1104,
+      "theoretical_loss": 4.382312968269276,
+      "tokens_seen": 200540160
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009486438773872573,
+      "loss": 0.1058,
+      "theoretical_loss": 4.381601686305098,
+      "tokens_seen": 200802304
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009485636334456749,
+      "loss": 0.1045,
+      "theoretical_loss": 4.38089159191413,
+      "tokens_seen": 201064448
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009484833895040924,
+      "loss": 0.1023,
+      "theoretical_loss": 4.380182681569111,
+      "tokens_seen": 201326592
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009484031455625101,
+      "loss": 0.1068,
+      "theoretical_loss": 4.379474951757829,
+      "tokens_seen": 201588736
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009483229016209276,
+      "loss": 0.107,
+      "theoretical_loss": 4.378768398983042,
+      "tokens_seen": 201850880
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009482426576793453,
+      "loss": 0.1039,
+      "theoretical_loss": 4.378063019762392,
+      "tokens_seen": 202113024
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009481624137377628,
+      "loss": 0.1046,
+      "theoretical_loss": 4.377358810628324,
+      "tokens_seen": 202375168
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009480821697961804,
+      "loss": 0.1048,
+      "theoretical_loss": 4.3766557681280025,
+      "tokens_seen": 202637312
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000948001925854598,
+      "loss": 0.1058,
+      "theoretical_loss": 4.375953888823233,
+      "tokens_seen": 202899456
+    },
+    {
+      "epoch": 0.06,
+      "objective/train/advantage_avg": 0.001105781877413392,
+      "objective/train/docs_used": 80605,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.189178228378296,
+      "objective/train/original_loss": 2.189178228378296,
+      "objective/train/theoretical_loss": 4.375253169290376,
+      "objective/train/tokens_used": 223621600,
+      "objective/train/value_avg": -0.01064300537109375,
+      "objective/train/value_loss": 0.00040862776222638786,
+      "objective/train/value_max": -0.0009965896606445312,
+      "objective/train/value_min": -0.306640625,
+      "objective/train/value_reward_corr": 0.6113931009977059,
+      "objective/train/value_std": 0.01380157470703125,
+      "objective/train/weight_avg": 1.0012942552566528,
+      "objective/train/weighted_lm_loss": 2.192763328552246,
+      "objective/train/weights_max": 1.2806123495101929,
+      "objective/train/weights_min": 0.37096279859542847,
+      "theoretical_loss": 4.375253169290376,
+      "tokens_seen": 203161600
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009479216819130155,
+      "loss": 0.105,
+      "theoretical_loss": 4.375253169290376,
+      "tokens_seen": 203161600
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009478414379714331,
+      "loss": 0.1059,
+      "theoretical_loss": 4.374553606120274,
+      "tokens_seen": 203423744
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009477611940298508,
+      "loss": 0.1047,
+      "theoretical_loss": 4.373855195918162,
+      "tokens_seen": 203685888
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009476809500882684,
+      "loss": 0.1076,
+      "theoretical_loss": 4.3731579353036,
+      "tokens_seen": 203948032
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009476007061466859,
+      "loss": 0.104,
+      "theoretical_loss": 4.372461820910382,
+      "tokens_seen": 204210176
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009475204622051036,
+      "loss": 0.1065,
+      "theoretical_loss": 4.371766849386468,
+      "tokens_seen": 204472320
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009474402182635211,
+      "loss": 0.1068,
+      "theoretical_loss": 4.3710730173939005,
+      "tokens_seen": 204734464
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009473599743219387,
+      "loss": 0.1063,
+      "theoretical_loss": 4.370380321608731,
+      "tokens_seen": 204996608
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009472797303803563,
+      "loss": 0.108,
+      "theoretical_loss": 4.369688758720937,
+      "tokens_seen": 205258752
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009471994864387738,
+      "loss": 0.1043,
+      "theoretical_loss": 4.368998325434355,
+      "tokens_seen": 205520896
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009471192424971916,
+      "loss": 0.1023,
+      "theoretical_loss": 4.3683090184666,
+      "tokens_seen": 205783040
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009470389985556091,
+      "loss": 0.1021,
+      "theoretical_loss": 4.367620834548987,
+      "tokens_seen": 206045184
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009469587546140267,
+      "loss": 0.1042,
+      "theoretical_loss": 4.3669337704264635,
+      "tokens_seen": 206307328
+    },
+    {
+      "epoch": 0.06,
+      "objective/train/advantage_avg": 0.0008768333937041461,
+      "objective/train/docs_used": 81724,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.068549871444702,
+      "objective/train/original_loss": 2.068549871444702,
+      "objective/train/theoretical_loss": 4.366590657274865,
+      "objective/train/tokens_used": 226898400,
+      "objective/train/value_avg": -0.0103302001953125,
+      "objective/train/value_loss": 0.00021303060930222273,
+      "objective/train/value_max": -0.0006880760192871094,
+      "objective/train/value_min": -0.205078125,
+      "objective/train/value_reward_corr": 0.6560487348637545,
+      "objective/train/value_std": 0.01080322265625,
+      "objective/train/weight_avg": 1.0009812116622925,
+      "objective/train/weighted_lm_loss": 2.0716254711151123,
+      "objective/train/weights_max": 1.1215885877609253,
+      "objective/train/weights_min": 0.8218191862106323,
+      "theoretical_loss": 4.366590657274865,
+      "tokens_seen": 206438400
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009468785106724443,
+      "loss": 0.1075,
+      "theoretical_loss": 4.366247822857533,
+      "tokens_seen": 206569472
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009467982667308619,
+      "loss": 0.1054,
+      "theoretical_loss": 4.365562988614176,
+      "tokens_seen": 206831616
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009467180227892794,
+      "loss": 0.1035,
+      "theoretical_loss": 4.364879264481787,
+      "tokens_seen": 207093760
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000946637778847697,
+      "loss": 0.1039,
+      "theoretical_loss": 4.364196647259092,
+      "tokens_seen": 207355904
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009465575349061146,
+      "loss": 0.1061,
+      "theoretical_loss": 4.363515133758084,
+      "tokens_seen": 207618048
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009464772909645321,
+      "loss": 0.1078,
+      "theoretical_loss": 4.3628347208039475,
+      "tokens_seen": 207880192
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009463970470229499,
+      "loss": 0.1021,
+      "theoretical_loss": 4.362155405234985,
+      "tokens_seen": 208142336
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009463168030813674,
+      "loss": 0.1093,
+      "theoretical_loss": 4.361477183902554,
+      "tokens_seen": 208404480
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000946236559139785,
+      "loss": 0.1055,
+      "theoretical_loss": 4.360800053670989,
+      "tokens_seen": 208666624
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009461563151982026,
+      "loss": 0.1034,
+      "theoretical_loss": 4.360124011417536,
+      "tokens_seen": 208928768
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009460760712566201,
+      "loss": 0.105,
+      "theoretical_loss": 4.359449054032282,
+      "tokens_seen": 209190912
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009459958273150377,
+      "loss": 0.1048,
+      "theoretical_loss": 4.358775178418089,
+      "tokens_seen": 209453056
+    },
+    {
+      "epoch": 0.06,
+      "objective/train/advantage_avg": 0.001143173431046307,
+      "objective/train/docs_used": 82917,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.0207271575927734,
+      "objective/train/original_loss": 2.0207271575927734,
+      "objective/train/theoretical_loss": 4.358102381490517,
+      "objective/train/tokens_used": 230175200,
+      "objective/train/value_avg": -0.009368896484375,
+      "objective/train/value_loss": 0.00035423377994447947,
+      "objective/train/value_max": -0.0006718635559082031,
+      "objective/train/value_min": -0.259033203125,
+      "objective/train/value_reward_corr": 0.4731410524064374,
+      "objective/train/value_std": 0.00994110107421875,
+      "objective/train/weight_avg": 1.001305103302002,
+      "objective/train/weighted_lm_loss": 2.0232884883880615,
+      "objective/train/weights_max": 1.2774591445922852,
+      "objective/train/weights_min": 0.3702460527420044,
+      "theoretical_loss": 4.358102381490517,
+      "tokens_seen": 209715200
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009459155833734553,
+      "loss": 0.1056,
+      "theoretical_loss": 4.358102381490517,
+      "tokens_seen": 209715200
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009458353394318729,
+      "loss": 0.1055,
+      "theoretical_loss": 4.3574306601777675,
+      "tokens_seen": 209977344
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009457550954902905,
+      "loss": 0.1016,
+      "theoretical_loss": 4.356760011420608,
+      "tokens_seen": 210239488
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009456748515487081,
+      "loss": 0.1089,
+      "theoretical_loss": 4.3560904321723095,
+      "tokens_seen": 210501632
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009455946076071257,
+      "loss": 0.106,
+      "theoretical_loss": 4.355421919398576,
+      "tokens_seen": 210763776
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009455143636655433,
+      "loss": 0.1057,
+      "theoretical_loss": 4.354754470077481,
+      "tokens_seen": 211025920
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009454341197239609,
+      "loss": 0.1032,
+      "theoretical_loss": 4.354088081199402,
+      "tokens_seen": 211288064
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009453538757823784,
+      "loss": 0.1041,
+      "theoretical_loss": 4.3534227497669535,
+      "tokens_seen": 211550208
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009452736318407961,
+      "loss": 0.1035,
+      "theoretical_loss": 4.352758472794923,
+      "tokens_seen": 211812352
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009451933878992136,
+      "loss": 0.1058,
+      "theoretical_loss": 4.352095247310208,
+      "tokens_seen": 212074496
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009451131439576312,
+      "loss": 0.1048,
+      "theoretical_loss": 4.351433070351748,
+      "tokens_seen": 212336640
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009450329000160488,
+      "loss": 0.1064,
+      "theoretical_loss": 4.350771938970466,
+      "tokens_seen": 212598784
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009449526560744663,
+      "loss": 0.1028,
+      "theoretical_loss": 4.350111850229202,
+      "tokens_seen": 212860928
+    },
+    {
+      "epoch": 0.06,
+      "objective/train/advantage_avg": 0.0043486496433615685,
+      "objective/train/docs_used": 84140,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9397648572921753,
+      "objective/train/original_loss": 1.9397649765014648,
+      "objective/train/theoretical_loss": 4.349782195933957,
+      "objective/train/tokens_used": 233452000,
+      "objective/train/value_avg": -0.01163482666015625,
+      "objective/train/value_loss": 0.00020028821018058807,
+      "objective/train/value_max": -0.0011072158813476562,
+      "objective/train/value_min": -0.1822509765625,
+      "objective/train/value_reward_corr": 0.6988513016032423,
+      "objective/train/value_std": 0.0160675048828125,
+      "objective/train/weight_avg": 1.0044445991516113,
+      "objective/train/weighted_lm_loss": 1.948439121246338,
+      "objective/train/weights_max": 1.1115705966949463,
+      "objective/train/weights_min": 0.3973115086555481,
+      "theoretical_loss": 4.349782195933957,
+      "tokens_seen": 212992000
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.000944872412132884,
+      "loss": 0.1017,
+      "theoretical_loss": 4.34945280120265,
+      "tokens_seen": 213123072
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009447921681913016,
+      "loss": 0.103,
+      "theoretical_loss": 4.348794788977298,
+      "tokens_seen": 213385216
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009447119242497192,
+      "loss": 0.1039,
+      "theoretical_loss": 4.348137810651366,
+      "tokens_seen": 213647360
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009446316803081368,
+      "loss": 0.102,
+      "theoretical_loss": 4.347481863334738,
+      "tokens_seen": 213909504
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009445514363665544,
+      "loss": 0.1045,
+      "theoretical_loss": 4.346826944148912,
+      "tokens_seen": 214171648
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 0.0009444711924249719,
+      "loss": 0.1049,
+      "theoretical_loss": 4.3461730502269305,
+      "tokens_seen": 214433792
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009443909484833896,
+      "loss": 0.1036,
+      "theoretical_loss": 4.345520178713323,
+      "tokens_seen": 214695936
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009443107045418071,
+      "loss": 0.1077,
+      "theoretical_loss": 4.344868326764045,
+      "tokens_seen": 214958080
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009442304606002246,
+      "loss": 0.1054,
+      "theoretical_loss": 4.344217491546422,
+      "tokens_seen": 215220224
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009441502166586424,
+      "loss": 0.1037,
+      "theoretical_loss": 4.343567670239084,
+      "tokens_seen": 215482368
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009440699727170599,
+      "loss": 0.1059,
+      "theoretical_loss": 4.342918860031914,
+      "tokens_seen": 215744512
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009439897287754775,
+      "loss": 0.1007,
+      "theoretical_loss": 4.342271058125983,
+      "tokens_seen": 216006656
+    },
+    {
+      "epoch": 0.07,
+      "objective/train/advantage_avg": 0.0013779407599940896,
+      "objective/train/docs_used": 85191,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.144592761993408,
+      "objective/train/original_loss": 2.1445932388305664,
+      "objective/train/theoretical_loss": 4.341624261733497,
+      "objective/train/tokens_used": 236728800,
+      "objective/train/value_avg": -0.01018524169921875,
+      "objective/train/value_loss": 0.0004004819202236831,
+      "objective/train/value_max": -0.0008134841918945312,
+      "objective/train/value_min": -0.294677734375,
+      "objective/train/value_reward_corr": 0.4947933347916151,
+      "objective/train/value_std": 0.01059722900390625,
+      "objective/train/weight_avg": 1.0015608072280884,
+      "objective/train/weighted_lm_loss": 2.14778208732605,
+      "objective/train/weights_max": 1.2878975868225098,
+      "objective/train/weights_min": 0.3711496591567993,
+      "theoretical_loss": 4.341624261733497,
+      "tokens_seen": 216268800
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009439094848338951,
+      "loss": 0.1053,
+      "theoretical_loss": 4.341624261733497,
+      "tokens_seen": 216268800
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009438292408923127,
+      "loss": 0.1064,
+      "theoretical_loss": 4.340978468077735,
+      "tokens_seen": 216530944
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009437489969507302,
+      "loss": 0.1038,
+      "theoretical_loss": 4.340333674392992,
+      "tokens_seen": 216793088
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009436687530091478,
+      "loss": 0.1046,
+      "theoretical_loss": 4.339689877924531,
+      "tokens_seen": 217055232
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009435885090675654,
+      "loss": 0.1038,
+      "theoretical_loss": 4.3390470759285105,
+      "tokens_seen": 217317376
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009435082651259829,
+      "loss": 0.1039,
+      "theoretical_loss": 4.338405265671941,
+      "tokens_seen": 217579520
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009434280211844006,
+      "loss": 0.0991,
+      "theoretical_loss": 4.337764444432625,
+      "tokens_seen": 217841664
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009433477772428182,
+      "loss": 0.1038,
+      "theoretical_loss": 4.337124609499101,
+      "tokens_seen": 218103808
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009432675333012359,
+      "loss": 0.1003,
+      "theoretical_loss": 4.336485758170589,
+      "tokens_seen": 218365952
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009431872893596534,
+      "loss": 0.106,
+      "theoretical_loss": 4.335847887756934,
+      "tokens_seen": 218628096
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009431070454180709,
+      "loss": 0.1021,
+      "theoretical_loss": 4.335210995578553,
+      "tokens_seen": 218890240
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009430268014764886,
+      "loss": 0.1041,
+      "theoretical_loss": 4.334575078966383,
+      "tokens_seen": 219152384
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009429465575349061,
+      "loss": 0.1049,
+      "theoretical_loss": 4.333940135261823,
+      "tokens_seen": 219414528
+    },
+    {
+      "epoch": 0.07,
+      "objective/train/advantage_avg": 0.002295425161719322,
+      "objective/train/docs_used": 86440,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9799681901931763,
+      "objective/train/original_loss": 1.9799679517745972,
+      "objective/train/theoretical_loss": 4.3336230274219325,
+      "objective/train/tokens_used": 240005600,
+      "objective/train/value_avg": -0.00746917724609375,
+      "objective/train/value_loss": 0.00019181027892045677,
+      "objective/train/value_max": -0.0007948875427246094,
+      "objective/train/value_min": -0.2081298828125,
+      "objective/train/value_reward_corr": 0.306679385928593,
+      "objective/train/value_std": 0.00490570068359375,
+      "objective/train/weight_avg": 1.002386212348938,
+      "objective/train/weighted_lm_loss": 1.9838651418685913,
+      "objective/train/weights_max": 1.0839669704437256,
+      "objective/train/weights_min": 0.37081003189086914,
+      "theoretical_loss": 4.3336230274219325,
+      "tokens_seen": 219545600
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009428663135933237,
+      "loss": 0.1019,
+      "theoretical_loss": 4.333306161816684,
+      "tokens_seen": 219676672
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009427860696517413,
+      "loss": 0.1057,
+      "theoretical_loss": 4.332673155993131,
+      "tokens_seen": 219938816
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.000942705825710159,
+      "loss": 0.1016,
+      "theoretical_loss": 4.332041115163636,
+      "tokens_seen": 220200960
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009426255817685765,
+      "loss": 0.1049,
+      "theoretical_loss": 4.331410036710925,
+      "tokens_seen": 220463104
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009425453378269941,
+      "loss": 0.1027,
+      "theoretical_loss": 4.330779918027919,
+      "tokens_seen": 220725248
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009424650938854117,
+      "loss": 0.1047,
+      "theoretical_loss": 4.330150756517692,
+      "tokens_seen": 220987392
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009423848499438292,
+      "loss": 0.1025,
+      "theoretical_loss": 4.3295225495934115,
+      "tokens_seen": 221249536
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009423046060022469,
+      "loss": 0.0982,
+      "theoretical_loss": 4.328895294678292,
+      "tokens_seen": 221511680
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009422243620606644,
+      "loss": 0.103,
+      "theoretical_loss": 4.32826898920554,
+      "tokens_seen": 221773824
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009421441181190821,
+      "loss": 0.1057,
+      "theoretical_loss": 4.3276436306183115,
+      "tokens_seen": 222035968
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009420638741774996,
+      "loss": 0.1003,
+      "theoretical_loss": 4.327019216369651,
+      "tokens_seen": 222298112
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009419836302359171,
+      "loss": 0.1012,
+      "theoretical_loss": 4.32639574392245,
+      "tokens_seen": 222560256
+    },
+    {
+      "epoch": 0.07,
+      "objective/train/advantage_avg": 0.0020489406306296587,
+      "objective/train/docs_used": 87601,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8337969779968262,
+      "objective/train/original_loss": 1.8337969779968262,
+      "objective/train/theoretical_loss": 4.325773210749392,
+      "objective/train/tokens_used": 243282400,
+      "objective/train/value_avg": -0.0086669921875,
+      "objective/train/value_loss": 0.000297574297292158,
+      "objective/train/value_max": -0.0009889602661132812,
+      "objective/train/value_min": -0.6943359375,
+      "objective/train/value_reward_corr": 0.5675373911477959,
+      "objective/train/value_std": 0.0108795166015625,
+      "objective/train/weight_avg": 1.0021827220916748,
+      "objective/train/weighted_lm_loss": 1.838815450668335,
+      "objective/train/weights_max": 1.4390860795974731,
+      "objective/train/weights_min": 0.37089207768440247,
+      "theoretical_loss": 4.325773210749392,
+      "tokens_seen": 222822400
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009419033862943349,
+      "loss": 0.0976,
+      "theoretical_loss": 4.325773210749392,
+      "tokens_seen": 222822400
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009418231423527524,
+      "loss": 0.1022,
+      "theoretical_loss": 4.325151614332908,
+      "tokens_seen": 223084544
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.00094174289841117,
+      "loss": 0.1034,
+      "theoretical_loss": 4.3245309521651265,
+      "tokens_seen": 223346688
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009416626544695876,
+      "loss": 0.103,
+      "theoretical_loss": 4.323911221747817,
+      "tokens_seen": 223608832
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009415824105280052,
+      "loss": 0.1048,
+      "theoretical_loss": 4.323292420592356,
+      "tokens_seen": 223870976
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009415021665864227,
+      "loss": 0.1024,
+      "theoretical_loss": 4.322674546219666,
+      "tokens_seen": 224133120
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009414219226448404,
+      "loss": 0.1028,
+      "theoretical_loss": 4.322057596160174,
+      "tokens_seen": 224395264
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009413416787032579,
+      "loss": 0.0994,
+      "theoretical_loss": 4.321441567953762,
+      "tokens_seen": 224657408
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009412614347616754,
+      "loss": 0.1017,
+      "theoretical_loss": 4.320826459149725,
+      "tokens_seen": 224919552
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009411811908200932,
+      "loss": 0.1075,
+      "theoretical_loss": 4.3202122673067125,
+      "tokens_seen": 225181696
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009411009468785107,
+      "loss": 0.1007,
+      "theoretical_loss": 4.319598989992695,
+      "tokens_seen": 225443840
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009410207029369283,
+      "loss": 0.1049,
+      "theoretical_loss": 4.318986624784908,
+      "tokens_seen": 225705984
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009409404589953459,
+      "loss": 0.1034,
+      "theoretical_loss": 4.318375169269813,
+      "tokens_seen": 225968128
+    },
+    {
+      "epoch": 0.07,
+      "objective/train/advantage_avg": 0.0023550805635750294,
+      "objective/train/docs_used": 88830,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9909827709197998,
+      "objective/train/original_loss": 1.9909827709197998,
+      "objective/train/theoretical_loss": 4.3180697818953035,
+      "objective/train/tokens_used": 246559200,
+      "objective/train/value_avg": -0.00814056396484375,
+      "objective/train/value_loss": 0.00033672992140054703,
+      "objective/train/value_max": -0.0005397796630859375,
+      "objective/train/value_min": -0.321533203125,
+      "objective/train/value_reward_corr": 0.5563445957193008,
+      "objective/train/value_std": 0.010406494140625,
+      "objective/train/weight_avg": 1.0025025606155396,
+      "objective/train/weighted_lm_loss": 1.9960408210754395,
+      "objective/train/weights_max": 1.258882761001587,
+      "objective/train/weights_min": 0.3709288537502289,
+      "theoretical_loss": 4.3180697818953035,
+      "tokens_seen": 226099200
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009408602150537635,
+      "loss": 0.1014,
+      "theoretical_loss": 4.317764621043046,
+      "tokens_seen": 226230272
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009407799711121811,
+      "loss": 0.0995,
+      "theoretical_loss": 4.317154977709375,
+      "tokens_seen": 226492416
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009406997271705986,
+      "loss": 0.1001,
+      "theoretical_loss": 4.3165462368826555,
+      "tokens_seen": 226754560
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009406194832290162,
+      "loss": 0.0995,
+      "theoretical_loss": 4.315938396185782,
+      "tokens_seen": 227016704
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009405392392874338,
+      "loss": 0.1001,
+      "theoretical_loss": 4.315331453250648,
+      "tokens_seen": 227278848
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009404589953458514,
+      "loss": 0.0994,
+      "theoretical_loss": 4.314725405718099,
+      "tokens_seen": 227540992
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.000940378751404269,
+      "loss": 0.0997,
+      "theoretical_loss": 4.314120251237887,
+      "tokens_seen": 227803136
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009402985074626867,
+      "loss": 0.103,
+      "theoretical_loss": 4.31351598746863,
+      "tokens_seen": 228065280
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009402182635211042,
+      "loss": 0.1013,
+      "theoretical_loss": 4.312912612077767,
+      "tokens_seen": 228327424
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009401380195795217,
+      "loss": 0.1022,
+      "theoretical_loss": 4.312310122741512,
+      "tokens_seen": 228589568
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009400577756379394,
+      "loss": 0.1028,
+      "theoretical_loss": 4.311708517144817,
+      "tokens_seen": 228851712
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009399775316963569,
+      "loss": 0.1024,
+      "theoretical_loss": 4.311107792981323,
+      "tokens_seen": 229113856
+    },
+    {
+      "epoch": 0.07,
+      "objective/train/advantage_avg": 0.0026854872703552246,
+      "objective/train/docs_used": 90061,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.144618511199951,
+      "objective/train/original_loss": 2.144618511199951,
+      "objective/train/theoretical_loss": 4.310507947953321,
+      "objective/train/tokens_used": 249836000,
+      "objective/train/value_avg": -0.00982666015625,
+      "objective/train/value_loss": 0.00028081017080694437,
+      "objective/train/value_max": -0.0007295608520507812,
+      "objective/train/value_min": -0.66845703125,
+      "objective/train/value_reward_corr": 0.6199593894084428,
+      "objective/train/value_std": 0.01155853271484375,
+      "objective/train/weight_avg": 1.0028172731399536,
+      "objective/train/weighted_lm_loss": 2.1502115726470947,
+      "objective/train/weights_max": 1.3387657403945923,
+      "objective/train/weights_min": 0.37122613191604614,
+      "theoretical_loss": 4.310507947953321,
+      "tokens_seen": 229376000
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009398972877547745,
+      "loss": 0.1031,
+      "theoretical_loss": 4.310507947953321,
+      "tokens_seen": 229376000
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009398170438131921,
+      "loss": 0.1045,
+      "theoretical_loss": 4.309908979771709,
+      "tokens_seen": 229638144
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009397367998716097,
+      "loss": 0.1044,
+      "theoretical_loss": 4.3093108861559495,
+      "tokens_seen": 229900288
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009396565559300272,
+      "loss": 0.1022,
+      "theoretical_loss": 4.308713664834029,
+      "tokens_seen": 230162432
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009395763119884449,
+      "loss": 0.0958,
+      "theoretical_loss": 4.308117313542413,
+      "tokens_seen": 230424576
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009394960680468625,
+      "loss": 0.1022,
+      "theoretical_loss": 4.30752183002601,
+      "tokens_seen": 230686720
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009394158241052801,
+      "loss": 0.099,
+      "theoretical_loss": 4.3069272120381275,
+      "tokens_seen": 230948864
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009393355801636977,
+      "loss": 0.102,
+      "theoretical_loss": 4.30633345734043,
+      "tokens_seen": 231211008
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009392553362221152,
+      "loss": 0.1004,
+      "theoretical_loss": 4.3057405637029,
+      "tokens_seen": 231473152
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009391750922805329,
+      "loss": 0.1015,
+      "theoretical_loss": 4.305148528903798,
+      "tokens_seen": 231735296
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009390948483389504,
+      "loss": 0.1023,
+      "theoretical_loss": 4.304557350729623,
+      "tokens_seen": 231997440
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009390146043973679,
+      "loss": 0.1002,
+      "theoretical_loss": 4.303967026975072,
+      "tokens_seen": 232259584
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009389343604557857,
+      "loss": 0.1024,
+      "theoretical_loss": 4.303377555442998,
+      "tokens_seen": 232521728
+    },
+    {
+      "epoch": 0.07,
+      "objective/train/advantage_avg": 0.0007922460208646953,
+      "objective/train/docs_used": 91258,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9420995712280273,
+      "objective/train/original_loss": 1.9420994520187378,
+      "objective/train/theoretical_loss": 4.303083138576003,
+      "objective/train/tokens_used": 253112800,
+      "objective/train/value_avg": -0.006786346435546875,
+      "objective/train/value_loss": 0.0005959350382909179,
+      "objective/train/value_max": -0.0007352828979492188,
+      "objective/train/value_min": -0.293701171875,
+      "objective/train/value_reward_corr": 0.39993119929541304,
+      "objective/train/value_std": 0.007472991943359375,
+      "objective/train/weight_avg": 1.0010305643081665,
+      "objective/train/weighted_lm_loss": 1.9428025484085083,
+      "objective/train/weights_max": 1.215839147567749,
+      "objective/train/weights_min": 0.2238616794347763,
+      "theoretical_loss": 4.303083138576003,
+      "tokens_seen": 232652800
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009388541165142032,
+      "loss": 0.098,
+      "theoretical_loss": 4.302788933944375,
+      "tokens_seen": 232783872
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009387738725726208,
+      "loss": 0.1014,
+      "theoretical_loss": 4.302201160298255,
+      "tokens_seen": 233046016
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009386936286310384,
+      "loss": 0.1019,
+      "theoretical_loss": 4.301614232331733,
+      "tokens_seen": 233308160
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.000938613384689456,
+      "loss": 0.101,
+      "theoretical_loss": 4.301028147879904,
+      "tokens_seen": 233570304
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009385331407478735,
+      "loss": 0.0983,
+      "theoretical_loss": 4.300442904785831,
+      "tokens_seen": 233832448
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009384528968062911,
+      "loss": 0.1014,
+      "theoretical_loss": 4.299858500900495,
+      "tokens_seen": 234094592
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009383726528647087,
+      "loss": 0.1024,
+      "theoretical_loss": 4.2992749340827725,
+      "tokens_seen": 234356736
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009382924089231262,
+      "loss": 0.1028,
+      "theoretical_loss": 4.298692202199386,
+      "tokens_seen": 234618880
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.000938212164981544,
+      "loss": 0.101,
+      "theoretical_loss": 4.298110303124871,
+      "tokens_seen": 234881024
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009381319210399615,
+      "loss": 0.1022,
+      "theoretical_loss": 4.29752923474154,
+      "tokens_seen": 235143168
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009380516770983792,
+      "loss": 0.1006,
+      "theoretical_loss": 4.29694899493944,
+      "tokens_seen": 235405312
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009379714331567967,
+      "loss": 0.1017,
+      "theoretical_loss": 4.2963695816163225,
+      "tokens_seen": 235667456
+    },
+    {
+      "epoch": 0.07,
+      "objective/train/advantage_avg": 0.0031256088986992836,
+      "objective/train/docs_used": 92471,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.102999448776245,
+      "objective/train/original_loss": 2.102999687194824,
+      "objective/train/theoretical_loss": 4.295790992677603,
+      "objective/train/tokens_used": 256389600,
+      "objective/train/value_avg": -0.01329803466796875,
+      "objective/train/value_loss": 0.0007132674218155444,
+      "objective/train/value_max": -0.0005130767822265625,
+      "objective/train/value_min": -0.8544921875,
+      "objective/train/value_reward_corr": 0.6135290805544735,
+      "objective/train/value_std": 0.0253448486328125,
+      "objective/train/weight_avg": 1.0034579038619995,
+      "objective/train/weighted_lm_loss": 2.107316493988037,
+      "objective/train/weights_max": 1.5163893699645996,
+      "objective/train/weights_min": 0.39436399936676025,
+      "theoretical_loss": 4.295790992677603,
+      "tokens_seen": 235929600
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009378911892152143,
+      "loss": 0.0997,
+      "theoretical_loss": 4.295790992677603,
+      "tokens_seen": 235929600
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009378109452736319,
+      "loss": 0.1009,
+      "theoretical_loss": 4.2952132260363225,
+      "tokens_seen": 236191744
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009377307013320494,
+      "loss": 0.0996,
+      "theoretical_loss": 4.294636279613117,
+      "tokens_seen": 236453888
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.000937650457390467,
+      "loss": 0.1036,
+      "theoretical_loss": 4.294060151336178,
+      "tokens_seen": 236716032
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009375702134488846,
+      "loss": 0.1019,
+      "theoretical_loss": 4.293484839141217,
+      "tokens_seen": 236978176
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009374899695073022,
+      "loss": 0.1,
+      "theoretical_loss": 4.29291034097143,
+      "tokens_seen": 237240320
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009374097255657198,
+      "loss": 0.097,
+      "theoretical_loss": 4.2923366547774595,
+      "tokens_seen": 237502464
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009373294816241375,
+      "loss": 0.1002,
+      "theoretical_loss": 4.2917637785173675,
+      "tokens_seen": 237764608
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.000937249237682555,
+      "loss": 0.1004,
+      "theoretical_loss": 4.291191710156591,
+      "tokens_seen": 238026752
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009371689937409725,
+      "loss": 0.1003,
+      "theoretical_loss": 4.290620447667912,
+      "tokens_seen": 238288896
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009370887497993902,
+      "loss": 0.0979,
+      "theoretical_loss": 4.290049989031424,
+      "tokens_seen": 238551040
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009370085058578077,
+      "loss": 0.1033,
+      "theoretical_loss": 4.289480332234493,
+      "tokens_seen": 238813184
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009369282619162254,
+      "loss": 0.1012,
+      "theoretical_loss": 4.288911475271731,
+      "tokens_seen": 239075328
+    },
+    {
+      "epoch": 0.07,
+      "objective/train/advantage_avg": 0.001886105164885521,
+      "objective/train/docs_used": 93570,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.0680296421051025,
+      "objective/train/original_loss": 2.0680294036865234,
+      "objective/train/theoretical_loss": 4.28862734610345,
+      "objective/train/tokens_used": 259666400,
+      "objective/train/value_avg": -0.008026123046875,
+      "objective/train/value_loss": 0.0002583606983534992,
+      "objective/train/value_max": -0.0006852149963378906,
+      "objective/train/value_min": -0.2454833984375,
+      "objective/train/value_reward_corr": 0.48885293063683233,
+      "objective/train/value_std": 0.007061004638671875,
+      "objective/train/weight_avg": 1.0020033121109009,
+      "objective/train/weighted_lm_loss": 2.0723109245300293,
+      "objective/train/weights_max": 1.1007682085037231,
+      "objective/train/weights_min": 0.3832702934741974,
+      "theoretical_loss": 4.28862734610345,
+      "tokens_seen": 239206400
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009368480179746429,
+      "loss": 0.1002,
+      "theoretical_loss": 4.288343416144952,
+      "tokens_seen": 239337472
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009367677740330605,
+      "loss": 0.1023,
+      "theoretical_loss": 4.287776152863146,
+      "tokens_seen": 239599616
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009366875300914782,
+      "loss": 0.1016,
+      "theoretical_loss": 4.287209683442444,
+      "tokens_seen": 239861760
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009366072861498957,
+      "loss": 0.1004,
+      "theoretical_loss": 4.286644005906081,
+      "tokens_seen": 240123904
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009365270422083133,
+      "loss": 0.101,
+      "theoretical_loss": 4.286079118284368,
+      "tokens_seen": 240386048
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009364467982667309,
+      "loss": 0.1034,
+      "theoretical_loss": 4.285515018614655,
+      "tokens_seen": 240648192
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009363665543251485,
+      "loss": 0.096,
+      "theoretical_loss": 4.2849517049412995,
+      "tokens_seen": 240910336
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.000936286310383566,
+      "loss": 0.0974,
+      "theoretical_loss": 4.284389175315636,
+      "tokens_seen": 241172480
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009362060664419837,
+      "loss": 0.0995,
+      "theoretical_loss": 4.283827427795939,
+      "tokens_seen": 241434624
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009361258225004012,
+      "loss": 0.0987,
+      "theoretical_loss": 4.283266460447394,
+      "tokens_seen": 241696768
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009360455785588187,
+      "loss": 0.0967,
+      "theoretical_loss": 4.282706271342066,
+      "tokens_seen": 241958912
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009359653346172365,
+      "loss": 0.1028,
+      "theoretical_loss": 4.282146858558866,
+      "tokens_seen": 242221056
+    },
+    {
+      "epoch": 0.07,
+      "objective/train/advantage_avg": 0.001827435800805688,
+      "objective/train/docs_used": 94878,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.0017893314361572,
+      "objective/train/original_loss": 2.001789093017578,
+      "objective/train/theoretical_loss": 4.281588220183519,
+      "objective/train/tokens_used": 262943200,
+      "objective/train/value_avg": -0.01013946533203125,
+      "objective/train/value_loss": 0.0004723604361061007,
+      "objective/train/value_max": -0.0006642341613769531,
+      "objective/train/value_min": -0.4013671875,
+      "objective/train/value_reward_corr": 0.613649177640665,
+      "objective/train/value_std": 0.0154876708984375,
+      "objective/train/weight_avg": 1.0020326375961304,
+      "objective/train/weighted_lm_loss": 2.0051186084747314,
+      "objective/train/weights_max": 1.3551794290542603,
+      "objective/train/weights_min": 0.37200865149497986,
+      "theoretical_loss": 4.281588220183519,
+      "tokens_seen": 242483200
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.000935885090675654,
+      "loss": 0.0977,
+      "theoretical_loss": 4.281588220183519,
+      "tokens_seen": 242483200
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009358048467340716,
+      "loss": 0.0995,
+      "theoretical_loss": 4.281030354308533,
+      "tokens_seen": 242745344
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009357246027924892,
+      "loss": 0.0989,
+      "theoretical_loss": 4.280473259033169,
+      "tokens_seen": 243007488
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009356443588509068,
+      "loss": 0.0982,
+      "theoretical_loss": 4.27991693246341,
+      "tokens_seen": 243269632
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009355641149093244,
+      "loss": 0.0986,
+      "theoretical_loss": 4.279361372711923,
+      "tokens_seen": 243531776
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009354838709677419,
+      "loss": 0.1022,
+      "theoretical_loss": 4.278806577898042,
+      "tokens_seen": 243793920
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009354036270261595,
+      "loss": 0.0979,
+      "theoretical_loss": 4.278252546147724,
+      "tokens_seen": 244056064
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009353233830845771,
+      "loss": 0.0995,
+      "theoretical_loss": 4.277699275593523,
+      "tokens_seen": 244318208
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009352431391429948,
+      "loss": 0.0949,
+      "theoretical_loss": 4.277146764374566,
+      "tokens_seen": 244580352
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009351628952014123,
+      "loss": 0.0999,
+      "theoretical_loss": 4.276595010636514,
+      "tokens_seen": 244842496
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.00093508265125983,
+      "loss": 0.0993,
+      "theoretical_loss": 4.276044012531534,
+      "tokens_seen": 245104640
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009350024073182475,
+      "loss": 0.1013,
+      "theoretical_loss": 4.275493768218274,
+      "tokens_seen": 245366784
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009349221633766651,
+      "loss": 0.0975,
+      "theoretical_loss": 4.274944275861828,
+      "tokens_seen": 245628928
+    },
+    {
+      "epoch": 0.07,
+      "objective/train/advantage_avg": 0.002574330661445856,
+      "objective/train/docs_used": 96014,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.0550389289855957,
+      "objective/train/original_loss": 2.055039405822754,
+      "objective/train/theoretical_loss": 4.274669811095759,
+      "objective/train/tokens_used": 266220000,
+      "objective/train/value_avg": -0.01256561279296875,
+      "objective/train/value_loss": 0.0013382832985371351,
+      "objective/train/value_max": -0.0008039474487304688,
+      "objective/train/value_min": -0.810546875,
+      "objective/train/value_reward_corr": 0.6458940332158234,
+      "objective/train/value_std": 0.0258026123046875,
+      "objective/train/weight_avg": 1.003141164779663,
+      "objective/train/weighted_lm_loss": 2.06048846244812,
+      "objective/train/weights_max": 1.4646140336990356,
+      "objective/train/weights_min": 0.36990442872047424,
+      "theoretical_loss": 4.274669811095759,
+      "tokens_seen": 245760000
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009348419194350827,
+      "loss": 0.098,
+      "theoretical_loss": 4.274395533633712,
+      "tokens_seen": 245891072
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009347616754935002,
+      "loss": 0.096,
+      "theoretical_loss": 4.273847539711825,
+      "tokens_seen": 246153216
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009346814315519178,
+      "loss": 0.1013,
+      "theoretical_loss": 4.273300292280435,
+      "tokens_seen": 246415360
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009346011876103354,
+      "loss": 0.0989,
+      "theoretical_loss": 4.272753789530134,
+      "tokens_seen": 246677504
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.000934520943668753,
+      "loss": 0.096,
+      "theoretical_loss": 4.272208029657822,
+      "tokens_seen": 246939648
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009344406997271707,
+      "loss": 0.0971,
+      "theoretical_loss": 4.271663010866669,
+      "tokens_seen": 247201792
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 0.0009343604557855883,
+      "loss": 0.1003,
+      "theoretical_loss": 4.2711187313660925,
+      "tokens_seen": 247463936
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009342802118440058,
+      "loss": 0.0983,
+      "theoretical_loss": 4.270575189371727,
+      "tokens_seen": 247726080
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009341999679024234,
+      "loss": 0.0991,
+      "theoretical_loss": 4.270032383105398,
+      "tokens_seen": 247988224
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000934119723960841,
+      "loss": 0.0973,
+      "theoretical_loss": 4.269490310795089,
+      "tokens_seen": 248250368
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009340394800192585,
+      "loss": 0.1025,
+      "theoretical_loss": 4.268948970674917,
+      "tokens_seen": 248512512
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009339592360776762,
+      "loss": 0.1013,
+      "theoretical_loss": 4.268408360985109,
+      "tokens_seen": 248774656
+    },
+    {
+      "epoch": 0.08,
+      "objective/train/advantage_avg": 0.00040733805508352816,
+      "objective/train/docs_used": 97104,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9560189247131348,
+      "objective/train/original_loss": 1.9560186862945557,
+      "objective/train/theoretical_loss": 4.267868479971966,
+      "objective/train/tokens_used": 269496800,
+      "objective/train/value_avg": -0.010101318359375,
+      "objective/train/value_loss": 0.0005951999919489026,
+      "objective/train/value_max": -0.0006189346313476562,
+      "objective/train/value_min": -0.81201171875,
+      "objective/train/value_reward_corr": 0.654036095652701,
+      "objective/train/value_std": 0.0165863037109375,
+      "objective/train/weight_avg": 1.0006780624389648,
+      "objective/train/weighted_lm_loss": 1.9563541412353516,
+      "objective/train/weights_max": 1.3073004484176636,
+      "objective/train/weights_min": 0.4099235534667969,
+      "theoretical_loss": 4.267868479971966,
+      "tokens_seen": 249036800
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009338789921360937,
+      "loss": 0.0996,
+      "theoretical_loss": 4.267868479971966,
+      "tokens_seen": 249036800
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009337987481945113,
+      "loss": 0.099,
+      "theoretical_loss": 4.267329325887841,
+      "tokens_seen": 249298944
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000933718504252929,
+      "loss": 0.0923,
+      "theoretical_loss": 4.266790896991109,
+      "tokens_seen": 249561088
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009336382603113465,
+      "loss": 0.1004,
+      "theoretical_loss": 4.266253191546146,
+      "tokens_seen": 249823232
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009335580163697641,
+      "loss": 0.0992,
+      "theoretical_loss": 4.265716207823292,
+      "tokens_seen": 250085376
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009334777724281817,
+      "loss": 0.0991,
+      "theoretical_loss": 4.2651799440988345,
+      "tokens_seen": 250347520
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009333975284865993,
+      "loss": 0.0974,
+      "theoretical_loss": 4.2646443986549745,
+      "tokens_seen": 250609664
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009333172845450168,
+      "loss": 0.099,
+      "theoretical_loss": 4.264109569779803,
+      "tokens_seen": 250871808
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009332370406034345,
+      "loss": 0.1013,
+      "theoretical_loss": 4.263575455767277,
+      "tokens_seen": 251133952
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000933156796661852,
+      "loss": 0.0997,
+      "theoretical_loss": 4.263042054917186,
+      "tokens_seen": 251396096
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009330765527202696,
+      "loss": 0.0991,
+      "theoretical_loss": 4.262509365535134,
+      "tokens_seen": 251658240
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009329963087786873,
+      "loss": 0.0967,
+      "theoretical_loss": 4.261977385932512,
+      "tokens_seen": 251920384
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009329160648371048,
+      "loss": 0.0985,
+      "theoretical_loss": 4.261446114426466,
+      "tokens_seen": 252182528
+    },
+    {
+      "epoch": 0.08,
+      "objective/train/advantage_avg": 0.002480955794453621,
+      "objective/train/docs_used": 98257,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.974042534828186,
+      "objective/train/original_loss": 1.974042534828186,
+      "objective/train/theoretical_loss": 4.261180743685337,
+      "objective/train/tokens_used": 272773600,
+      "objective/train/value_avg": -0.00753021240234375,
+      "objective/train/value_loss": 0.00019367090135347098,
+      "objective/train/value_max": -0.0005908012390136719,
+      "objective/train/value_min": -0.181640625,
+      "objective/train/value_reward_corr": 0.6212227796200605,
+      "objective/train/value_std": 0.008819580078125,
+      "objective/train/weight_avg": 1.0025759935379028,
+      "objective/train/weighted_lm_loss": 1.9780480861663818,
+      "objective/train/weights_max": 1.16512930393219,
+      "objective/train/weights_min": 0.7835448384284973,
+      "theoretical_loss": 4.261180743685337,
+      "tokens_seen": 252313600
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009328358208955225,
+      "loss": 0.0973,
+      "theoretical_loss": 4.260915549339879,
+      "tokens_seen": 252444672
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.00093275557695394,
+      "loss": 0.097,
+      "theoretical_loss": 4.2603856890013425,
+      "tokens_seen": 252706816
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009326753330123576,
+      "loss": 0.0991,
+      "theoretical_loss": 4.25985653174513,
+      "tokens_seen": 252968960
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009325950890707752,
+      "loss": 0.0948,
+      "theoretical_loss": 4.259328075911173,
+      "tokens_seen": 253231104
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009325148451291927,
+      "loss": 0.0989,
+      "theoretical_loss": 4.258800319845038,
+      "tokens_seen": 253493248
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009324346011876103,
+      "loss": 0.0974,
+      "theoretical_loss": 4.258273261897896,
+      "tokens_seen": 253755392
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009323543572460279,
+      "loss": 0.097,
+      "theoretical_loss": 4.257746900426506,
+      "tokens_seen": 254017536
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009322741133044456,
+      "loss": 0.0975,
+      "theoretical_loss": 4.25722123379318,
+      "tokens_seen": 254279680
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009321938693628631,
+      "loss": 0.1009,
+      "theoretical_loss": 4.256696260365768,
+      "tokens_seen": 254541824
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009321136254212808,
+      "loss": 0.0996,
+      "theoretical_loss": 4.256171978517629,
+      "tokens_seen": 254803968
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009320333814796983,
+      "loss": 0.0952,
+      "theoretical_loss": 4.255648386627607,
+      "tokens_seen": 255066112
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000931953137538116,
+      "loss": 0.0987,
+      "theoretical_loss": 4.255125483080007,
+      "tokens_seen": 255328256
+    },
+    {
+      "epoch": 0.08,
+      "objective/train/advantage_avg": 0.0014403994427993894,
+      "objective/train/docs_used": 99403,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.000425100326538,
+      "objective/train/original_loss": 2.000425338745117,
+      "objective/train/theoretical_loss": 4.254603266264572,
+      "objective/train/tokens_used": 276050400,
+      "objective/train/value_avg": -0.00920867919921875,
+      "objective/train/value_loss": 0.0005131934303790331,
+      "objective/train/value_max": -0.0005726814270019531,
+      "objective/train/value_min": -0.29736328125,
+      "objective/train/value_reward_corr": 0.55161832441128,
+      "objective/train/value_std": 0.011138916015625,
+      "objective/train/weight_avg": 1.0016599893569946,
+      "objective/train/weighted_lm_loss": 2.0043511390686035,
+      "objective/train/weights_max": 1.3048828840255737,
+      "objective/train/weights_min": 0.3703209161758423,
+      "theoretical_loss": 4.254603266264572,
+      "tokens_seen": 255590400
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009318728935965335,
+      "loss": 0.0987,
+      "theoretical_loss": 4.254603266264572,
+      "tokens_seen": 255590400
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000931792649654951,
+      "loss": 0.0957,
+      "theoretical_loss": 4.254081734576458,
+      "tokens_seen": 255852544
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009317124057133687,
+      "loss": 0.0968,
+      "theoretical_loss": 4.253560886416212,
+      "tokens_seen": 256114688
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009316321617717862,
+      "loss": 0.0972,
+      "theoretical_loss": 4.253040720189746,
+      "tokens_seen": 256376832
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009315519178302038,
+      "loss": 0.0973,
+      "theoretical_loss": 4.252521234308315,
+      "tokens_seen": 256638976
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009314716738886215,
+      "loss": 0.0967,
+      "theoretical_loss": 4.2520024271884935,
+      "tokens_seen": 256901120
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009313914299470391,
+      "loss": 0.0981,
+      "theoretical_loss": 4.251484297252151,
+      "tokens_seen": 257163264
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009313111860054566,
+      "loss": 0.1018,
+      "theoretical_loss": 4.250966842926434,
+      "tokens_seen": 257425408
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009312309420638742,
+      "loss": 0.0991,
+      "theoretical_loss": 4.250450062643734,
+      "tokens_seen": 257687552
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009311506981222918,
+      "loss": 0.0955,
+      "theoretical_loss": 4.249933954841672,
+      "tokens_seen": 257949696
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009310704541807093,
+      "loss": 0.1025,
+      "theoretical_loss": 4.2494185179630755,
+      "tokens_seen": 258211840
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000930990210239127,
+      "loss": 0.0993,
+      "theoretical_loss": 4.24890375045595,
+      "tokens_seen": 258473984
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009309099662975445,
+      "loss": 0.0974,
+      "theoretical_loss": 4.248389650773463,
+      "tokens_seen": 258736128
+    },
+    {
+      "epoch": 0.08,
+      "objective/train/advantage_avg": 0.0018602707423269749,
+      "objective/train/docs_used": 100739,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8192548751831055,
+      "objective/train/original_loss": 1.8192548751831055,
+      "objective/train/theoretical_loss": 4.24813285088448,
+      "objective/train/tokens_used": 279327200,
+      "objective/train/value_avg": -0.00965118408203125,
+      "objective/train/value_loss": 0.00019530275312718004,
+      "objective/train/value_max": -0.0006694793701171875,
+      "objective/train/value_min": -0.4150390625,
+      "objective/train/value_reward_corr": 0.596449167871979,
+      "objective/train/value_std": 0.01064300537109375,
+      "objective/train/weight_avg": 1.0019567012786865,
+      "objective/train/weighted_lm_loss": 1.8218411207199097,
+      "objective/train/weights_max": 1.1794391870498657,
+      "objective/train/weights_min": 0.8217251300811768,
+      "theoretical_loss": 4.24813285088448,
+      "tokens_seen": 258867200
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009308297223559621,
+      "loss": 0.099,
+      "theoretical_loss": 4.24787621737392,
+      "tokens_seen": 258998272
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009307494784143798,
+      "loss": 0.0972,
+      "theoretical_loss": 4.247363448720739,
+      "tokens_seen": 259260416
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009306692344727973,
+      "loss": 0.0988,
+      "theoretical_loss": 4.246851343282432,
+      "tokens_seen": 259522560
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000930588990531215,
+      "loss": 0.1011,
+      "theoretical_loss": 4.246339899532582,
+      "tokens_seen": 259784704
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009305087465896325,
+      "loss": 0.0971,
+      "theoretical_loss": 4.245829115949818,
+      "tokens_seen": 260046848
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009304285026480501,
+      "loss": 0.1012,
+      "theoretical_loss": 4.245318991017802,
+      "tokens_seen": 260308992
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009303482587064677,
+      "loss": 0.0982,
+      "theoretical_loss": 4.244809523225195,
+      "tokens_seen": 260571136
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009302680147648853,
+      "loss": 0.1,
+      "theoretical_loss": 4.244300711065646,
+      "tokens_seen": 260833280
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009301877708233028,
+      "loss": 0.0938,
+      "theoretical_loss": 4.243792553037767,
+      "tokens_seen": 261095424
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009301075268817204,
+      "loss": 0.0983,
+      "theoretical_loss": 4.243285047645106,
+      "tokens_seen": 261357568
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000930027282940138,
+      "loss": 0.1027,
+      "theoretical_loss": 4.242778193396136,
+      "tokens_seen": 261619712
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009299470389985556,
+      "loss": 0.0991,
+      "theoretical_loss": 4.242271988804228,
+      "tokens_seen": 261881856
+    },
+    {
+      "epoch": 0.08,
+      "objective/train/advantage_avg": 0.003497189376503229,
+      "objective/train/docs_used": 101934,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.0956976413726807,
+      "objective/train/original_loss": 2.0956978797912598,
+      "objective/train/theoretical_loss": 4.241766432387629,
+      "objective/train/tokens_used": 282604000,
+      "objective/train/value_avg": -0.0075836181640625,
+      "objective/train/value_loss": 0.00010112895688507706,
+      "objective/train/value_max": -0.00064849853515625,
+      "objective/train/value_min": -0.1431884765625,
+      "objective/train/value_reward_corr": 0.3128660984446543,
+      "objective/train/value_std": 0.005107879638671875,
+      "objective/train/weight_avg": 1.0035475492477417,
+      "objective/train/weighted_lm_loss": 2.104475498199463,
+      "objective/train/weights_max": 1.1146957874298096,
+      "objective/train/weights_min": 0.901778519153595,
+      "theoretical_loss": 4.241766432387629,
+      "tokens_seen": 262144000
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009298667950569733,
+      "loss": 0.0955,
+      "theoretical_loss": 4.241766432387629,
+      "tokens_seen": 262144000
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009297865511153908,
+      "loss": 0.0972,
+      "theoretical_loss": 4.241261522669445,
+      "tokens_seen": 262406144
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009297063071738084,
+      "loss": 0.0996,
+      "theoretical_loss": 4.240757258177617,
+      "tokens_seen": 262668288
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000929626063232226,
+      "loss": 0.0959,
+      "theoretical_loss": 4.240253637444903,
+      "tokens_seen": 262930432
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009295458192906435,
+      "loss": 0.0955,
+      "theoretical_loss": 4.239750659008854,
+      "tokens_seen": 263192576
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009294655753490611,
+      "loss": 0.0971,
+      "theoretical_loss": 4.2392483214117975,
+      "tokens_seen": 263454720
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009293853314074787,
+      "loss": 0.0999,
+      "theoretical_loss": 4.238746623200815,
+      "tokens_seen": 263716864
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009293050874658963,
+      "loss": 0.0985,
+      "theoretical_loss": 4.238245562927722,
+      "tokens_seen": 263979008
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000929224843524314,
+      "loss": 0.0991,
+      "theoretical_loss": 4.237745139149047,
+      "tokens_seen": 264241152
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009291445995827316,
+      "loss": 0.0964,
+      "theoretical_loss": 4.237245350426015,
+      "tokens_seen": 264503296
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009290643556411491,
+      "loss": 0.0942,
+      "theoretical_loss": 4.236746195324523,
+      "tokens_seen": 264765440
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009289841116995667,
+      "loss": 0.1008,
+      "theoretical_loss": 4.2362476724151215,
+      "tokens_seen": 265027584
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009289038677579843,
+      "loss": 0.0977,
+      "theoretical_loss": 4.235749780272998,
+      "tokens_seen": 265289728
+    },
+    {
+      "epoch": 0.08,
+      "objective/train/advantage_avg": 0.0023133521899580956,
+      "objective/train/docs_used": 103026,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7841287851333618,
+      "objective/train/original_loss": 1.7841286659240723,
+      "objective/train/theoretical_loss": 4.235501070295674,
+      "objective/train/tokens_used": 285880800,
+      "objective/train/value_avg": -0.007236480712890625,
+      "objective/train/value_loss": 0.00020630829385481775,
+      "objective/train/value_max": -0.0005254745483398438,
+      "objective/train/value_min": -0.1910400390625,
+      "objective/train/value_reward_corr": 0.3257111556668698,
+      "objective/train/value_std": 0.00516510009765625,
+      "objective/train/weight_avg": 1.0024032592773438,
+      "objective/train/weighted_lm_loss": 1.7887521982192993,
+      "objective/train/weights_max": 1.092002034187317,
+      "objective/train/weights_min": 0.36931154131889343,
+      "theoretical_loss": 4.235501070295674,
+      "tokens_seen": 265420800
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009288236238164018,
+      "loss": 0.1003,
+      "theoretical_loss": 4.235252517477956,
+      "tokens_seen": 265551872
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009287433798748195,
+      "loss": 0.0946,
+      "theoretical_loss": 4.23475588261439,
+      "tokens_seen": 265814016
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000928663135933237,
+      "loss": 0.0971,
+      "theoretical_loss": 4.234259874271275,
+      "tokens_seen": 266076160
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009285828919916546,
+      "loss": 0.0972,
+      "theoretical_loss": 4.23376449104214,
+      "tokens_seen": 266338304
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009285026480500723,
+      "loss": 0.1004,
+      "theoretical_loss": 4.233269731525055,
+      "tokens_seen": 266600448
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009284224041084899,
+      "loss": 0.0981,
+      "theoretical_loss": 4.232775594322605,
+      "tokens_seen": 266862592
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009283421601669074,
+      "loss": 0.0975,
+      "theoretical_loss": 4.232282078041876,
+      "tokens_seen": 267124736
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000928261916225325,
+      "loss": 0.0978,
+      "theoretical_loss": 4.231789181294436,
+      "tokens_seen": 267386880
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009281816722837426,
+      "loss": 0.1,
+      "theoretical_loss": 4.231296902696314,
+      "tokens_seen": 267649024
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009281014283421601,
+      "loss": 0.0954,
+      "theoretical_loss": 4.230805240867982,
+      "tokens_seen": 267911168
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009280211844005778,
+      "loss": 0.0972,
+      "theoretical_loss": 4.230314194434336,
+      "tokens_seen": 268173312
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009279409404589953,
+      "loss": 0.0976,
+      "theoretical_loss": 4.229823762024681,
+      "tokens_seen": 268435456
+    },
+    {
+      "epoch": 0.08,
+      "objective/train/advantage_avg": 0.001721393782645464,
+      "objective/train/docs_used": 104173,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.79710054397583,
+      "objective/train/original_loss": 1.79710054397583,
+      "objective/train/theoretical_loss": 4.2293339422727065,
+      "objective/train/tokens_used": 289157600,
+      "objective/train/value_avg": -0.00934600830078125,
+      "objective/train/value_loss": 0.00033259327756240964,
+      "objective/train/value_max": -0.0006589889526367188,
+      "objective/train/value_min": -0.2215576171875,
+      "objective/train/value_reward_corr": 0.6127857051291834,
+      "objective/train/value_std": 0.0121307373046875,
+      "objective/train/weight_avg": 1.0018748044967651,
+      "objective/train/weighted_lm_loss": 1.7999058961868286,
+      "objective/train/weights_max": 1.2480190992355347,
+      "objective/train/weights_min": 0.40334662795066833,
+      "theoretical_loss": 4.2293339422727065,
+      "tokens_seen": 268697600
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000927860696517413,
+      "loss": 0.0974,
+      "theoretical_loss": 4.2293339422727065,
+      "tokens_seen": 268697600
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009277804525758306,
+      "loss": 0.0988,
+      "theoretical_loss": 4.228844733816474,
+      "tokens_seen": 268959744
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009277002086342481,
+      "loss": 0.0948,
+      "theoretical_loss": 4.228356135298394,
+      "tokens_seen": 269221888
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009276199646926658,
+      "loss": 0.0941,
+      "theoretical_loss": 4.227868145365211,
+      "tokens_seen": 269484032
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009275397207510833,
+      "loss": 0.0966,
+      "theoretical_loss": 4.227380762667987,
+      "tokens_seen": 269746176
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009274594768095009,
+      "loss": 0.0953,
+      "theoretical_loss": 4.226893985862076,
+      "tokens_seen": 270008320
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009273792328679185,
+      "loss": 0.099,
+      "theoretical_loss": 4.226407813607116,
+      "tokens_seen": 270270464
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009272989889263361,
+      "loss": 0.0969,
+      "theoretical_loss": 4.2259222445670055,
+      "tokens_seen": 270532608
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009272187449847536,
+      "loss": 0.0936,
+      "theoretical_loss": 4.225437277409885,
+      "tokens_seen": 270794752
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009271385010431712,
+      "loss": 0.0955,
+      "theoretical_loss": 4.224952910808122,
+      "tokens_seen": 271056896
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009270582571015889,
+      "loss": 0.0976,
+      "theoretical_loss": 4.224469143438294,
+      "tokens_seen": 271319040
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009269780131600064,
+      "loss": 0.1001,
+      "theoretical_loss": 4.223985973981171,
+      "tokens_seen": 271581184
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009268977692184241,
+      "loss": 0.0996,
+      "theoretical_loss": 4.223503401121693,
+      "tokens_seen": 271843328
+    },
+    {
+      "epoch": 0.08,
+      "objective/train/advantage_avg": 0.002592921955510974,
+      "objective/train/docs_used": 105443,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8855316638946533,
+      "objective/train/original_loss": 1.8855319023132324,
+      "objective/train/theoretical_loss": 4.223262338006254,
+      "objective/train/tokens_used": 292434400,
+      "objective/train/value_avg": -0.00720977783203125,
+      "objective/train/value_loss": 0.00012640572094824165,
+      "objective/train/value_max": -0.00036835670471191406,
+      "objective/train/value_min": -0.2186279296875,
+      "objective/train/value_reward_corr": 0.5305388442877321,
+      "objective/train/value_std": 0.00785064697265625,
+      "objective/train/weight_avg": 1.0026540756225586,
+      "objective/train/weighted_lm_loss": 1.8909047842025757,
+      "objective/train/weights_max": 1.1803033351898193,
+      "objective/train/weights_min": 0.4927148222923279,
+      "theoretical_loss": 4.223262338006254,
+      "tokens_seen": 271974400
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009268175252768416,
+      "loss": 0.0944,
+      "theoretical_loss": 4.223021423548962,
+      "tokens_seen": 272105472
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009267372813352593,
+      "loss": 0.0964,
+      "theoretical_loss": 4.222540039956215,
+      "tokens_seen": 272367616
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009266570373936768,
+      "loss": 0.0953,
+      "theoretical_loss": 4.222059249040814,
+      "tokens_seen": 272629760
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009265767934520943,
+      "loss": 0.0991,
+      "theoretical_loss": 4.2215790495042285,
+      "tokens_seen": 272891904
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000926496549510512,
+      "loss": 0.0986,
+      "theoretical_loss": 4.221099440052014,
+      "tokens_seen": 273154048
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009264163055689295,
+      "loss": 0.0959,
+      "theoretical_loss": 4.220620419393799,
+      "tokens_seen": 273416192
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009263360616273471,
+      "loss": 0.0982,
+      "theoretical_loss": 4.220141986243268,
+      "tokens_seen": 273678336
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009262558176857648,
+      "loss": 0.0984,
+      "theoretical_loss": 4.219664139318145,
+      "tokens_seen": 273940480
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009261755737441824,
+      "loss": 0.0973,
+      "theoretical_loss": 4.219186877340174,
+      "tokens_seen": 274202624
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009260953298025999,
+      "loss": 0.0947,
+      "theoretical_loss": 4.218710199035108,
+      "tokens_seen": 274464768
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009260150858610175,
+      "loss": 0.0947,
+      "theoretical_loss": 4.218234103132686,
+      "tokens_seen": 274726912
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009259348419194351,
+      "loss": 0.0974,
+      "theoretical_loss": 4.217758588366623,
+      "tokens_seen": 274989056
+    },
+    {
+      "epoch": 0.08,
+      "objective/train/advantage_avg": 0.002839514520019293,
+      "objective/train/docs_used": 106703,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8725626468658447,
+      "objective/train/original_loss": 1.8725626468658447,
+      "objective/train/theoretical_loss": 4.217283653474588,
+      "objective/train/tokens_used": 295711200,
+      "objective/train/value_avg": -0.01064300537109375,
+      "objective/train/value_loss": 0.00048408194561488926,
+      "objective/train/value_max": -0.0005173683166503906,
+      "objective/train/value_min": -0.8681640625,
+      "objective/train/value_reward_corr": 0.614751588044035,
+      "objective/train/value_std": 0.016937255859375,
+      "objective/train/weight_avg": 1.003056287765503,
+      "objective/train/weighted_lm_loss": 1.8782776594161987,
+      "objective/train/weights_max": 1.6007287502288818,
+      "objective/train/weights_min": 0.3700864911079407,
+      "theoretical_loss": 4.217283653474588,
+      "tokens_seen": 275251200
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009258545979778526,
+      "loss": 0.094,
+      "theoretical_loss": 4.217283653474588,
+      "tokens_seen": 275251200
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009257743540362703,
+      "loss": 0.0967,
+      "theoretical_loss": 4.216809297198195,
+      "tokens_seen": 275513344
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009256941100946878,
+      "loss": 0.0972,
+      "theoretical_loss": 4.21633551828298,
+      "tokens_seen": 275775488
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009256138661531054,
+      "loss": 0.0984,
+      "theoretical_loss": 4.215862315478388,
+      "tokens_seen": 276037632
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009255336222115231,
+      "loss": 0.0994,
+      "theoretical_loss": 4.2153896875377574,
+      "tokens_seen": 276299776
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009254533782699407,
+      "loss": 0.0979,
+      "theoretical_loss": 4.214917633218304,
+      "tokens_seen": 276561920
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009253731343283583,
+      "loss": 0.098,
+      "theoretical_loss": 4.214446151281106,
+      "tokens_seen": 276824064
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009252928903867758,
+      "loss": 0.0947,
+      "theoretical_loss": 4.213975240491084,
+      "tokens_seen": 277086208
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009252126464451934,
+      "loss": 0.0965,
+      "theoretical_loss": 4.213504899616995,
+      "tokens_seen": 277348352
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.000925132402503611,
+      "loss": 0.0967,
+      "theoretical_loss": 4.213035127431402,
+      "tokens_seen": 277610496
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009250521585620286,
+      "loss": 0.0971,
+      "theoretical_loss": 4.212565922710677,
+      "tokens_seen": 277872640
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009249719146204461,
+      "loss": 0.0967,
+      "theoretical_loss": 4.21209728423497,
+      "tokens_seen": 278134784
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009248916706788639,
+      "loss": 0.0989,
+      "theoretical_loss": 4.2116292107882,
+      "tokens_seen": 278396928
+    },
+    {
+      "epoch": 0.08,
+      "objective/train/advantage_avg": 0.0019138733623549342,
+      "objective/train/docs_used": 107929,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.952666163444519,
+      "objective/train/original_loss": 1.9526660442352295,
+      "objective/train/theoretical_loss": 4.211395385571668,
+      "objective/train/tokens_used": 298988000,
+      "objective/train/value_avg": -0.00894927978515625,
+      "objective/train/value_loss": 0.0007483014487661421,
+      "objective/train/value_max": -0.00033664703369140625,
+      "objective/train/value_min": -0.8779296875,
+      "objective/train/value_reward_corr": 0.5763799754051704,
+      "objective/train/value_std": 0.0159759521484375,
+      "objective/train/weight_avg": 1.002224087715149,
+      "objective/train/weighted_lm_loss": 1.956479549407959,
+      "objective/train/weights_max": 1.5199860334396362,
+      "objective/train/weights_min": 0.37053996324539185,
+      "theoretical_loss": 4.211395385571668,
+      "tokens_seen": 278528000
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009248114267372814,
+      "loss": 0.0977,
+      "theoretical_loss": 4.211161701158042,
+      "tokens_seen": 278659072
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009247311827956989,
+      "loss": 0.0988,
+      "theoretical_loss": 4.2106947541359085,
+      "tokens_seen": 278921216
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009246509388541166,
+      "loss": 0.0962,
+      "theoretical_loss": 4.210228368516935,
+      "tokens_seen": 279183360
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009245706949125341,
+      "loss": 0.1018,
+      "theoretical_loss": 4.209762543099966,
+      "tokens_seen": 279445504
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009244904509709517,
+      "loss": 0.0973,
+      "theoretical_loss": 4.209297276687541,
+      "tokens_seen": 279707648
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009244102070293693,
+      "loss": 0.1001,
+      "theoretical_loss": 4.2088325680858745,
+      "tokens_seen": 279969792
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 0.0009243299630877869,
+      "loss": 0.0956,
+      "theoretical_loss": 4.208368416104849,
+      "tokens_seen": 280231936
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009242497191462045,
+      "loss": 0.0991,
+      "theoretical_loss": 4.207904819557995,
+      "tokens_seen": 280494080
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.000924169475204622,
+      "loss": 0.0967,
+      "theoretical_loss": 4.207441777262477,
+      "tokens_seen": 280756224
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009240892312630397,
+      "loss": 0.0966,
+      "theoretical_loss": 4.206979288039081,
+      "tokens_seen": 281018368
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009240089873214573,
+      "loss": 0.0953,
+      "theoretical_loss": 4.206517350712199,
+      "tokens_seen": 281280512
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009239287433798749,
+      "loss": 0.0974,
+      "theoretical_loss": 4.206055964109813,
+      "tokens_seen": 281542656
+    },
+    {
+      "epoch": 0.09,
+      "objective/train/advantage_avg": 0.002816372085362673,
+      "objective/train/docs_used": 109222,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8340928554534912,
+      "objective/train/original_loss": 1.8340927362442017,
+      "objective/train/theoretical_loss": 4.205595127063485,
+      "objective/train/tokens_used": 302264800,
+      "objective/train/value_avg": -0.00853729248046875,
+      "objective/train/value_loss": 0.0001849338150350377,
+      "objective/train/value_max": -0.0005483627319335938,
+      "objective/train/value_min": -0.266357421875,
+      "objective/train/value_reward_corr": 0.5370670101322965,
+      "objective/train/value_std": 0.0100250244140625,
+      "objective/train/weight_avg": 1.0029041767120361,
+      "objective/train/weighted_lm_loss": 1.8380697965621948,
+      "objective/train/weights_max": 1.208823800086975,
+      "objective/train/weights_min": 0.3825574517250061,
+      "theoretical_loss": 4.205595127063485,
+      "tokens_seen": 281804800
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009238484994382924,
+      "loss": 0.0958,
+      "theoretical_loss": 4.205595127063485,
+      "tokens_seen": 281804800
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009237682554967101,
+      "loss": 0.0947,
+      "theoretical_loss": 4.205134838408337,
+      "tokens_seen": 282066944
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009236880115551276,
+      "loss": 0.0966,
+      "theoretical_loss": 4.20467509698304,
+      "tokens_seen": 282329088
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009236077676135451,
+      "loss": 0.0962,
+      "theoretical_loss": 4.204215901629803,
+      "tokens_seen": 282591232
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009235275236719628,
+      "loss": 0.0969,
+      "theoretical_loss": 4.203757251194353,
+      "tokens_seen": 282853376
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009234472797303803,
+      "loss": 0.096,
+      "theoretical_loss": 4.203299144525923,
+      "tokens_seen": 283115520
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.000923367035788798,
+      "loss": 0.0966,
+      "theoretical_loss": 4.202841580477241,
+      "tokens_seen": 283377664
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009232867918472156,
+      "loss": 0.0995,
+      "theoretical_loss": 4.202384557904513,
+      "tokens_seen": 283639808
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009232065479056332,
+      "loss": 0.096,
+      "theoretical_loss": 4.201928075667411,
+      "tokens_seen": 283901952
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009231263039640507,
+      "loss": 0.0955,
+      "theoretical_loss": 4.201472132629057,
+      "tokens_seen": 284164096
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009230460600224683,
+      "loss": 0.1011,
+      "theoretical_loss": 4.201016727656012,
+      "tokens_seen": 284426240
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009229658160808859,
+      "loss": 0.0984,
+      "theoretical_loss": 4.2005618596182615,
+      "tokens_seen": 284688384
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009228855721393035,
+      "loss": 0.0986,
+      "theoretical_loss": 4.200107527389202,
+      "tokens_seen": 284950528
+    },
+    {
+      "epoch": 0.09,
+      "objective/train/advantage_avg": 0.00423348369076848,
+      "objective/train/docs_used": 110428,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.866652250289917,
+      "objective/train/original_loss": 1.866652488708496,
+      "objective/train/theoretical_loss": 4.1998805618517965,
+      "objective/train/tokens_used": 305541600,
+      "objective/train/value_avg": -0.00862884521484375,
+      "objective/train/value_loss": 0.00014239439042285085,
+      "objective/train/value_max": -0.0005507469177246094,
+      "objective/train/value_min": -0.1759033203125,
+      "objective/train/value_reward_corr": 0.4883459042398012,
+      "objective/train/value_std": 0.0081787109375,
+      "objective/train/weight_avg": 1.0043009519577026,
+      "objective/train/weighted_lm_loss": 1.8749480247497559,
+      "objective/train/weights_max": 1.1676921844482422,
+      "objective/train/weights_min": 0.3836681544780731,
+      "theoretical_loss": 4.1998805618517965,
+      "tokens_seen": 285081600
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009228053281977211,
+      "loss": 0.0967,
+      "theoretical_loss": 4.199653729845626,
+      "tokens_seen": 285212672
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009227250842561386,
+      "loss": 0.0968,
+      "theoretical_loss": 4.199200465867714,
+      "tokens_seen": 285474816
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009226448403145564,
+      "loss": 0.0961,
+      "theoretical_loss": 4.198747734339013,
+      "tokens_seen": 285736960
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009225645963729739,
+      "loss": 0.0945,
+      "theoretical_loss": 4.198295534146429,
+      "tokens_seen": 285999104
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009224843524313914,
+      "loss": 0.0952,
+      "theoretical_loss": 4.197843864180214,
+      "tokens_seen": 286261248
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009224041084898091,
+      "loss": 0.0955,
+      "theoretical_loss": 4.197392723333951,
+      "tokens_seen": 286523392
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009223238645482266,
+      "loss": 0.098,
+      "theoretical_loss": 4.196942110504538,
+      "tokens_seen": 286785536
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009222436206066442,
+      "loss": 0.0955,
+      "theoretical_loss": 4.196492024592183,
+      "tokens_seen": 287047680
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009221633766650618,
+      "loss": 0.0951,
+      "theoretical_loss": 4.196042464500382,
+      "tokens_seen": 287309824
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009220831327234794,
+      "loss": 0.0948,
+      "theoretical_loss": 4.195593429135916,
+      "tokens_seen": 287571968
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009220028887818969,
+      "loss": 0.1016,
+      "theoretical_loss": 4.195144917408828,
+      "tokens_seen": 287834112
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009219226448403146,
+      "loss": 0.095,
+      "theoretical_loss": 4.194696928232417,
+      "tokens_seen": 288096256
+    },
+    {
+      "epoch": 0.09,
+      "objective/train/advantage_avg": 0.003713709767907858,
+      "objective/train/docs_used": 111591,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8504618406295776,
+      "objective/train/original_loss": 1.850461721420288,
+      "objective/train/theoretical_loss": 4.194249460523222,
+      "objective/train/tokens_used": 308818400,
+      "objective/train/value_avg": -0.0093841552734375,
+      "objective/train/value_loss": 0.0001626669109100476,
+      "objective/train/value_max": -0.00070953369140625,
+      "objective/train/value_min": -0.247802734375,
+      "objective/train/value_reward_corr": 0.5734128643584034,
+      "objective/train/value_std": 0.01049041748046875,
+      "objective/train/weight_avg": 1.0037906169891357,
+      "objective/train/weighted_lm_loss": 1.857253909111023,
+      "objective/train/weights_max": 1.1863479614257812,
+      "objective/train/weights_min": 0.3687554597854614,
+      "theoretical_loss": 4.194249460523222,
+      "tokens_seen": 288358400
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009218424008987322,
+      "loss": 0.0936,
+      "theoretical_loss": 4.194249460523222,
+      "tokens_seen": 288358400
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009217621569571497,
+      "loss": 0.0967,
+      "theoretical_loss": 4.193802513201015,
+      "tokens_seen": 288620544
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009216819130155674,
+      "loss": 0.0998,
+      "theoretical_loss": 4.193356085188778,
+      "tokens_seen": 288882688
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009216016690739849,
+      "loss": 0.0987,
+      "theoretical_loss": 4.1929101754127025,
+      "tokens_seen": 289144832
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009215214251324026,
+      "loss": 0.0962,
+      "theoretical_loss": 4.192464782802167,
+      "tokens_seen": 289406976
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009214411811908201,
+      "loss": 0.096,
+      "theoretical_loss": 4.192019906289733,
+      "tokens_seen": 289669120
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009213609372492377,
+      "loss": 0.099,
+      "theoretical_loss": 4.1915755448111245,
+      "tokens_seen": 289931264
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009212806933076553,
+      "loss": 0.0974,
+      "theoretical_loss": 4.191131697305222,
+      "tokens_seen": 290193408
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009212004493660728,
+      "loss": 0.0959,
+      "theoretical_loss": 4.1906883627140505,
+      "tokens_seen": 290455552
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009211202054244905,
+      "loss": 0.0929,
+      "theoretical_loss": 4.19024553998276,
+      "tokens_seen": 290717696
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009210399614829081,
+      "loss": 0.0956,
+      "theoretical_loss": 4.189803228059623,
+      "tokens_seen": 290979840
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009209597175413257,
+      "loss": 0.096,
+      "theoretical_loss": 4.189361425896016,
+      "tokens_seen": 291241984
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009208794735997432,
+      "loss": 0.0929,
+      "theoretical_loss": 4.188920132446411,
+      "tokens_seen": 291504128
+    },
+    {
+      "epoch": 0.09,
+      "objective/train/advantage_avg": 0.0008069810573942959,
+      "objective/train/docs_used": 112844,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8364742994308472,
+      "objective/train/original_loss": 1.8364744186401367,
+      "objective/train/theoretical_loss": 4.188699676163473,
+      "objective/train/tokens_used": 312095200,
+      "objective/train/value_avg": -0.008056640625,
+      "objective/train/value_loss": 0.00030392000917345285,
+      "objective/train/value_max": -0.0004012584686279297,
+      "objective/train/value_min": -0.474609375,
+      "objective/train/value_reward_corr": 0.44846135234212026,
+      "objective/train/value_std": 0.00936126708984375,
+      "objective/train/weight_avg": 1.0009410381317139,
+      "objective/train/weighted_lm_loss": 1.8375290632247925,
+      "objective/train/weights_max": 1.5633509159088135,
+      "objective/train/weights_min": 0.3711496591567993,
+      "theoretical_loss": 4.188699676163473,
+      "tokens_seen": 291635200
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009207992296581609,
+      "loss": 0.0937,
+      "theoretical_loss": 4.188479346668359,
+      "tokens_seen": 291766272
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009207189857165784,
+      "loss": 0.0943,
+      "theoretical_loss": 4.188039067522484,
+      "tokens_seen": 292028416
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009206387417749959,
+      "loss": 0.0952,
+      "theoretical_loss": 4.18759929397247,
+      "tokens_seen": 292290560
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009205584978334136,
+      "loss": 0.0959,
+      "theoretical_loss": 4.187160024985044,
+      "tokens_seen": 292552704
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009204782538918311,
+      "loss": 0.0975,
+      "theoretical_loss": 4.1867212595299685,
+      "tokens_seen": 292814848
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009203980099502489,
+      "loss": 0.0961,
+      "theoretical_loss": 4.186282996580034,
+      "tokens_seen": 293076992
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009203177660086664,
+      "loss": 0.0984,
+      "theoretical_loss": 4.185845235111037,
+      "tokens_seen": 293339136
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.000920237522067084,
+      "loss": 0.0962,
+      "theoretical_loss": 4.185407974101779,
+      "tokens_seen": 293601280
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009201572781255016,
+      "loss": 0.095,
+      "theoretical_loss": 4.184971212534048,
+      "tokens_seen": 293863424
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009200770341839191,
+      "loss": 0.096,
+      "theoretical_loss": 4.184534949392611,
+      "tokens_seen": 294125568
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009199967902423367,
+      "loss": 0.0951,
+      "theoretical_loss": 4.184099183665199,
+      "tokens_seen": 294387712
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009199165463007543,
+      "loss": 0.0985,
+      "theoretical_loss": 4.1836639143425,
+      "tokens_seen": 294649856
+    },
+    {
+      "epoch": 0.09,
+      "objective/train/advantage_avg": 0.002410832792520523,
+      "objective/train/docs_used": 114046,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9235732555389404,
+      "objective/train/original_loss": 1.92357337474823,
+      "objective/train/theoretical_loss": 4.1832291404181445,
+      "objective/train/tokens_used": 315372000,
+      "objective/train/value_avg": -0.00838470458984375,
+      "objective/train/value_loss": 0.00013771136582363397,
+      "objective/train/value_max": -0.0008459091186523438,
+      "objective/train/value_min": -0.28466796875,
+      "objective/train/value_reward_corr": 0.5819581613531781,
+      "objective/train/value_std": 0.0088043212890625,
+      "objective/train/weight_avg": 1.0024785995483398,
+      "objective/train/weighted_lm_loss": 1.928615689277649,
+      "objective/train/weights_max": 1.112996220588684,
+      "objective/train/weights_min": 0.6144814491271973,
+      "theoretical_loss": 4.1832291404181445,
+      "tokens_seen": 294912000
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009198363023591719,
+      "loss": 0.0955,
+      "theoretical_loss": 4.1832291404181445,
+      "tokens_seen": 294912000
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009197560584175894,
+      "loss": 0.0958,
+      "theoretical_loss": 4.182794860888696,
+      "tokens_seen": 295174144
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009196758144760072,
+      "loss": 0.0974,
+      "theoretical_loss": 4.18236107475364,
+      "tokens_seen": 295436288
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009195955705344247,
+      "loss": 0.0974,
+      "theoretical_loss": 4.18192778101537,
+      "tokens_seen": 295698432
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009195153265928422,
+      "loss": 0.0943,
+      "theoretical_loss": 4.181494978679181,
+      "tokens_seen": 295960576
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009194350826512599,
+      "loss": 0.0952,
+      "theoretical_loss": 4.181062666753256,
+      "tokens_seen": 296222720
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009193548387096774,
+      "loss": 0.0945,
+      "theoretical_loss": 4.180630844248653,
+      "tokens_seen": 296484864
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.000919274594768095,
+      "loss": 0.0977,
+      "theoretical_loss": 4.180199510179299,
+      "tokens_seen": 296747008
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009191943508265126,
+      "loss": 0.0951,
+      "theoretical_loss": 4.179768663561975,
+      "tokens_seen": 297009152
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009191141068849302,
+      "loss": 0.0925,
+      "theoretical_loss": 4.1793383034163085,
+      "tokens_seen": 297271296
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009190338629433478,
+      "loss": 0.0942,
+      "theoretical_loss": 4.178908428764759,
+      "tokens_seen": 297533440
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009189536190017654,
+      "loss": 0.0948,
+      "theoretical_loss": 4.17847903863261,
+      "tokens_seen": 297795584
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.000918873375060183,
+      "loss": 0.0953,
+      "theoretical_loss": 4.178050132047958,
+      "tokens_seen": 298057728
+    },
+    {
+      "epoch": 0.09,
+      "objective/train/advantage_avg": 0.0007046199752949178,
+      "objective/train/docs_used": 115247,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8377935886383057,
+      "objective/train/original_loss": 1.8377935886383057,
+      "objective/train/theoretical_loss": 4.1778358597829905,
+      "objective/train/tokens_used": 318648800,
+      "objective/train/value_avg": -0.0088348388671875,
+      "objective/train/value_loss": 0.0002799044887069613,
+      "objective/train/value_max": -0.00047469139099121094,
+      "objective/train/value_min": -0.80615234375,
+      "objective/train/value_reward_corr": 0.4904552822474981,
+      "objective/train/value_std": 0.01067352294921875,
+      "objective/train/weight_avg": 1.0008306503295898,
+      "objective/train/weighted_lm_loss": 1.8391979932785034,
+      "objective/train/weights_max": 1.235874056816101,
+      "objective/train/weights_min": 0.37098827958106995,
+      "theoretical_loss": 4.1778358597829905,
+      "tokens_seen": 298188800
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009187931311186006,
+      "loss": 0.0941,
+      "theoretical_loss": 4.177621708041703,
+      "tokens_seen": 298319872
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009187128871770182,
+      "loss": 0.0969,
+      "theoretical_loss": 4.177193765647534,
+      "tokens_seen": 298582016
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009186326432354357,
+      "loss": 0.0977,
+      "theoretical_loss": 4.176766303901922,
+      "tokens_seen": 298844160
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009185523992938534,
+      "loss": 0.0959,
+      "theoretical_loss": 4.17633932184411,
+      "tokens_seen": 299106304
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009184721553522709,
+      "loss": 0.0955,
+      "theoretical_loss": 4.1759128185161005,
+      "tokens_seen": 299368448
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009183919114106885,
+      "loss": 0.0921,
+      "theoretical_loss": 4.175486792962646,
+      "tokens_seen": 299630592
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009183116674691061,
+      "loss": 0.0956,
+      "theoretical_loss": 4.175061244231237,
+      "tokens_seen": 299892736
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009182314235275236,
+      "loss": 0.0972,
+      "theoretical_loss": 4.174636171372097,
+      "tokens_seen": 300154880
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009181511795859412,
+      "loss": 0.0917,
+      "theoretical_loss": 4.174211573438166,
+      "tokens_seen": 300417024
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009180709356443589,
+      "loss": 0.0939,
+      "theoretical_loss": 4.173787449485094,
+      "tokens_seen": 300679168
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009179906917027765,
+      "loss": 0.098,
+      "theoretical_loss": 4.17336379857123,
+      "tokens_seen": 300941312
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009179104477611941,
+      "loss": 0.0945,
+      "theoretical_loss": 4.172940619757611,
+      "tokens_seen": 301203456
+    },
+    {
+      "epoch": 0.09,
+      "objective/train/advantage_avg": 0.0009184295777231455,
+      "objective/train/docs_used": 116418,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7909634113311768,
+      "objective/train/original_loss": 1.7909634113311768,
+      "objective/train/theoretical_loss": 4.172517912107954,
+      "objective/train/tokens_used": 321925600,
+      "objective/train/value_avg": -0.0111846923828125,
+      "objective/train/value_loss": 0.00032702440512366593,
+      "objective/train/value_max": -0.0006213188171386719,
+      "objective/train/value_min": -0.42529296875,
+      "objective/train/value_reward_corr": 0.6038215415930122,
+      "objective/train/value_std": 0.013702392578125,
+      "objective/train/weight_avg": 1.0010701417922974,
+      "objective/train/weighted_lm_loss": 1.792219877243042,
+      "objective/train/weights_max": 1.2377305030822754,
+      "objective/train/weights_min": 0.3794703483581543,
+      "theoretical_loss": 4.172517912107954,
+      "tokens_seen": 301465600
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009178302038196117,
+      "loss": 0.0937,
+      "theoretical_loss": 4.172517912107954,
+      "tokens_seen": 301465600
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009177499598780292,
+      "loss": 0.0927,
+      "theoretical_loss": 4.172095674688645,
+      "tokens_seen": 301727744
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009176697159364468,
+      "loss": 0.0956,
+      "theoretical_loss": 4.171673906568729,
+      "tokens_seen": 301989888
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009175894719948644,
+      "loss": 0.0972,
+      "theoretical_loss": 4.171252606819899,
+      "tokens_seen": 302252032
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009175092280532819,
+      "loss": 0.0951,
+      "theoretical_loss": 4.170831774516489,
+      "tokens_seen": 302514176
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009174289841116997,
+      "loss": 0.0955,
+      "theoretical_loss": 4.170411408735461,
+      "tokens_seen": 302776320
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009173487401701172,
+      "loss": 0.0954,
+      "theoretical_loss": 4.169991508556398,
+      "tokens_seen": 303038464
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009172684962285348,
+      "loss": 0.0944,
+      "theoretical_loss": 4.169572073061493,
+      "tokens_seen": 303300608
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009171882522869524,
+      "loss": 0.0975,
+      "theoretical_loss": 4.16915310133554,
+      "tokens_seen": 303562752
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009171080083453699,
+      "loss": 0.0963,
+      "theoretical_loss": 4.1687345924659205,
+      "tokens_seen": 303824896
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009170277644037875,
+      "loss": 0.094,
+      "theoretical_loss": 4.168316545542602,
+      "tokens_seen": 304087040
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009169475204622051,
+      "loss": 0.0935,
+      "theoretical_loss": 4.167898959658121,
+      "tokens_seen": 304349184
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009168672765206227,
+      "loss": 0.0902,
+      "theoretical_loss": 4.167481833907576,
+      "tokens_seen": 304611328
+    },
+    {
+      "epoch": 0.09,
+      "objective/train/advantage_avg": 0.0005431807949207723,
+      "objective/train/docs_used": 117554,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7858117818832397,
+      "objective/train/original_loss": 1.7858116626739502,
+      "objective/train/theoretical_loss": 4.167273443300447,
+      "objective/train/tokens_used": 325202400,
+      "objective/train/value_avg": -0.0095367431640625,
+      "objective/train/value_loss": 0.00023923047410789877,
+      "objective/train/value_max": -0.0003447532653808594,
+      "objective/train/value_min": -0.414794921875,
+      "objective/train/value_reward_corr": 0.6826350512891199,
+      "objective/train/value_std": 0.01375579833984375,
+      "objective/train/weight_avg": 1.0006574392318726,
+      "objective/train/weighted_lm_loss": 1.7867764234542847,
+      "objective/train/weights_max": 1.3765175342559814,
+      "objective/train/weights_min": 0.3716852366924286,
+      "theoretical_loss": 4.167273443300447,
+      "tokens_seen": 304742400
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009167870325790402,
+      "loss": 0.0982,
+      "theoretical_loss": 4.16706516738862,
+      "tokens_seen": 304873472
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.000916706788637458,
+      "loss": 0.0941,
+      "theoretical_loss": 4.166648959201449,
+      "tokens_seen": 305135616
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009166265446958755,
+      "loss": 0.0978,
+      "theoretical_loss": 4.166233208448794,
+      "tokens_seen": 305397760
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009165463007542931,
+      "loss": 0.0968,
+      "theoretical_loss": 4.165817914235908,
+      "tokens_seen": 305659904
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009164660568127107,
+      "loss": 0.0952,
+      "theoretical_loss": 4.165403075670562,
+      "tokens_seen": 305922048
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009163858128711282,
+      "loss": 0.0942,
+      "theoretical_loss": 4.164988691863032,
+      "tokens_seen": 306184192
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009163055689295459,
+      "loss": 0.0944,
+      "theoretical_loss": 4.164574761926092,
+      "tokens_seen": 306446336
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009162253249879634,
+      "loss": 0.0949,
+      "theoretical_loss": 4.164161284975005,
+      "tokens_seen": 306708480
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.000916145081046381,
+      "loss": 0.0941,
+      "theoretical_loss": 4.1637482601275115,
+      "tokens_seen": 306970624
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009160648371047986,
+      "loss": 0.0983,
+      "theoretical_loss": 4.163335686503822,
+      "tokens_seen": 307232768
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009159845931632162,
+      "loss": 0.0961,
+      "theoretical_loss": 4.162923563226607,
+      "tokens_seen": 307494912
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009159043492216338,
+      "loss": 0.0935,
+      "theoretical_loss": 4.1625118894209905,
+      "tokens_seen": 307757056
+    },
+    {
+      "epoch": 0.09,
+      "objective/train/advantage_avg": 0.0024794533383101225,
+      "objective/train/docs_used": 118729,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.930225133895874,
+      "objective/train/original_loss": 1.930225133895874,
+      "objective/train/theoretical_loss": 4.16210066421454,
+      "objective/train/tokens_used": 328479200,
+      "objective/train/value_avg": -0.01091766357421875,
+      "objective/train/value_loss": 0.0003562370839063078,
+      "objective/train/value_max": -0.0003101825714111328,
+      "objective/train/value_min": -0.7587890625,
+      "objective/train/value_reward_corr": 0.711995979515493,
+      "objective/train/value_std": 0.02008056640625,
+      "objective/train/weight_avg": 1.002644658088684,
+      "objective/train/weighted_lm_loss": 1.9346448183059692,
+      "objective/train/weights_max": 1.8252228498458862,
+      "objective/train/weights_min": 0.36886805295944214,
+      "theoretical_loss": 4.16210066421454,
+      "tokens_seen": 308019200
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009158241052800514,
+      "loss": 0.0936,
+      "theoretical_loss": 4.16210066421454,
+      "tokens_seen": 308019200
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.000915743861338469,
+      "loss": 0.0949,
+      "theoretical_loss": 4.161689886737255,
+      "tokens_seen": 308281344
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009156636173968865,
+      "loss": 0.0969,
+      "theoretical_loss": 4.161279556121562,
+      "tokens_seen": 308543488
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009155833734553042,
+      "loss": 0.0955,
+      "theoretical_loss": 4.160869671502302,
+      "tokens_seen": 308805632
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009155031295137217,
+      "loss": 0.0963,
+      "theoretical_loss": 4.160460232016725,
+      "tokens_seen": 309067776
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009154228855721394,
+      "loss": 0.0969,
+      "theoretical_loss": 4.16005123680448,
+      "tokens_seen": 309329920
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009153426416305569,
+      "loss": 0.0937,
+      "theoretical_loss": 4.159642685007606,
+      "tokens_seen": 309592064
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009152623976889744,
+      "loss": 0.0927,
+      "theoretical_loss": 4.1592345757705225,
+      "tokens_seen": 309854208
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009151821537473922,
+      "loss": 0.0921,
+      "theoretical_loss": 4.158826908240022,
+      "tokens_seen": 310116352
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009151019098058097,
+      "loss": 0.0931,
+      "theoretical_loss": 4.158419681565265,
+      "tokens_seen": 310378496
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009150216658642273,
+      "loss": 0.0949,
+      "theoretical_loss": 4.1580128948977615,
+      "tokens_seen": 310640640
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009149414219226449,
+      "loss": 0.094,
+      "theoretical_loss": 4.157606547391374,
+      "tokens_seen": 310902784
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009148611779810625,
+      "loss": 0.0976,
+      "theoretical_loss": 4.157200638202301,
+      "tokens_seen": 311164928
+    },
+    {
+      "epoch": 0.09,
+      "objective/train/advantage_avg": 0.002907957648858428,
+      "objective/train/docs_used": 120076,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.0078868865966797,
+      "objective/train/original_loss": 2.0078868865966797,
+      "objective/train/theoretical_loss": 4.156997847713709,
+      "objective/train/tokens_used": 331756000,
+      "objective/train/value_avg": -0.010467529296875,
+      "objective/train/value_loss": 0.0005483797285705805,
+      "objective/train/value_max": -0.0007014274597167969,
+      "objective/train/value_min": -0.6416015625,
+      "objective/train/value_reward_corr": 0.563955012185076,
+      "objective/train/value_std": 0.0168914794921875,
+      "objective/train/weight_avg": 1.0031453371047974,
+      "objective/train/weighted_lm_loss": 2.013608694076538,
+      "objective/train/weights_max": 1.2697027921676636,
+      "objective/train/weights_min": 0.37099963426589966,
+      "theoretical_loss": 4.156997847713709,
+      "tokens_seen": 311296000
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.00091478093403948,
+      "loss": 0.0907,
+      "theoretical_loss": 4.156795166489074,
+      "tokens_seen": 311427072
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009147006900978976,
+      "loss": 0.0952,
+      "theoretical_loss": 4.156390131412543,
+      "tokens_seen": 311689216
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009146204461563152,
+      "loss": 0.0937,
+      "theoretical_loss": 4.155985532135875,
+      "tokens_seen": 311951360
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009145402022147327,
+      "loss": 0.097,
+      "theoretical_loss": 4.1555813678245395,
+      "tokens_seen": 312213504
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009144599582731505,
+      "loss": 0.0941,
+      "theoretical_loss": 4.155177637646306,
+      "tokens_seen": 312475648
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.000914379714331568,
+      "loss": 0.0916,
+      "theoretical_loss": 4.154774340771228,
+      "tokens_seen": 312737792
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009142994703899856,
+      "loss": 0.0953,
+      "theoretical_loss": 4.154371476371646,
+      "tokens_seen": 312999936
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 0.0009142192264484032,
+      "loss": 0.0935,
+      "theoretical_loss": 4.153969043622169,
+      "tokens_seen": 313262080
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009141389825068207,
+      "loss": 0.0967,
+      "theoretical_loss": 4.15356704169967,
+      "tokens_seen": 313524224
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009140587385652384,
+      "loss": 0.0973,
+      "theoretical_loss": 4.153165469783279,
+      "tokens_seen": 313786368
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009139784946236559,
+      "loss": 0.0972,
+      "theoretical_loss": 4.152764327054376,
+      "tokens_seen": 314048512
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009138982506820735,
+      "loss": 0.095,
+      "theoretical_loss": 4.152363612696579,
+      "tokens_seen": 314310656
+    },
+    {
+      "epoch": 0.1,
+      "objective/train/advantage_avg": 0.0016880643088370562,
+      "objective/train/docs_used": 121169,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8933781385421753,
+      "objective/train/original_loss": 1.8933782577514648,
+      "objective/train/theoretical_loss": 4.151963325895737,
+      "objective/train/tokens_used": 335032800,
+      "objective/train/value_avg": -0.00879669189453125,
+      "objective/train/value_loss": 0.0004417987947817892,
+      "objective/train/value_max": -0.0004973411560058594,
+      "objective/train/value_min": -0.265380859375,
+      "objective/train/value_reward_corr": 0.4943290748479078,
+      "objective/train/value_std": 0.01016998291015625,
+      "objective/train/weight_avg": 1.0018750429153442,
+      "objective/train/weighted_lm_loss": 1.8967622518539429,
+      "objective/train/weights_max": 1.139946699142456,
+      "objective/train/weights_min": 0.36912843585014343,
+      "theoretical_loss": 4.151963325895737,
+      "tokens_seen": 314572800
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009138180067404911,
+      "loss": 0.0935,
+      "theoretical_loss": 4.151963325895737,
+      "tokens_seen": 314572800
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009137377627989088,
+      "loss": 0.0948,
+      "theoretical_loss": 4.151563465839927,
+      "tokens_seen": 314834944
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009136575188573263,
+      "loss": 0.0952,
+      "theoretical_loss": 4.151164031719437,
+      "tokens_seen": 315097088
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009135772749157439,
+      "loss": 0.0959,
+      "theoretical_loss": 4.15076502272677,
+      "tokens_seen": 315359232
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009134970309741615,
+      "loss": 0.0937,
+      "theoretical_loss": 4.150366438056622,
+      "tokens_seen": 315621376
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000913416787032579,
+      "loss": 0.0926,
+      "theoretical_loss": 4.149968276905888,
+      "tokens_seen": 315883520
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009133365430909967,
+      "loss": 0.099,
+      "theoretical_loss": 4.149570538473644,
+      "tokens_seen": 316145664
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009132562991494142,
+      "loss": 0.0913,
+      "theoretical_loss": 4.149173221961146,
+      "tokens_seen": 316407808
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009131760552078318,
+      "loss": 0.094,
+      "theoretical_loss": 4.1487763265718165,
+      "tokens_seen": 316669952
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009130958112662494,
+      "loss": 0.0944,
+      "theoretical_loss": 4.148379851511241,
+      "tokens_seen": 316932096
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009130155673246669,
+      "loss": 0.0942,
+      "theoretical_loss": 4.147983795987161,
+      "tokens_seen": 317194240
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009129353233830846,
+      "loss": 0.0937,
+      "theoretical_loss": 4.14758815920946,
+      "tokens_seen": 317456384
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009128550794415022,
+      "loss": 0.0916,
+      "theoretical_loss": 4.147192940390165,
+      "tokens_seen": 317718528
+    },
+    {
+      "epoch": 0.1,
+      "objective/train/advantage_avg": 0.00349609088152647,
+      "objective/train/docs_used": 122440,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.709967017173767,
+      "objective/train/original_loss": 1.7099668979644775,
+      "objective/train/theoretical_loss": 4.146995487469262,
+      "objective/train/tokens_used": 338309600,
+      "objective/train/value_avg": -0.00899505615234375,
+      "objective/train/value_loss": 0.00012839387636631727,
+      "objective/train/value_max": -0.0005526542663574219,
+      "objective/train/value_min": -0.255615234375,
+      "objective/train/value_reward_corr": 0.6779735039057588,
+      "objective/train/value_std": 0.0102081298828125,
+      "objective/train/weight_avg": 1.0035591125488281,
+      "objective/train/weighted_lm_loss": 1.7162126302719116,
+      "objective/train/weights_max": 1.1629977226257324,
+      "objective/train/weights_min": 0.6108145713806152,
+      "theoretical_loss": 4.146995487469262,
+      "tokens_seen": 317849600
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009127748354999198,
+      "loss": 0.0971,
+      "theoretical_loss": 4.146798138743433,
+      "tokens_seen": 317980672
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009126945915583374,
+      "loss": 0.0904,
+      "theoretical_loss": 4.146403753485544,
+      "tokens_seen": 318242816
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000912614347616755,
+      "loss": 0.093,
+      "theoretical_loss": 4.146009783834892,
+      "tokens_seen": 318504960
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009125341036751725,
+      "loss": 0.0953,
+      "theoretical_loss": 4.145616229011987,
+      "tokens_seen": 318767104
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009124538597335902,
+      "loss": 0.0957,
+      "theoretical_loss": 4.145223088239432,
+      "tokens_seen": 319029248
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009123736157920077,
+      "loss": 0.0923,
+      "theoretical_loss": 4.14483036074193,
+      "tokens_seen": 319291392
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009122933718504252,
+      "loss": 0.0962,
+      "theoretical_loss": 4.14443804574627,
+      "tokens_seen": 319553536
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000912213127908843,
+      "loss": 0.0927,
+      "theoretical_loss": 4.144046142481317,
+      "tokens_seen": 319815680
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009121328839672605,
+      "loss": 0.0968,
+      "theoretical_loss": 4.143654650178012,
+      "tokens_seen": 320077824
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009120526400256781,
+      "loss": 0.097,
+      "theoretical_loss": 4.143263568069358,
+      "tokens_seen": 320339968
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009119723960840957,
+      "loss": 0.0928,
+      "theoretical_loss": 4.142872895390417,
+      "tokens_seen": 320602112
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009118921521425133,
+      "loss": 0.0939,
+      "theoretical_loss": 4.142482631378303,
+      "tokens_seen": 320864256
+    },
+    {
+      "epoch": 0.1,
+      "objective/train/advantage_avg": 0.0006320319953374565,
+      "objective/train/docs_used": 123582,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8734241724014282,
+      "objective/train/original_loss": 1.8734240531921387,
+      "objective/train/theoretical_loss": 4.142092775272169,
+      "objective/train/tokens_used": 341586400,
+      "objective/train/value_avg": -0.01175689697265625,
+      "objective/train/value_loss": 0.00034092742134816945,
+      "objective/train/value_max": -0.0005526542663574219,
+      "objective/train/value_min": -0.2489013671875,
+      "objective/train/value_reward_corr": 0.649642412628804,
+      "objective/train/value_std": 0.01458740234375,
+      "objective/train/weight_avg": 1.0007905960083008,
+      "objective/train/weighted_lm_loss": 1.8751580715179443,
+      "objective/train/weights_max": 1.2165815830230713,
+      "objective/train/weights_min": 0.37836602330207825,
+      "theoretical_loss": 4.142092775272169,
+      "tokens_seen": 321126400
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009118119082009308,
+      "loss": 0.0918,
+      "theoretical_loss": 4.142092775272169,
+      "tokens_seen": 321126400
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009117316642593484,
+      "loss": 0.0898,
+      "theoretical_loss": 4.141703326313209,
+      "tokens_seen": 321388544
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000911651420317766,
+      "loss": 0.0933,
+      "theoretical_loss": 4.141314283744643,
+      "tokens_seen": 321650688
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009115711763761835,
+      "loss": 0.093,
+      "theoretical_loss": 4.140925646811714,
+      "tokens_seen": 321912832
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009114909324346013,
+      "loss": 0.0948,
+      "theoretical_loss": 4.1405374147616785,
+      "tokens_seen": 322174976
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009114106884930188,
+      "loss": 0.0929,
+      "theoretical_loss": 4.140149586843803,
+      "tokens_seen": 322437120
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009113304445514365,
+      "loss": 0.094,
+      "theoretical_loss": 4.139762162309355,
+      "tokens_seen": 322699264
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000911250200609854,
+      "loss": 0.0934,
+      "theoretical_loss": 4.139375140411592,
+      "tokens_seen": 322961408
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009111699566682715,
+      "loss": 0.0913,
+      "theoretical_loss": 4.138988520405764,
+      "tokens_seen": 323223552
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009110897127266892,
+      "loss": 0.0922,
+      "theoretical_loss": 4.138602301549097,
+      "tokens_seen": 323485696
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009110094687851067,
+      "loss": 0.0922,
+      "theoretical_loss": 4.138216483100795,
+      "tokens_seen": 323747840
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009109292248435243,
+      "loss": 0.0901,
+      "theoretical_loss": 4.137831064322021,
+      "tokens_seen": 324009984
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009108489809019419,
+      "loss": 0.0906,
+      "theoretical_loss": 4.1374460444759045,
+      "tokens_seen": 324272128
+    },
+    {
+      "epoch": 0.1,
+      "objective/train/advantage_avg": 0.0036898814141750336,
+      "objective/train/docs_used": 124782,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.784644365310669,
+      "objective/train/original_loss": 1.784644603729248,
+      "objective/train/theoretical_loss": 4.137253683922854,
+      "objective/train/tokens_used": 344863200,
+      "objective/train/value_avg": -0.00811767578125,
+      "objective/train/value_loss": 0.0003189310082234442,
+      "objective/train/value_max": -0.0005593299865722656,
+      "objective/train/value_min": -0.67138671875,
+      "objective/train/value_reward_corr": 0.6118297905242213,
+      "objective/train/value_std": 0.0111846923828125,
+      "objective/train/weight_avg": 1.0038195848464966,
+      "objective/train/weighted_lm_loss": 1.790905475616455,
+      "objective/train/weights_max": 1.1770083904266357,
+      "objective/train/weights_min": 0.23969075083732605,
+      "theoretical_loss": 4.137253683922854,
+      "tokens_seen": 324403200
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009107687369603596,
+      "loss": 0.0923,
+      "theoretical_loss": 4.137061422827525,
+      "tokens_seen": 324534272
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009106884930187771,
+      "loss": 0.0939,
+      "theoretical_loss": 4.136677198643908,
+      "tokens_seen": 324796416
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009106082490771947,
+      "loss": 0.0912,
+      "theoretical_loss": 4.13629337119402,
+      "tokens_seen": 325058560
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009105280051356123,
+      "loss": 0.0923,
+      "theoretical_loss": 4.135909939748757,
+      "tokens_seen": 325320704
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009104477611940298,
+      "loss": 0.0918,
+      "theoretical_loss": 4.135526903580946,
+      "tokens_seen": 325582848
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009103675172524475,
+      "loss": 0.0938,
+      "theoretical_loss": 4.135144261965327,
+      "tokens_seen": 325844992
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000910287273310865,
+      "loss": 0.0923,
+      "theoretical_loss": 4.134762014178559,
+      "tokens_seen": 326107136
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009102070293692827,
+      "loss": 0.0925,
+      "theoretical_loss": 4.134380159499204,
+      "tokens_seen": 326369280
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009101267854277002,
+      "loss": 0.0931,
+      "theoretical_loss": 4.1339986972077245,
+      "tokens_seen": 326631424
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009100465414861177,
+      "loss": 0.0885,
+      "theoretical_loss": 4.133617626586475,
+      "tokens_seen": 326893568
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009099662975445355,
+      "loss": 0.0916,
+      "theoretical_loss": 4.133236946919698,
+      "tokens_seen": 327155712
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000909886053602953,
+      "loss": 0.0972,
+      "theoretical_loss": 4.132856657493516,
+      "tokens_seen": 327417856
+    },
+    {
+      "debugging/Compilability": 1.0,
+      "debugging/distinct-1-grams": 0.7246169962973871,
+      "debugging/entropy-1-grams": 5.13190351820591,
+      "debugging/length": 472.8,
+      "debugging/num_segments": 10,
+      "debugging/raw_token_scores_avg": 0.008663635700941086,
+      "debugging/raw_token_scores_std": 0.023182114586234093,
+      "debugging/score": 0.0033104506818385365,
+      "debugging/score_std": 0.0017105448100298692,
+      "epoch": 0.1,
+      "objective/train/advantage_avg": 0.0008121287100948393,
+      "objective/train/docs_used": 126013,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9080229997634888,
+      "objective/train/original_loss": 1.9080228805541992,
+      "objective/train/theoretical_loss": 4.132476757595925,
+      "objective/train/tokens_used": 348140000,
+      "objective/train/value_avg": -0.0094757080078125,
+      "objective/train/value_loss": 0.0002959860139526427,
+      "objective/train/value_max": -0.0003554821014404297,
+      "objective/train/value_min": -0.2890625,
+      "objective/train/value_reward_corr": 0.6876529624313351,
+      "objective/train/value_std": 0.01248931884765625,
+      "objective/train/weight_avg": 1.000952959060669,
+      "objective/train/weighted_lm_loss": 1.9104235172271729,
+      "objective/train/weights_max": 1.1095370054244995,
+      "objective/train/weights_min": 0.376959890127182,
+      "theoretical_loss": 4.132476757595925,
+      "tokens_seen": 327680000
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009098058096613706,
+      "loss": 0.0954,
+      "theoretical_loss": 4.132476757595925,
+      "tokens_seen": 327680000
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009097255657197882,
+      "loss": 0.0948,
+      "theoretical_loss": 4.132097246516788,
+      "tokens_seen": 327942144
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009096453217782058,
+      "loss": 0.0945,
+      "theoretical_loss": 4.131718123547829,
+      "tokens_seen": 328204288
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009095650778366233,
+      "loss": 0.0931,
+      "theoretical_loss": 4.131339387982628,
+      "tokens_seen": 328466432
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000909484833895041,
+      "loss": 0.0913,
+      "theoretical_loss": 4.1309610391166105,
+      "tokens_seen": 328728576
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009094045899534585,
+      "loss": 0.0902,
+      "theoretical_loss": 4.1305830762470475,
+      "tokens_seen": 328990720
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000909324346011876,
+      "loss": 0.093,
+      "theoretical_loss": 4.1302054986730425,
+      "tokens_seen": 329252864
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009092441020702938,
+      "loss": 0.0922,
+      "theoretical_loss": 4.129828305695531,
+      "tokens_seen": 329515008
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009091638581287113,
+      "loss": 0.0919,
+      "theoretical_loss": 4.129451496617269,
+      "tokens_seen": 329777152
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009090836141871289,
+      "loss": 0.0938,
+      "theoretical_loss": 4.129075070742831,
+      "tokens_seen": 330039296
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009090033702455465,
+      "loss": 0.093,
+      "theoretical_loss": 4.128699027378604,
+      "tokens_seen": 330301440
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009089231263039641,
+      "loss": 0.0926,
+      "theoretical_loss": 4.128323365832777,
+      "tokens_seen": 330563584
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009088428823623817,
+      "loss": 0.0938,
+      "theoretical_loss": 4.127948085415338,
+      "tokens_seen": 330825728
+    },
+    {
+      "epoch": 0.1,
+      "objective/train/advantage_avg": 0.003476484213024378,
+      "objective/train/docs_used": 127130,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9096784591674805,
+      "objective/train/original_loss": 1.90967857837677,
+      "objective/train/theoretical_loss": 4.127760587914626,
+      "objective/train/tokens_used": 351416800,
+      "objective/train/value_avg": -0.009552001953125,
+      "objective/train/value_loss": 0.00040583781083114445,
+      "objective/train/value_max": -0.0003006458282470703,
+      "objective/train/value_min": -0.32568359375,
+      "objective/train/value_reward_corr": 0.46149172405686373,
+      "objective/train/value_std": 0.01270294189453125,
+      "objective/train/weight_avg": 1.0036547183990479,
+      "objective/train/weighted_lm_loss": 1.9170717000961304,
+      "objective/train/weights_max": 1.1853569746017456,
+      "objective/train/weights_min": 0.36846548318862915,
+      "theoretical_loss": 4.127760587914626,
+      "tokens_seen": 330956800
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009087626384207992,
+      "loss": 0.0939,
+      "theoretical_loss": 4.127573185438068,
+      "tokens_seen": 331087872
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009086823944792168,
+      "loss": 0.0963,
+      "theoretical_loss": 4.127198665214536,
+      "tokens_seen": 331350016
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009086021505376344,
+      "loss": 0.0901,
+      "theoretical_loss": 4.126824524060088,
+      "tokens_seen": 331612160
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000908521906596052,
+      "loss": 0.093,
+      "theoretical_loss": 4.126450761291847,
+      "tokens_seen": 331874304
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009084416626544696,
+      "loss": 0.0952,
+      "theoretical_loss": 4.126077376228702,
+      "tokens_seen": 332136448
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009083614187128873,
+      "loss": 0.0878,
+      "theoretical_loss": 4.1257043681913075,
+      "tokens_seen": 332398592
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009082811747713048,
+      "loss": 0.092,
+      "theoretical_loss": 4.125331736502073,
+      "tokens_seen": 332660736
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009082009308297223,
+      "loss": 0.0937,
+      "theoretical_loss": 4.124959480485156,
+      "tokens_seen": 332922880
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.00090812068688814,
+      "loss": 0.0899,
+      "theoretical_loss": 4.124587599466462,
+      "tokens_seen": 333185024
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009080404429465575,
+      "loss": 0.0936,
+      "theoretical_loss": 4.124216092773635,
+      "tokens_seen": 333447168
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009079601990049751,
+      "loss": 0.0898,
+      "theoretical_loss": 4.123844959736049,
+      "tokens_seen": 333709312
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009078799550633927,
+      "loss": 0.091,
+      "theoretical_loss": 4.123474199684807,
+      "tokens_seen": 333971456
+    },
+    {
+      "epoch": 0.1,
+      "objective/train/advantage_avg": 0.0011445061536505818,
+      "objective/train/docs_used": 128373,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 2.0021450519561768,
+      "objective/train/original_loss": 2.0021448135375977,
+      "objective/train/theoretical_loss": 4.123103811952736,
+      "objective/train/tokens_used": 354693600,
+      "objective/train/value_avg": -0.010498046875,
+      "objective/train/value_loss": 0.0002171178930439055,
+      "objective/train/value_max": -0.0004494190216064453,
+      "objective/train/value_min": -0.269287109375,
+      "objective/train/value_reward_corr": 0.6008675056499843,
+      "objective/train/value_std": 0.011932373046875,
+      "objective/train/weight_avg": 1.0012482404708862,
+      "objective/train/weighted_lm_loss": 2.003540515899658,
+      "objective/train/weights_max": 1.1421287059783936,
+      "objective/train/weights_min": 0.3697309195995331,
+      "theoretical_loss": 4.123103811952736,
+      "tokens_seen": 334233600
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009077997111218103,
+      "loss": 0.0955,
+      "theoretical_loss": 4.123103811952736,
+      "tokens_seen": 334233600
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000907719467180228,
+      "loss": 0.0893,
+      "theoretical_loss": 4.122733795874372,
+      "tokens_seen": 334495744
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009076392232386455,
+      "loss": 0.0945,
+      "theoretical_loss": 4.122364150785966,
+      "tokens_seen": 334757888
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009075589792970631,
+      "loss": 0.0945,
+      "theoretical_loss": 4.1219948760254725,
+      "tokens_seen": 335020032
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009074787353554807,
+      "loss": 0.0932,
+      "theoretical_loss": 4.121625970932542,
+      "tokens_seen": 335282176
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009073984914138983,
+      "loss": 0.093,
+      "theoretical_loss": 4.121257434848519,
+      "tokens_seen": 335544320
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009073182474723158,
+      "loss": 0.0925,
+      "theoretical_loss": 4.120889267116435,
+      "tokens_seen": 335806464
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009072380035307335,
+      "loss": 0.0924,
+      "theoretical_loss": 4.1205214670810015,
+      "tokens_seen": 336068608
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000907157759589151,
+      "loss": 0.0926,
+      "theoretical_loss": 4.120154034088609,
+      "tokens_seen": 336330752
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009070775156475685,
+      "loss": 0.0924,
+      "theoretical_loss": 4.119786967487314,
+      "tokens_seen": 336592896
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009069972717059863,
+      "loss": 0.0936,
+      "theoretical_loss": 4.11942026662684,
+      "tokens_seen": 336855040
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009069170277644038,
+      "loss": 0.0898,
+      "theoretical_loss": 4.11905393085857,
+      "tokens_seen": 337117184
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009068367838228214,
+      "loss": 0.093,
+      "theoretical_loss": 4.118687959535539,
+      "tokens_seen": 337379328
+    },
+    {
+      "epoch": 0.1,
+      "objective/train/advantage_avg": 0.0014116659294813871,
+      "objective/train/docs_used": 129675,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7438528537750244,
+      "objective/train/original_loss": 1.7438528537750244,
+      "objective/train/theoretical_loss": 4.118505110339262,
+      "objective/train/tokens_used": 357970400,
+      "objective/train/value_avg": -0.0096435546875,
+      "objective/train/value_loss": 0.00027853285428136587,
+      "objective/train/value_max": -0.00034737586975097656,
+      "objective/train/value_min": -0.552734375,
+      "objective/train/value_reward_corr": 0.6218103415544912,
+      "objective/train/value_std": 0.0139312744140625,
+      "objective/train/weight_avg": 1.001541256904602,
+      "objective/train/weighted_lm_loss": 1.745504379272461,
+      "objective/train/weights_max": 1.441696286201477,
+      "objective/train/weights_min": 0.36845633387565613,
+      "theoretical_loss": 4.118505110339262,
+      "tokens_seen": 337510400
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000906756539881239,
+      "loss": 0.091,
+      "theoretical_loss": 4.118322352012429,
+      "tokens_seen": 337641472
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009066762959396566,
+      "loss": 0.0905,
+      "theoretical_loss": 4.117957107645569,
+      "tokens_seen": 337903616
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009065960519980741,
+      "loss": 0.0931,
+      "theoretical_loss": 4.1175922257929205,
+      "tokens_seen": 338165760
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009065158080564918,
+      "loss": 0.0947,
+      "theoretical_loss": 4.117227705814078,
+      "tokens_seen": 338427904
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009064355641149093,
+      "loss": 0.0977,
+      "theoretical_loss": 4.116863547070264,
+      "tokens_seen": 338690048
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009063553201733269,
+      "loss": 0.0932,
+      "theoretical_loss": 4.116499748924319,
+      "tokens_seen": 338952192
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009062750762317446,
+      "loss": 0.0935,
+      "theoretical_loss": 4.116136310740702,
+      "tokens_seen": 339214336
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009061948322901621,
+      "loss": 0.0875,
+      "theoretical_loss": 4.115773231885479,
+      "tokens_seen": 339476480
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009061145883485798,
+      "loss": 0.0908,
+      "theoretical_loss": 4.115410511726323,
+      "tokens_seen": 339738624
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009060343444069973,
+      "loss": 0.0922,
+      "theoretical_loss": 4.115048149632507,
+      "tokens_seen": 340000768
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009059541004654149,
+      "loss": 0.0936,
+      "theoretical_loss": 4.114686144974897,
+      "tokens_seen": 340262912
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009058738565238325,
+      "loss": 0.0925,
+      "theoretical_loss": 4.114324497125947,
+      "tokens_seen": 340525056
+    },
+    {
+      "epoch": 0.1,
+      "objective/train/advantage_avg": 0.0013680076226592064,
+      "objective/train/docs_used": 130755,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8868423700332642,
+      "objective/train/original_loss": 1.8868422508239746,
+      "objective/train/theoretical_loss": 4.113963205459697,
+      "objective/train/tokens_used": 361247200,
+      "objective/train/value_avg": -0.007740020751953125,
+      "objective/train/value_loss": 0.00031318547553382814,
+      "objective/train/value_max": -0.0003654956817626953,
+      "objective/train/value_min": -0.4619140625,
+      "objective/train/value_reward_corr": 0.5194909400023118,
+      "objective/train/value_std": 0.0107574462890625,
+      "objective/train/weight_avg": 1.0015056133270264,
+      "objective/train/weighted_lm_loss": 1.8895421028137207,
+      "objective/train/weights_max": 1.373528242111206,
+      "objective/train/weights_min": 0.38126373291015625,
+      "theoretical_loss": 4.113963205459697,
+      "tokens_seen": 340787200
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.00090579361258225,
+      "loss": 0.0913,
+      "theoretical_loss": 4.113963205459697,
+      "tokens_seen": 340787200
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009057133686406676,
+      "loss": 0.0915,
+      "theoretical_loss": 4.113602269351765,
+      "tokens_seen": 341049344
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009056331246990852,
+      "loss": 0.0906,
+      "theoretical_loss": 4.113241688179341,
+      "tokens_seen": 341311488
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009055528807575029,
+      "loss": 0.0934,
+      "theoretical_loss": 4.1128814613211855,
+      "tokens_seen": 341573632
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009054726368159204,
+      "loss": 0.0902,
+      "theoretical_loss": 4.1125215881576205,
+      "tokens_seen": 341835776
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009053923928743381,
+      "loss": 0.0901,
+      "theoretical_loss": 4.112162068070525,
+      "tokens_seen": 342097920
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009053121489327556,
+      "loss": 0.0916,
+      "theoretical_loss": 4.111802900443333,
+      "tokens_seen": 342360064
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009052319049911731,
+      "loss": 0.0895,
+      "theoretical_loss": 4.111444084661026,
+      "tokens_seen": 342622208
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009051516610495908,
+      "loss": 0.0891,
+      "theoretical_loss": 4.111085620110127,
+      "tokens_seen": 342884352
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009050714171080083,
+      "loss": 0.0918,
+      "theoretical_loss": 4.110727506178697,
+      "tokens_seen": 343146496
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.000904991173166426,
+      "loss": 0.0935,
+      "theoretical_loss": 4.110369742256329,
+      "tokens_seen": 343408640
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009049109292248435,
+      "loss": 0.0902,
+      "theoretical_loss": 4.110012327734145,
+      "tokens_seen": 343670784
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009048306852832611,
+      "loss": 0.0916,
+      "theoretical_loss": 4.1096552620047895,
+      "tokens_seen": 343932928
+    },
+    {
+      "epoch": 0.1,
+      "objective/train/advantage_avg": 0.00047141360118985176,
+      "objective/train/docs_used": 131936,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.86632239818573,
+      "objective/train/original_loss": 1.8663225173950195,
+      "objective/train/theoretical_loss": 4.109476859748038,
+      "objective/train/tokens_used": 364524000,
+      "objective/train/value_avg": -0.00713348388671875,
+      "objective/train/value_loss": 0.00014549301704391837,
+      "objective/train/value_max": -0.0002472400665283203,
+      "objective/train/value_min": -0.466064453125,
+      "objective/train/value_reward_corr": 0.7043933872865291,
+      "objective/train/value_std": 0.01116180419921875,
+      "objective/train/weight_avg": 1.0005429983139038,
+      "objective/train/weighted_lm_loss": 1.8668056726455688,
+      "objective/train/weights_max": 1.1419428586959839,
+      "objective/train/weights_min": 0.7390621304512024,
+      "theoretical_loss": 4.109476859748038,
+      "tokens_seen": 344064000
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009047504413416788,
+      "loss": 0.0924,
+      "theoretical_loss": 4.109298544462423,
+      "tokens_seen": 344195072
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009046701974000963,
+      "loss": 0.087,
+      "theoretical_loss": 4.108942174502721,
+      "tokens_seen": 344457216
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009045899534585139,
+      "loss": 0.0912,
+      "theoretical_loss": 4.108586151522863,
+      "tokens_seen": 344719360
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009045097095169315,
+      "loss": 0.0936,
+      "theoretical_loss": 4.1082304749215375,
+      "tokens_seen": 344981504
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009044294655753491,
+      "loss": 0.0912,
+      "theoretical_loss": 4.107875144098925,
+      "tokens_seen": 345243648
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009043492216337666,
+      "loss": 0.0954,
+      "theoretical_loss": 4.107520158456703,
+      "tokens_seen": 345505792
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009042689776921843,
+      "loss": 0.0906,
+      "theoretical_loss": 4.107165517398034,
+      "tokens_seen": 345767936
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009041887337506018,
+      "loss": 0.093,
+      "theoretical_loss": 4.106811220327568,
+      "tokens_seen": 346030080
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 0.0009041084898090193,
+      "loss": 0.0911,
+      "theoretical_loss": 4.10645726665143,
+      "tokens_seen": 346292224
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009040282458674371,
+      "loss": 0.0936,
+      "theoretical_loss": 4.10610365577722,
+      "tokens_seen": 346554368
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009039480019258546,
+      "loss": 0.0919,
+      "theoretical_loss": 4.105750387114009,
+      "tokens_seen": 346816512
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009038677579842723,
+      "loss": 0.0909,
+      "theoretical_loss": 4.105397460072329,
+      "tokens_seen": 347078656
+    },
+    {
+      "epoch": 0.11,
+      "objective/train/advantage_avg": 0.0021935408003628254,
+      "objective/train/docs_used": 133174,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.799157738685608,
+      "objective/train/original_loss": 1.7991578578948975,
+      "objective/train/theoretical_loss": 4.105044874064174,
+      "objective/train/tokens_used": 367800800,
+      "objective/train/value_avg": -0.00794219970703125,
+      "objective/train/value_loss": 0.00020322235650382936,
+      "objective/train/value_max": -0.0003459453582763672,
+      "objective/train/value_min": -0.2293701171875,
+      "objective/train/value_reward_corr": 0.5725621260668602,
+      "objective/train/value_std": 0.0089569091796875,
+      "objective/train/weight_avg": 1.0022897720336914,
+      "objective/train/weighted_lm_loss": 1.801876425743103,
+      "objective/train/weights_max": 1.1363343000411987,
+      "objective/train/weights_min": 0.3731314241886139,
+      "theoretical_loss": 4.105044874064174,
+      "tokens_seen": 347340800
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009037875140426898,
+      "loss": 0.0874,
+      "theoretical_loss": 4.105044874064174,
+      "tokens_seen": 347340800
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009037072701011074,
+      "loss": 0.0912,
+      "theoretical_loss": 4.104692628502993,
+      "tokens_seen": 347602944
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.000903627026159525,
+      "loss": 0.0899,
+      "theoretical_loss": 4.104340722803683,
+      "tokens_seen": 347865088
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009035467822179425,
+      "loss": 0.0907,
+      "theoretical_loss": 4.103989156382589,
+      "tokens_seen": 348127232
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009034665382763601,
+      "loss": 0.0876,
+      "theoretical_loss": 4.103637928657495,
+      "tokens_seen": 348389376
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009033862943347777,
+      "loss": 0.0955,
+      "theoretical_loss": 4.103287039047622,
+      "tokens_seen": 348651520
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009033060503931954,
+      "loss": 0.0904,
+      "theoretical_loss": 4.102936486973624,
+      "tokens_seen": 348913664
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009032258064516129,
+      "loss": 0.09,
+      "theoretical_loss": 4.102586271857579,
+      "tokens_seen": 349175808
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009031455625100306,
+      "loss": 0.0886,
+      "theoretical_loss": 4.102236393122989,
+      "tokens_seen": 349437952
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009030653185684481,
+      "loss": 0.092,
+      "theoretical_loss": 4.101886850194775,
+      "tokens_seen": 349700096
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009029850746268657,
+      "loss": 0.0886,
+      "theoretical_loss": 4.10153764249927,
+      "tokens_seen": 349962240
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009029048306852833,
+      "loss": 0.0889,
+      "theoretical_loss": 4.1011887694642155,
+      "tokens_seen": 350224384
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009028245867437008,
+      "loss": 0.09,
+      "theoretical_loss": 4.100840230518759,
+      "tokens_seen": 350486528
+    },
+    {
+      "epoch": 0.11,
+      "objective/train/advantage_avg": -0.0007634600042365491,
+      "objective/train/docs_used": 134382,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8593584299087524,
+      "objective/train/original_loss": 1.859358549118042,
+      "objective/train/theoretical_loss": 4.10066608615162,
+      "objective/train/tokens_used": 371077600,
+      "objective/train/value_avg": -0.00992584228515625,
+      "objective/train/value_loss": 0.00040058817830868065,
+      "objective/train/value_max": -0.000415802001953125,
+      "objective/train/value_min": -0.1810302734375,
+      "objective/train/value_reward_corr": 0.653844216492223,
+      "objective/train/value_std": 0.01154327392578125,
+      "objective/train/weight_avg": 0.9994103908538818,
+      "objective/train/weighted_lm_loss": 1.8588926792144775,
+      "objective/train/weights_max": 1.1080909967422485,
+      "objective/train/weights_min": 0.22362525761127472,
+      "theoretical_loss": 4.10066608615162,
+      "tokens_seen": 350617600
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009027443428021184,
+      "loss": 0.0923,
+      "theoretical_loss": 4.100492025093445,
+      "tokens_seen": 350748672
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.000902664098860536,
+      "loss": 0.0924,
+      "theoretical_loss": 4.100144152620215,
+      "tokens_seen": 351010816
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009025838549189537,
+      "loss": 0.0908,
+      "theoretical_loss": 4.099796612532403,
+      "tokens_seen": 351272960
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009025036109773713,
+      "loss": 0.09,
+      "theoretical_loss": 4.0994494042647265,
+      "tokens_seen": 351535104
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009024233670357889,
+      "loss": 0.0908,
+      "theoretical_loss": 4.099102527253285,
+      "tokens_seen": 351797248
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009023431230942064,
+      "loss": 0.0926,
+      "theoretical_loss": 4.098755980935557,
+      "tokens_seen": 352059392
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.000902262879152624,
+      "loss": 0.0895,
+      "theoretical_loss": 4.0984097647503965,
+      "tokens_seen": 352321536
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009021826352110416,
+      "loss": 0.0906,
+      "theoretical_loss": 4.09806387813802,
+      "tokens_seen": 352583680
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009021023912694591,
+      "loss": 0.0878,
+      "theoretical_loss": 4.0977183205400145,
+      "tokens_seen": 352845824
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009020221473278768,
+      "loss": 0.0901,
+      "theoretical_loss": 4.097373091399324,
+      "tokens_seen": 353107968
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009019419033862943,
+      "loss": 0.0898,
+      "theoretical_loss": 4.097028190160249,
+      "tokens_seen": 353370112
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.000901861659444712,
+      "loss": 0.0912,
+      "theoretical_loss": 4.096683616268442,
+      "tokens_seen": 353632256
+    },
+    {
+      "epoch": 0.11,
+      "objective/train/advantage_avg": 0.0033219335600733757,
+      "objective/train/docs_used": 135614,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5801607370376587,
+      "objective/train/original_loss": 1.5801606178283691,
+      "objective/train/theoretical_loss": 4.096339369170902,
+      "objective/train/tokens_used": 374354400,
+      "objective/train/value_avg": -0.01206207275390625,
+      "objective/train/value_loss": 0.00040383703890256584,
+      "objective/train/value_max": -0.0004992485046386719,
+      "objective/train/value_min": -0.5283203125,
+      "objective/train/value_reward_corr": 0.6677181799151066,
+      "objective/train/value_std": 0.0159149169921875,
+      "objective/train/weight_avg": 1.003501296043396,
+      "objective/train/weighted_lm_loss": 1.5844190120697021,
+      "objective/train/weights_max": 1.4516695737838745,
+      "objective/train/weights_min": 0.37713250517845154,
+      "theoretical_loss": 4.096339369170902,
+      "tokens_seen": 353894400
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009017814155031296,
+      "loss": 0.0904,
+      "theoretical_loss": 4.096339369170902,
+      "tokens_seen": 353894400
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009017011715615471,
+      "loss": 0.0911,
+      "theoretical_loss": 4.095995448315972,
+      "tokens_seen": 354156544
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009016209276199647,
+      "loss": 0.0903,
+      "theoretical_loss": 4.095651853153331,
+      "tokens_seen": 354418688
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009015406836783823,
+      "loss": 0.0916,
+      "theoretical_loss": 4.095308583133997,
+      "tokens_seen": 354680832
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009014604397367999,
+      "loss": 0.0917,
+      "theoretical_loss": 4.094965637710314,
+      "tokens_seen": 354942976
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009013801957952175,
+      "loss": 0.0907,
+      "theoretical_loss": 4.094623016335954,
+      "tokens_seen": 355205120
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009012999518536351,
+      "loss": 0.0904,
+      "theoretical_loss": 4.094280718465911,
+      "tokens_seen": 355467264
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009012197079120526,
+      "loss": 0.0904,
+      "theoretical_loss": 4.093938743556496,
+      "tokens_seen": 355729408
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009011394639704702,
+      "loss": 0.0891,
+      "theoretical_loss": 4.093597091065333,
+      "tokens_seen": 355991552
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009010592200288879,
+      "loss": 0.0892,
+      "theoretical_loss": 4.093255760451357,
+      "tokens_seen": 356253696
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009009789760873054,
+      "loss": 0.0887,
+      "theoretical_loss": 4.092914751174808,
+      "tokens_seen": 356515840
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009008987321457231,
+      "loss": 0.0903,
+      "theoretical_loss": 4.092574062697225,
+      "tokens_seen": 356777984
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009008184882041406,
+      "loss": 0.0885,
+      "theoretical_loss": 4.092233694481447,
+      "tokens_seen": 357040128
+    },
+    {
+      "epoch": 0.11,
+      "objective/train/advantage_avg": 0.0004918540944345295,
+      "objective/train/docs_used": 136331,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9003441333770752,
+      "objective/train/original_loss": 1.9003441333770752,
+      "objective/train/theoretical_loss": 4.092063630304224,
+      "objective/train/tokens_used": 377631200,
+      "objective/train/value_avg": -0.007537841796875,
+      "objective/train/value_loss": 0.00020429695723578334,
+      "objective/train/value_max": -0.0003597736358642578,
+      "objective/train/value_min": -0.2430419921875,
+      "objective/train/value_reward_corr": 0.5656587517444805,
+      "objective/train/value_std": 0.00870513916015625,
+      "objective/train/weight_avg": 1.0005881786346436,
+      "objective/train/weighted_lm_loss": 1.9004318714141846,
+      "objective/train/weights_max": 1.1491543054580688,
+      "objective/train/weights_min": 0.3702276945114136,
+      "theoretical_loss": 4.092063630304224,
+      "tokens_seen": 357171200
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009007382442625582,
+      "loss": 0.0933,
+      "theoretical_loss": 4.091893645991604,
+      "tokens_seen": 357302272
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009006580003209758,
+      "loss": 0.0918,
+      "theoretical_loss": 4.091553916693115,
+      "tokens_seen": 357564416
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009005777563793933,
+      "loss": 0.0863,
+      "theoretical_loss": 4.091214506052687,
+      "tokens_seen": 357826560
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009004975124378109,
+      "loss": 0.0846,
+      "theoretical_loss": 4.090875413538302,
+      "tokens_seen": 358088704
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009004172684962285,
+      "loss": 0.0925,
+      "theoretical_loss": 4.090536638619224,
+      "tokens_seen": 358350848
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009003370245546462,
+      "loss": 0.0879,
+      "theoretical_loss": 4.090198180765989,
+      "tokens_seen": 358612992
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009002567806130637,
+      "loss": 0.0895,
+      "theoretical_loss": 4.0898600394504,
+      "tokens_seen": 358875136
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009001765366714814,
+      "loss": 0.0891,
+      "theoretical_loss": 4.089522214145525,
+      "tokens_seen": 359137280
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009000962927298989,
+      "loss": 0.0917,
+      "theoretical_loss": 4.089184704325695,
+      "tokens_seen": 359399424
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0009000160487883166,
+      "loss": 0.0891,
+      "theoretical_loss": 4.088847509466497,
+      "tokens_seen": 359661568
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008999358048467341,
+      "loss": 0.0889,
+      "theoretical_loss": 4.088510629044771,
+      "tokens_seen": 359923712
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008998555609051516,
+      "loss": 0.0889,
+      "theoretical_loss": 4.088174062538605,
+      "tokens_seen": 360185856
+    },
+    {
+      "epoch": 0.11,
+      "objective/train/advantage_avg": 0.0018875771202147007,
+      "objective/train/docs_used": 137426,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.9378498792648315,
+      "objective/train/original_loss": 1.937849998474121,
+      "objective/train/theoretical_loss": 4.087837809427334,
+      "objective/train/tokens_used": 380908000,
+      "objective/train/value_avg": -0.01007843017578125,
+      "objective/train/value_loss": 0.00019329431233927608,
+      "objective/train/value_max": -0.0002779960632324219,
+      "objective/train/value_min": -0.277099609375,
+      "objective/train/value_reward_corr": 0.6944034668612225,
+      "objective/train/value_std": 0.01515960693359375,
+      "objective/train/weight_avg": 1.0019793510437012,
+      "objective/train/weighted_lm_loss": 1.9406569004058838,
+      "objective/train/weights_max": 1.115104079246521,
+      "objective/train/weights_min": 0.36884480714797974,
+      "theoretical_loss": 4.087837809427334,
+      "tokens_seen": 360448000
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008997753169635693,
+      "loss": 0.0936,
+      "theoretical_loss": 4.087837809427334,
+      "tokens_seen": 360448000
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008996950730219868,
+      "loss": 0.0886,
+      "theoretical_loss": 4.087501869191536,
+      "tokens_seen": 360710144
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008996148290804045,
+      "loss": 0.0882,
+      "theoretical_loss": 4.087166241313023,
+      "tokens_seen": 360972288
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008995345851388221,
+      "loss": 0.093,
+      "theoretical_loss": 4.086830925274842,
+      "tokens_seen": 361234432
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008994543411972397,
+      "loss": 0.0892,
+      "theoretical_loss": 4.08649592056127,
+      "tokens_seen": 361496576
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008993740972556572,
+      "loss": 0.093,
+      "theoretical_loss": 4.086161226657811,
+      "tokens_seen": 361758720
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008992938533140748,
+      "loss": 0.0863,
+      "theoretical_loss": 4.085826843051191,
+      "tokens_seen": 362020864
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008992136093724924,
+      "loss": 0.092,
+      "theoretical_loss": 4.0854927692293534,
+      "tokens_seen": 362283008
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008991333654309099,
+      "loss": 0.0896,
+      "theoretical_loss": 4.085159004681458,
+      "tokens_seen": 362545152
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008990531214893276,
+      "loss": 0.0908,
+      "theoretical_loss": 4.084825548897873,
+      "tokens_seen": 362807296
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008989728775477451,
+      "loss": 0.0933,
+      "theoretical_loss": 4.084492401370177,
+      "tokens_seen": 363069440
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008988926336061629,
+      "loss": 0.0915,
+      "theoretical_loss": 4.08415956159115,
+      "tokens_seen": 363331584
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008988123896645804,
+      "loss": 0.0896,
+      "theoretical_loss": 4.083827029054773,
+      "tokens_seen": 363593728
+    },
+    {
+      "epoch": 0.11,
+      "objective/train/advantage_avg": 0.000944749335758388,
+      "objective/train/docs_used": 138686,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8428866863250732,
+      "objective/train/original_loss": 1.8428866863250732,
+      "objective/train/theoretical_loss": 4.083660877844774,
+      "objective/train/tokens_used": 384184800,
+      "objective/train/value_avg": -0.007354736328125,
+      "objective/train/value_loss": 0.0001762463798513636,
+      "objective/train/value_max": -0.00036406517028808594,
+      "objective/train/value_min": -0.16845703125,
+      "objective/train/value_reward_corr": 0.5633609780519697,
+      "objective/train/value_std": 0.008056640625,
+      "objective/train/weight_avg": 1.0010242462158203,
+      "objective/train/weighted_lm_loss": 1.8447622060775757,
+      "objective/train/weights_max": 1.0969452857971191,
+      "objective/train/weights_min": 0.3732965886592865,
+      "theoretical_loss": 4.083660877844774,
+      "tokens_seen": 363724800
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008987321457229979,
+      "loss": 0.0914,
+      "theoretical_loss": 4.0834948032562215,
+      "tokens_seen": 363855872
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008986519017814156,
+      "loss": 0.0914,
+      "theoretical_loss": 4.083162883691864,
+      "tokens_seen": 364118016
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008985716578398331,
+      "loss": 0.0915,
+      "theoretical_loss": 4.082831269859261,
+      "tokens_seen": 364380160
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008984914138982507,
+      "loss": 0.0904,
+      "theoretical_loss": 4.0824999612571515,
+      "tokens_seen": 364642304
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008984111699566683,
+      "loss": 0.0894,
+      "theoretical_loss": 4.082168957385462,
+      "tokens_seen": 364904448
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008983309260150859,
+      "loss": 0.092,
+      "theoretical_loss": 4.081838257745293,
+      "tokens_seen": 365166592
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008982506820735034,
+      "loss": 0.0877,
+      "theoretical_loss": 4.081507861838922,
+      "tokens_seen": 365428736
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.000898170438131921,
+      "loss": 0.0904,
+      "theoretical_loss": 4.081177769169795,
+      "tokens_seen": 365690880
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008980901941903387,
+      "loss": 0.0874,
+      "theoretical_loss": 4.080847979242526,
+      "tokens_seen": 365953024
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008980099502487562,
+      "loss": 0.0904,
+      "theoretical_loss": 4.0805184915628905,
+      "tokens_seen": 366215168
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008979297063071739,
+      "loss": 0.0896,
+      "theoretical_loss": 4.080189305637827,
+      "tokens_seen": 366477312
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008978494623655914,
+      "loss": 0.0898,
+      "theoretical_loss": 4.079860420975429,
+      "tokens_seen": 366739456
+    },
+    {
+      "epoch": 0.11,
+      "objective/train/advantage_avg": 0.0033440319821238518,
+      "objective/train/docs_used": 139849,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8659143447875977,
+      "objective/train/original_loss": 1.8659144639968872,
+      "objective/train/theoretical_loss": 4.07953183708494,
+      "objective/train/tokens_used": 387461600,
+      "objective/train/value_avg": -0.00817108154296875,
+      "objective/train/value_loss": 0.0002236411819467321,
+      "objective/train/value_max": -0.00025916099548339844,
+      "objective/train/value_min": -0.228515625,
+      "objective/train/value_reward_corr": 0.38717246449564546,
+      "objective/train/value_std": 0.00867462158203125,
+      "objective/train/weight_avg": 1.0034469366073608,
+      "objective/train/weighted_lm_loss": 1.8702045679092407,
+      "objective/train/weights_max": 1.2567331790924072,
+      "objective/train/weights_min": 0.3701472282409668,
+      "theoretical_loss": 4.07953183708494,
+      "tokens_seen": 367001600
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.000897769218424009,
+      "loss": 0.0949,
+      "theoretical_loss": 4.07953183708494,
+      "tokens_seen": 367001600
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008976889744824266,
+      "loss": 0.0914,
+      "theoretical_loss": 4.079203553476759,
+      "tokens_seen": 367263744
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008976087305408441,
+      "loss": 0.0919,
+      "theoretical_loss": 4.078875569662424,
+      "tokens_seen": 367525888
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008975284865992618,
+      "loss": 0.0886,
+      "theoretical_loss": 4.07854788515462,
+      "tokens_seen": 367788032
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008974482426576793,
+      "loss": 0.0896,
+      "theoretical_loss": 4.078220499467168,
+      "tokens_seen": 368050176
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.000897367998716097,
+      "loss": 0.09,
+      "theoretical_loss": 4.077893412115025,
+      "tokens_seen": 368312320
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008972877547745146,
+      "loss": 0.0878,
+      "theoretical_loss": 4.077566622614281,
+      "tokens_seen": 368574464
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008972075108329322,
+      "loss": 0.0894,
+      "theoretical_loss": 4.077240130482153,
+      "tokens_seen": 368836608
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008971272668913497,
+      "loss": 0.0914,
+      "theoretical_loss": 4.076913935236982,
+      "tokens_seen": 369098752
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008970470229497674,
+      "loss": 0.0894,
+      "theoretical_loss": 4.076588036398235,
+      "tokens_seen": 369360896
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008969667790081849,
+      "loss": 0.0891,
+      "theoretical_loss": 4.076262433486491,
+      "tokens_seen": 369623040
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008968865350666024,
+      "loss": 0.0913,
+      "theoretical_loss": 4.075937126023448,
+      "tokens_seen": 369885184
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008968062911250201,
+      "loss": 0.0888,
+      "theoretical_loss": 4.075612113531915,
+      "tokens_seen": 370147328
+    },
+    {
+      "epoch": 0.11,
+      "objective/train/advantage_avg": 0.0013215008657425642,
+      "objective/train/docs_used": 141078,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.754763126373291,
+      "objective/train/original_loss": 1.754763126373291,
+      "objective/train/theoretical_loss": 4.0754497177516456,
+      "objective/train/tokens_used": 390738400,
+      "objective/train/value_avg": -0.0116424560546875,
+      "objective/train/value_loss": 0.00028973660664632916,
+      "objective/train/value_max": -0.00014770030975341797,
+      "objective/train/value_min": -0.254150390625,
+      "objective/train/value_reward_corr": 0.5981599021798076,
+      "objective/train/value_std": 0.013458251953125,
+      "objective/train/weight_avg": 1.001459002494812,
+      "objective/train/weighted_lm_loss": 1.7561613321304321,
+      "objective/train/weights_max": 1.1360948085784912,
+      "objective/train/weights_min": 0.38701966404914856,
+      "theoretical_loss": 4.0754497177516456,
+      "tokens_seen": 370278400
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008967260471834376,
+      "loss": 0.0888,
+      "theoretical_loss": 4.075287395535807,
+      "tokens_seen": 370409472
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008966458032418553,
+      "loss": 0.0897,
+      "theoretical_loss": 4.074962971560145,
+      "tokens_seen": 370671616
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008965655593002729,
+      "loss": 0.0885,
+      "theoretical_loss": 4.074638841131052,
+      "tokens_seen": 370933760
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008964853153586905,
+      "loss": 0.0905,
+      "theoretical_loss": 4.074315003775746,
+      "tokens_seen": 371195904
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.000896405071417108,
+      "loss": 0.0917,
+      "theoretical_loss": 4.073991459022544,
+      "tokens_seen": 371458048
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008963248274755256,
+      "loss": 0.0871,
+      "theoretical_loss": 4.073668206400851,
+      "tokens_seen": 371720192
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008962445835339432,
+      "loss": 0.0954,
+      "theoretical_loss": 4.0733452454411605,
+      "tokens_seen": 371982336
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008961643395923608,
+      "loss": 0.0904,
+      "theoretical_loss": 4.0730225756750515,
+      "tokens_seen": 372244480
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008960840956507784,
+      "loss": 0.0879,
+      "theoretical_loss": 4.072700196635185,
+      "tokens_seen": 372506624
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008960038517091959,
+      "loss": 0.0906,
+      "theoretical_loss": 4.072378107855299,
+      "tokens_seen": 372768768
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008959236077676137,
+      "loss": 0.0866,
+      "theoretical_loss": 4.0720563088702075,
+      "tokens_seen": 373030912
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008958433638260312,
+      "loss": 0.089,
+      "theoretical_loss": 4.071734799215796,
+      "tokens_seen": 373293056
+    },
+    {
+      "epoch": 0.11,
+      "objective/train/advantage_avg": 0.0027087300550192595,
+      "objective/train/docs_used": 142333,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8659402132034302,
+      "objective/train/original_loss": 1.8659402132034302,
+      "objective/train/theoretical_loss": 4.071413578429017,
+      "objective/train/tokens_used": 394015200,
+      "objective/train/value_avg": -0.01194000244140625,
+      "objective/train/value_loss": 0.0003572187852114439,
+      "objective/train/value_max": -0.0005660057067871094,
+      "objective/train/value_min": -0.54833984375,
+      "objective/train/value_reward_corr": 0.6665229811822021,
+      "objective/train/value_std": 0.0186767578125,
+      "objective/train/weight_avg": 1.002875804901123,
+      "objective/train/weighted_lm_loss": 1.8703423738479614,
+      "objective/train/weights_max": 1.7303780317306519,
+      "objective/train/weights_min": 0.39453253149986267,
+      "theoretical_loss": 4.071413578429017,
+      "tokens_seen": 373555200
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008957631198844487,
+      "loss": 0.088,
+      "theoretical_loss": 4.071413578429017,
+      "tokens_seen": 373555200
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008956828759428664,
+      "loss": 0.0903,
+      "theoretical_loss": 4.071092646047892,
+      "tokens_seen": 373817344
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008956026320012839,
+      "loss": 0.0885,
+      "theoretical_loss": 4.0707720016115,
+      "tokens_seen": 374079488
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008955223880597015,
+      "loss": 0.0874,
+      "theoretical_loss": 4.0704516446599825,
+      "tokens_seen": 374341632
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008954421441181191,
+      "loss": 0.092,
+      "theoretical_loss": 4.070131574734536,
+      "tokens_seen": 374603776
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008953619001765367,
+      "loss": 0.0897,
+      "theoretical_loss": 4.069811791377409,
+      "tokens_seen": 374865920
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008952816562349542,
+      "loss": 0.0887,
+      "theoretical_loss": 4.0694922941319,
+      "tokens_seen": 375128064
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008952014122933718,
+      "loss": 0.0914,
+      "theoretical_loss": 4.069173082542351,
+      "tokens_seen": 375390208
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008951211683517895,
+      "loss": 0.0939,
+      "theoretical_loss": 4.068854156154154,
+      "tokens_seen": 375652352
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.000895040924410207,
+      "loss": 0.0905,
+      "theoretical_loss": 4.068535514513734,
+      "tokens_seen": 375914496
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008949606804686247,
+      "loss": 0.0918,
+      "theoretical_loss": 4.068217157168556,
+      "tokens_seen": 376176640
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008948804365270422,
+      "loss": 0.0897,
+      "theoretical_loss": 4.06789908366712,
+      "tokens_seen": 376438784
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008948001925854599,
+      "loss": 0.0918,
+      "theoretical_loss": 4.067581293558955,
+      "tokens_seen": 376700928
+    },
+    {
+      "epoch": 0.11,
+      "objective/train/advantage_avg": 0.0015242223162204027,
+      "objective/train/docs_used": 143470,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8477563858032227,
+      "objective/train/original_loss": 1.8477565050125122,
+      "objective/train/theoretical_loss": 4.067422504636857,
+      "objective/train/tokens_used": 397292000,
+      "objective/train/value_avg": -0.00897216796875,
+      "objective/train/value_loss": 0.00023800358758307993,
+      "objective/train/value_max": -0.00040459632873535156,
+      "objective/train/value_min": -0.291748046875,
+      "objective/train/value_reward_corr": 0.5821468077761229,
+      "objective/train/value_std": 0.0121917724609375,
+      "objective/train/weight_avg": 1.0016366243362427,
+      "objective/train/weighted_lm_loss": 1.8498717546463013,
+      "objective/train/weights_max": 1.1463937759399414,
+      "objective/train/weights_min": 0.3839375674724579,
+      "theoretical_loss": 4.067422504636857,
+      "tokens_seen": 376832000
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008947199486438774,
+      "loss": 0.0918,
+      "theoretical_loss": 4.0672637863946175,
+      "tokens_seen": 376963072
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008946397047022949,
+      "loss": 0.0885,
+      "theoretical_loss": 4.0669465617256915,
+      "tokens_seen": 377225216
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008945594607607126,
+      "loss": 0.0878,
+      "theoretical_loss": 4.06662961910478,
+      "tokens_seen": 377487360
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008944792168191301,
+      "loss": 0.0873,
+      "theoretical_loss": 4.066312958085503,
+      "tokens_seen": 377749504
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008943989728775478,
+      "loss": 0.088,
+      "theoretical_loss": 4.065996578222502,
+      "tokens_seen": 378011648
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008943187289359654,
+      "loss": 0.0904,
+      "theoretical_loss": 4.065680479071426,
+      "tokens_seen": 378273792
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.000894238484994383,
+      "loss": 0.0875,
+      "theoretical_loss": 4.0653646601889335,
+      "tokens_seen": 378535936
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008941582410528005,
+      "loss": 0.0891,
+      "theoretical_loss": 4.065049121132693,
+      "tokens_seen": 378798080
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008940779971112181,
+      "loss": 0.0881,
+      "theoretical_loss": 4.0647338614613755,
+      "tokens_seen": 379060224
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 0.0008939977531696357,
+      "loss": 0.0862,
+      "theoretical_loss": 4.0644188807346495,
+      "tokens_seen": 379322368
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008939175092280532,
+      "loss": 0.0906,
+      "theoretical_loss": 4.064104178513186,
+      "tokens_seen": 379584512
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008938372652864709,
+      "loss": 0.089,
+      "theoretical_loss": 4.0637897543586465,
+      "tokens_seen": 379846656
+    },
+    {
+      "epoch": 0.12,
+      "objective/train/advantage_avg": 0.002633684780448675,
+      "objective/train/docs_used": 144698,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.705493688583374,
+      "objective/train/original_loss": 1.705493688583374,
+      "objective/train/theoretical_loss": 4.063475607833687,
+      "objective/train/tokens_used": 400568800,
+      "objective/train/value_avg": -0.00835418701171875,
+      "objective/train/value_loss": 0.0002564003807492554,
+      "objective/train/value_max": -0.00031757354736328125,
+      "objective/train/value_min": -0.32568359375,
+      "objective/train/value_reward_corr": 0.6303237316396728,
+      "objective/train/value_std": 0.01141357421875,
+      "objective/train/weight_avg": 1.002751350402832,
+      "objective/train/weighted_lm_loss": 1.7094148397445679,
+      "objective/train/weights_max": 1.104806661605835,
+      "objective/train/weights_min": 0.3750951290130615,
+      "theoretical_loss": 4.063475607833687,
+      "tokens_seen": 380108800
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008937570213448884,
+      "loss": 0.087,
+      "theoretical_loss": 4.063475607833687,
+      "tokens_seen": 380108800
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008936767774033062,
+      "loss": 0.0893,
+      "theoretical_loss": 4.063161738501951,
+      "tokens_seen": 380370944
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008935965334617237,
+      "loss": 0.0892,
+      "theoretical_loss": 4.0628481459280685,
+      "tokens_seen": 380633088
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008935162895201413,
+      "loss": 0.0872,
+      "theoretical_loss": 4.062534829677653,
+      "tokens_seen": 380895232
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008934360455785589,
+      "loss": 0.0894,
+      "theoretical_loss": 4.062221789317297,
+      "tokens_seen": 381157376
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008933558016369764,
+      "loss": 0.0874,
+      "theoretical_loss": 4.061909024414572,
+      "tokens_seen": 381419520
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000893275557695394,
+      "loss": 0.088,
+      "theoretical_loss": 4.061596534538021,
+      "tokens_seen": 381681664
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008931953137538116,
+      "loss": 0.0913,
+      "theoretical_loss": 4.061284319257162,
+      "tokens_seen": 381943808
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008931150698122292,
+      "loss": 0.0907,
+      "theoretical_loss": 4.060972378142479,
+      "tokens_seen": 382205952
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008930348258706467,
+      "loss": 0.0902,
+      "theoretical_loss": 4.060660710765423,
+      "tokens_seen": 382468096
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008929545819290645,
+      "loss": 0.0914,
+      "theoretical_loss": 4.060349316698408,
+      "tokens_seen": 382730240
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000892874337987482,
+      "loss": 0.0879,
+      "theoretical_loss": 4.060038195514808,
+      "tokens_seen": 382992384
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008927940940458995,
+      "loss": 0.0868,
+      "theoretical_loss": 4.059727346788955,
+      "tokens_seen": 383254528
+    },
+    {
+      "epoch": 0.12,
+      "objective/train/advantage_avg": 0.0013565245317295194,
+      "objective/train/docs_used": 145827,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7567694187164307,
+      "objective/train/original_loss": 1.7567695379257202,
+      "objective/train/theoretical_loss": 4.059572024464923,
+      "objective/train/tokens_used": 403845600,
+      "objective/train/value_avg": -0.0077972412109375,
+      "objective/train/value_loss": 0.00035242707235738635,
+      "objective/train/value_max": -0.00033283233642578125,
+      "objective/train/value_min": -0.4189453125,
+      "objective/train/value_reward_corr": 0.5299108861611858,
+      "objective/train/value_std": 0.01129150390625,
+      "objective/train/weight_avg": 1.0015063285827637,
+      "objective/train/weighted_lm_loss": 1.7594711780548096,
+      "objective/train/weights_max": 1.2279499769210815,
+      "objective/train/weights_min": 0.36844298243522644,
+      "theoretical_loss": 4.059572024464923,
+      "tokens_seen": 383385600
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008927138501043172,
+      "loss": 0.0874,
+      "theoretical_loss": 4.059416770096134,
+      "tokens_seen": 383516672
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008926336061627347,
+      "loss": 0.0888,
+      "theoretical_loss": 4.059106465012583,
+      "tokens_seen": 383778816
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008925533622211523,
+      "loss": 0.0881,
+      "theoretical_loss": 4.058796431115489,
+      "tokens_seen": 384040960
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008924731182795699,
+      "loss": 0.0897,
+      "theoretical_loss": 4.058486667982986,
+      "tokens_seen": 384303104
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008923928743379875,
+      "loss": 0.0877,
+      "theoretical_loss": 4.058177175194148,
+      "tokens_seen": 384565248
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008923126303964051,
+      "loss": 0.0886,
+      "theoretical_loss": 4.057867952328994,
+      "tokens_seen": 384827392
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008922323864548226,
+      "loss": 0.0901,
+      "theoretical_loss": 4.057558998968479,
+      "tokens_seen": 385089536
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008921521425132403,
+      "loss": 0.0907,
+      "theoretical_loss": 4.0572503146944925,
+      "tokens_seen": 385351680
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008920718985716579,
+      "loss": 0.0887,
+      "theoretical_loss": 4.056941899089858,
+      "tokens_seen": 385613824
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008919916546300755,
+      "loss": 0.0885,
+      "theoretical_loss": 4.056633751738328,
+      "tokens_seen": 385875968
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000891911410688493,
+      "loss": 0.0879,
+      "theoretical_loss": 4.0563258722245825,
+      "tokens_seen": 386138112
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008918311667469107,
+      "loss": 0.0875,
+      "theoretical_loss": 4.056018260134226,
+      "tokens_seen": 386400256
+    },
+    {
+      "epoch": 0.12,
+      "objective/train/advantage_avg": 0.0018099743174389005,
+      "objective/train/docs_used": 146981,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.903164267539978,
+      "objective/train/original_loss": 1.9031643867492676,
+      "objective/train/theoretical_loss": 4.055710915053783,
+      "objective/train/tokens_used": 407122400,
+      "objective/train/value_avg": -0.0086517333984375,
+      "objective/train/value_loss": 0.0001866398670244962,
+      "objective/train/value_max": -0.0004425048828125,
+      "objective/train/value_min": -0.2122802734375,
+      "objective/train/value_reward_corr": 0.5988158937852228,
+      "objective/train/value_std": 0.01146697998046875,
+      "objective/train/weight_avg": 1.0019017457962036,
+      "objective/train/weighted_lm_loss": 1.9061700105667114,
+      "objective/train/weights_max": 1.1358146667480469,
+      "objective/train/weights_min": 0.6089951395988464,
+      "theoretical_loss": 4.055710915053783,
+      "tokens_seen": 386662400
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008917509228053282,
+      "loss": 0.0884,
+      "theoretical_loss": 4.055710915053783,
+      "tokens_seen": 386662400
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008916706788637457,
+      "loss": 0.0871,
+      "theoretical_loss": 4.055403836570701,
+      "tokens_seen": 386924544
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008915904349221634,
+      "loss": 0.0915,
+      "theoretical_loss": 4.05509702427334,
+      "tokens_seen": 387186688
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008915101909805809,
+      "loss": 0.0866,
+      "theoretical_loss": 4.054790477750974,
+      "tokens_seen": 387448832
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008914299470389986,
+      "loss": 0.0881,
+      "theoretical_loss": 4.054484196593791,
+      "tokens_seen": 387710976
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008913497030974162,
+      "loss": 0.0891,
+      "theoretical_loss": 4.054178180392885,
+      "tokens_seen": 387973120
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008912694591558338,
+      "loss": 0.0835,
+      "theoretical_loss": 4.053872428740256,
+      "tokens_seen": 388235264
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008911892152142514,
+      "loss": 0.0894,
+      "theoretical_loss": 4.053566941228809,
+      "tokens_seen": 388497408
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008911089712726689,
+      "loss": 0.0895,
+      "theoretical_loss": 4.053261717452346,
+      "tokens_seen": 388759552
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008910287273310865,
+      "loss": 0.0865,
+      "theoretical_loss": 4.052956757005573,
+      "tokens_seen": 389021696
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008909484833895041,
+      "loss": 0.0864,
+      "theoretical_loss": 4.0526520594840845,
+      "tokens_seen": 389283840
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008908682394479217,
+      "loss": 0.0906,
+      "theoretical_loss": 4.052347624484373,
+      "tokens_seen": 389545984
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008907879955063392,
+      "loss": 0.0874,
+      "theoretical_loss": 4.052043451603818,
+      "tokens_seen": 389808128
+    },
+    {
+      "epoch": 0.12,
+      "objective/train/advantage_avg": 0.002918938407674432,
+      "objective/train/docs_used": 148229,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8769011497497559,
+      "objective/train/original_loss": 1.8769011497497559,
+      "objective/train/theoretical_loss": 4.051891463332648,
+      "objective/train/tokens_used": 410399200,
+      "objective/train/value_avg": -0.00745391845703125,
+      "objective/train/value_loss": 0.00021266264957375824,
+      "objective/train/value_max": -0.0002779960632324219,
+      "objective/train/value_min": -0.27587890625,
+      "objective/train/value_reward_corr": 0.5017984739541499,
+      "objective/train/value_std": 0.0098876953125,
+      "objective/train/weight_avg": 1.003010869026184,
+      "objective/train/weighted_lm_loss": 1.8814412355422974,
+      "objective/train/weights_max": 1.1254287958145142,
+      "objective/train/weights_min": 0.380566269159317,
+      "theoretical_loss": 4.051891463332648,
+      "tokens_seen": 389939200
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000890707751564757,
+      "loss": 0.0897,
+      "theoretical_loss": 4.051739540440688,
+      "tokens_seen": 390070272
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008906275076231745,
+      "loss": 0.0865,
+      "theoretical_loss": 4.0514358905941386,
+      "tokens_seen": 390332416
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008905472636815921,
+      "loss": 0.0881,
+      "theoretical_loss": 4.051132501664204,
+      "tokens_seen": 390594560
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008904670197400097,
+      "loss": 0.0898,
+      "theoretical_loss": 4.050829373251803,
+      "tokens_seen": 390856704
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008903867757984272,
+      "loss": 0.0905,
+      "theoretical_loss": 4.050526504958727,
+      "tokens_seen": 391118848
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008903065318568448,
+      "loss": 0.0853,
+      "theoretical_loss": 4.050223896387647,
+      "tokens_seen": 391380992
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008902262879152624,
+      "loss": 0.0879,
+      "theoretical_loss": 4.0499215471421035,
+      "tokens_seen": 391643136
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.00089014604397368,
+      "loss": 0.0888,
+      "theoretical_loss": 4.049619456826508,
+      "tokens_seen": 391905280
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008900658000320975,
+      "loss": 0.0875,
+      "theoretical_loss": 4.0493176250461405,
+      "tokens_seen": 392167424
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008899855560905153,
+      "loss": 0.0879,
+      "theoretical_loss": 4.049016051407145,
+      "tokens_seen": 392429568
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008899053121489328,
+      "loss": 0.087,
+      "theoretical_loss": 4.048714735516527,
+      "tokens_seen": 392691712
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008898250682073504,
+      "loss": 0.0909,
+      "theoretical_loss": 4.048413676982155,
+      "tokens_seen": 392953856
+    },
+    {
+      "epoch": 0.12,
+      "objective/train/advantage_avg": 0.0008797993650659919,
+      "objective/train/docs_used": 149327,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8922388553619385,
+      "objective/train/original_loss": 1.8922390937805176,
+      "objective/train/theoretical_loss": 4.048112875412752,
+      "objective/train/tokens_used": 413676000,
+      "objective/train/value_avg": -0.0099639892578125,
+      "objective/train/value_loss": 0.00043581624049693346,
+      "objective/train/value_max": -0.0004076957702636719,
+      "objective/train/value_min": -0.87744140625,
+      "objective/train/value_reward_corr": 0.6595999360248977,
+      "objective/train/value_std": 0.01849365234375,
+      "objective/train/weight_avg": 1.001081109046936,
+      "objective/train/weighted_lm_loss": 1.8936957120895386,
+      "objective/train/weights_max": 1.6510648727416992,
+      "objective/train/weights_min": 0.36863452196121216,
+      "theoretical_loss": 4.048112875412752,
+      "tokens_seen": 393216000
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000889744824265768,
+      "loss": 0.0877,
+      "theoretical_loss": 4.048112875412752,
+      "tokens_seen": 393216000
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008896645803241855,
+      "loss": 0.0895,
+      "theoretical_loss": 4.0478123304179,
+      "tokens_seen": 393478144
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008895843363826032,
+      "loss": 0.0908,
+      "theoretical_loss": 4.047512041608029,
+      "tokens_seen": 393740288
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008895040924410207,
+      "loss": 0.0891,
+      "theoretical_loss": 4.047212008594424,
+      "tokens_seen": 394002432
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008894238484994383,
+      "loss": 0.0877,
+      "theoretical_loss": 4.046912230989214,
+      "tokens_seen": 394264576
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008893436045578559,
+      "loss": 0.0898,
+      "theoretical_loss": 4.0466127084053785,
+      "tokens_seen": 394526720
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008892633606162734,
+      "loss": 0.0882,
+      "theoretical_loss": 4.046313440456733,
+      "tokens_seen": 394788864
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008891831166746911,
+      "loss": 0.0893,
+      "theoretical_loss": 4.0460144267579405,
+      "tokens_seen": 395051008
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008891028727331087,
+      "loss": 0.0901,
+      "theoretical_loss": 4.045715666924499,
+      "tokens_seen": 395313152
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008890226287915263,
+      "loss": 0.0885,
+      "theoretical_loss": 4.045417160572743,
+      "tokens_seen": 395575296
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008889423848499438,
+      "loss": 0.0904,
+      "theoretical_loss": 4.045118907319839,
+      "tokens_seen": 395837440
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008888621409083615,
+      "loss": 0.0868,
+      "theoretical_loss": 4.04482090678379,
+      "tokens_seen": 396099584
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000888781896966779,
+      "loss": 0.0911,
+      "theoretical_loss": 4.044523158583421,
+      "tokens_seen": 396361728
+    },
+    {
+      "epoch": 0.12,
+      "objective/train/advantage_avg": 0.0021504259202629328,
+      "objective/train/docs_used": 150450,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7327624559402466,
+      "objective/train/original_loss": 1.732762336730957,
+      "objective/train/theoretical_loss": 4.044374378990227,
+      "objective/train/tokens_used": 416952800,
+      "objective/train/value_avg": -0.00965118408203125,
+      "objective/train/value_loss": 0.00036191148683428764,
+      "objective/train/value_max": -0.00023055076599121094,
+      "objective/train/value_min": -0.8271484375,
+      "objective/train/value_reward_corr": 0.6954772035577824,
+      "objective/train/value_std": 0.01971435546875,
+      "objective/train/weight_avg": 1.0023201704025269,
+      "objective/train/weighted_lm_loss": 1.7352375984191895,
+      "objective/train/weights_max": 1.7499204874038696,
+      "objective/train/weights_min": 0.37855079770088196,
+      "theoretical_loss": 4.044374378990227,
+      "tokens_seen": 396492800
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008887016530251965,
+      "loss": 0.0896,
+      "theoretical_loss": 4.044225662338388,
+      "tokens_seen": 396623872
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008886214090836142,
+      "loss": 0.0886,
+      "theoretical_loss": 4.04392841766917,
+      "tokens_seen": 396886016
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008885411651420317,
+      "loss": 0.0865,
+      "theoretical_loss": 4.043631424197067,
+      "tokens_seen": 397148160
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008884609212004495,
+      "loss": 0.09,
+      "theoretical_loss": 4.0433346815442,
+      "tokens_seen": 397410304
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000888380677258867,
+      "loss": 0.0919,
+      "theoretical_loss": 4.043038189333508,
+      "tokens_seen": 397672448
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008883004333172846,
+      "loss": 0.0924,
+      "theoretical_loss": 4.042741947188741,
+      "tokens_seen": 397934592
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008882201893757022,
+      "loss": 0.0876,
+      "theoretical_loss": 4.042445954734466,
+      "tokens_seen": 398196736
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008881399454341197,
+      "loss": 0.0926,
+      "theoretical_loss": 4.0421502115960575,
+      "tokens_seen": 398458880
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008880597014925373,
+      "loss": 0.0855,
+      "theoretical_loss": 4.0418547173997,
+      "tokens_seen": 398721024
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008879794575509549,
+      "loss": 0.0885,
+      "theoretical_loss": 4.041559471772382,
+      "tokens_seen": 398983168
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008878992136093725,
+      "loss": 0.0909,
+      "theoretical_loss": 4.041264474341896,
+      "tokens_seen": 399245312
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.00088781896966779,
+      "loss": 0.0891,
+      "theoretical_loss": 4.040969724736838,
+      "tokens_seen": 399507456
+    },
+    {
+      "epoch": 0.12,
+      "objective/train/advantage_avg": 0.001554798916913569,
+      "objective/train/docs_used": 151634,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7531509399414062,
+      "objective/train/original_loss": 1.7531508207321167,
+      "objective/train/theoretical_loss": 4.040675222586599,
+      "objective/train/tokens_used": 420229600,
+      "objective/train/value_avg": -0.009552001953125,
+      "objective/train/value_loss": 0.00028178023057989776,
+      "objective/train/value_max": -0.0002892017364501953,
+      "objective/train/value_min": -0.87255859375,
+      "objective/train/value_reward_corr": 0.7821445899209426,
+      "objective/train/value_std": 0.0219268798828125,
+      "objective/train/weight_avg": 1.0016918182373047,
+      "objective/train/weighted_lm_loss": 1.755144476890564,
+      "objective/train/weights_max": 1.6336557865142822,
+      "objective/train/weights_min": 0.541962206363678,
+      "theoretical_loss": 4.040675222586599,
+      "tokens_seen": 399769600
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008877387257262078,
+      "loss": 0.0863,
+      "theoretical_loss": 4.040675222586599,
+      "tokens_seen": 399769600
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008876584817846253,
+      "loss": 0.0854,
+      "theoretical_loss": 4.04038096752137,
+      "tokens_seen": 400031744
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008875782378430429,
+      "loss": 0.0886,
+      "theoretical_loss": 4.040086959172136,
+      "tokens_seen": 400293888
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008874979939014605,
+      "loss": 0.085,
+      "theoretical_loss": 4.039793197170672,
+      "tokens_seen": 400556032
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000887417749959878,
+      "loss": 0.0898,
+      "theoretical_loss": 4.039499681149547,
+      "tokens_seen": 400818176
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008873375060182957,
+      "loss": 0.0889,
+      "theoretical_loss": 4.039206410742114,
+      "tokens_seen": 401080320
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008872572620767132,
+      "loss": 0.0868,
+      "theoretical_loss": 4.038913385582515,
+      "tokens_seen": 401342464
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008871770181351308,
+      "loss": 0.0871,
+      "theoretical_loss": 4.038620605305673,
+      "tokens_seen": 401604608
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008870967741935484,
+      "loss": 0.0847,
+      "theoretical_loss": 4.038328069547293,
+      "tokens_seen": 401866752
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000887016530251966,
+      "loss": 0.0879,
+      "theoretical_loss": 4.03803577794386,
+      "tokens_seen": 402128896
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008869362863103836,
+      "loss": 0.0887,
+      "theoretical_loss": 4.037743730132635,
+      "tokens_seen": 402391040
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008868560423688012,
+      "loss": 0.0855,
+      "theoretical_loss": 4.037451925751654,
+      "tokens_seen": 402653184
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008867757984272188,
+      "loss": 0.088,
+      "theoretical_loss": 4.0371603644397265,
+      "tokens_seen": 402915328
+    },
+    {
+      "epoch": 0.12,
+      "objective/train/advantage_avg": 0.0015503467293456197,
+      "objective/train/docs_used": 152845,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6842749118804932,
+      "objective/train/original_loss": 1.6842749118804932,
+      "objective/train/theoretical_loss": 4.037014674821996,
+      "objective/train/tokens_used": 423506400,
+      "objective/train/value_avg": -0.013092041015625,
+      "objective/train/value_loss": 0.0004618663515429944,
+      "objective/train/value_max": -0.00034737586975097656,
+      "objective/train/value_min": -0.5283203125,
+      "objective/train/value_reward_corr": 0.5689597227558094,
+      "objective/train/value_std": 0.016571044921875,
+      "objective/train/weight_avg": 1.0017578601837158,
+      "objective/train/weighted_lm_loss": 1.6876652240753174,
+      "objective/train/weights_max": 1.291870355606079,
+      "objective/train/weights_min": 0.36923545598983765,
+      "theoretical_loss": 4.037014674821996,
+      "tokens_seen": 403046400
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008866955544856363,
+      "loss": 0.0901,
+      "theoretical_loss": 4.03686904583643,
+      "tokens_seen": 403177472
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000886615310544054,
+      "loss": 0.0872,
+      "theoretical_loss": 4.036577969582114,
+      "tokens_seen": 403439616
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008865350666024715,
+      "loss": 0.0862,
+      "theoretical_loss": 4.03628713531789,
+      "tokens_seen": 403701760
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008864548226608891,
+      "loss": 0.0896,
+      "theoretical_loss": 4.035996542685638,
+      "tokens_seen": 403963904
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008863745787193067,
+      "loss": 0.0864,
+      "theoretical_loss": 4.0357061913279955,
+      "tokens_seen": 404226048
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008862943347777242,
+      "loss": 0.0873,
+      "theoretical_loss": 4.035416080888364,
+      "tokens_seen": 404488192
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008862140908361419,
+      "loss": 0.0896,
+      "theoretical_loss": 4.035126211010899,
+      "tokens_seen": 404750336
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008861338468945595,
+      "loss": 0.0856,
+      "theoretical_loss": 4.034836581340515,
+      "tokens_seen": 405012480
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008860536029529771,
+      "loss": 0.0872,
+      "theoretical_loss": 4.034547191522877,
+      "tokens_seen": 405274624
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008859733590113947,
+      "loss": 0.0932,
+      "theoretical_loss": 4.034258041204404,
+      "tokens_seen": 405536768
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008858931150698123,
+      "loss": 0.0875,
+      "theoretical_loss": 4.033969130032263,
+      "tokens_seen": 405798912
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008858128711282298,
+      "loss": 0.0861,
+      "theoretical_loss": 4.033680457654368,
+      "tokens_seen": 406061056
+    },
+    {
+      "epoch": 0.12,
+      "objective/train/advantage_avg": 0.0022317736875265837,
+      "objective/train/docs_used": 154148,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6351722478866577,
+      "objective/train/original_loss": 1.6351724863052368,
+      "objective/train/theoretical_loss": 4.0333920237193785,
+      "objective/train/tokens_used": 426783200,
+      "objective/train/value_avg": -0.00635528564453125,
+      "objective/train/value_loss": 0.00012135026918258518,
+      "objective/train/value_max": -0.00023055076599121094,
+      "objective/train/value_min": -0.303955078125,
+      "objective/train/value_reward_corr": 0.5124807247227108,
+      "objective/train/value_std": 0.0070343017578125,
+      "objective/train/weight_avg": 1.00228750705719,
+      "objective/train/weighted_lm_loss": 1.6392003297805786,
+      "objective/train/weights_max": 1.1095505952835083,
+      "objective/train/weights_min": 0.37000179290771484,
+      "theoretical_loss": 4.0333920237193785,
+      "tokens_seen": 406323200
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008857326271866474,
+      "loss": 0.0859,
+      "theoretical_loss": 4.0333920237193785,
+      "tokens_seen": 406323200
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000885652383245065,
+      "loss": 0.0823,
+      "theoretical_loss": 4.0331038278766975,
+      "tokens_seen": 406585344
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008855721393034825,
+      "loss": 0.0895,
+      "theoretical_loss": 4.032815869776471,
+      "tokens_seen": 406847488
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008854918953619003,
+      "loss": 0.089,
+      "theoretical_loss": 4.032528149069579,
+      "tokens_seen": 407109632
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008854116514203178,
+      "loss": 0.0891,
+      "theoretical_loss": 4.0322406654076435,
+      "tokens_seen": 407371776
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008853314074787354,
+      "loss": 0.0893,
+      "theoretical_loss": 4.0319534184430195,
+      "tokens_seen": 407633920
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000885251163537153,
+      "loss": 0.0867,
+      "theoretical_loss": 4.031666407828795,
+      "tokens_seen": 407896064
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008851709195955705,
+      "loss": 0.0867,
+      "theoretical_loss": 4.03137963321879,
+      "tokens_seen": 408158208
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008850906756539881,
+      "loss": 0.089,
+      "theoretical_loss": 4.0310930942675505,
+      "tokens_seen": 408420352
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008850104317124057,
+      "loss": 0.0878,
+      "theoretical_loss": 4.030806790630353,
+      "tokens_seen": 408682496
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008849301877708233,
+      "loss": 0.0856,
+      "theoretical_loss": 4.030520721963199,
+      "tokens_seen": 408944640
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008848499438292408,
+      "loss": 0.0855,
+      "theoretical_loss": 4.030234887922808,
+      "tokens_seen": 409206784
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008847696998876586,
+      "loss": 0.0877,
+      "theoretical_loss": 4.029949288166627,
+      "tokens_seen": 409468928
+    },
+    {
+      "epoch": 0.12,
+      "objective/train/advantage_avg": -0.0004230485938023776,
+      "objective/train/docs_used": 155224,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8668181896209717,
+      "objective/train/original_loss": 1.8668180704116821,
+      "objective/train/theoretical_loss": 4.029806576038263,
+      "objective/train/tokens_used": 430060000,
+      "objective/train/value_avg": -0.00667572021484375,
+      "objective/train/value_loss": 0.00026823318330571055,
+      "objective/train/value_max": -0.0002512931823730469,
+      "objective/train/value_min": -0.2183837890625,
+      "objective/train/value_reward_corr": 0.5231429303787112,
+      "objective/train/value_std": 0.0097808837890625,
+      "objective/train/weight_avg": 0.9997006058692932,
+      "objective/train/weighted_lm_loss": 1.864938735961914,
+      "objective/train/weights_max": 1.1423146724700928,
+      "objective/train/weights_min": 0.36977463960647583,
+      "theoretical_loss": 4.029806576038263,
+      "tokens_seen": 409600000
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008846894559460761,
+      "loss": 0.0827,
+      "theoretical_loss": 4.0296639223528175,
+      "tokens_seen": 409731072
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008846092120044937,
+      "loss": 0.0863,
+      "theoretical_loss": 4.029378790140261,
+      "tokens_seen": 409993216
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008845289680629113,
+      "loss": 0.0867,
+      "theoretical_loss": 4.029093891188552,
+      "tokens_seen": 410255360
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008844487241213288,
+      "loss": 0.0881,
+      "theoretical_loss": 4.028809225158,
+      "tokens_seen": 410517504
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008843684801797465,
+      "loss": 0.0862,
+      "theoretical_loss": 4.028524791709621,
+      "tokens_seen": 410779648
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000884288236238164,
+      "loss": 0.0867,
+      "theoretical_loss": 4.028240590505148,
+      "tokens_seen": 411041792
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008842079922965816,
+      "loss": 0.0864,
+      "theoretical_loss": 4.027956621207015,
+      "tokens_seen": 411303936
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008841277483549992,
+      "loss": 0.0861,
+      "theoretical_loss": 4.027672883478364,
+      "tokens_seen": 411566080
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008840475044134169,
+      "loss": 0.0897,
+      "theoretical_loss": 4.027389376983041,
+      "tokens_seen": 411828224
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.0008839672604718344,
+      "loss": 0.0898,
+      "theoretical_loss": 4.02710610138559,
+      "tokens_seen": 412090368
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 0.000883887016530252,
+      "loss": 0.0868,
+      "theoretical_loss": 4.02682305635126,
+      "tokens_seen": 412352512
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008838067725886696,
+      "loss": 0.0879,
+      "theoretical_loss": 4.026540241545994,
+      "tokens_seen": 412614656
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": 0.0008909243042580783,
+      "objective/train/docs_used": 156351,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.63358473777771,
+      "objective/train/original_loss": 1.633584976196289,
+      "objective/train/theoretical_loss": 4.026257656636431,
+      "objective/train/tokens_used": 433336800,
+      "objective/train/value_avg": -0.00994110107421875,
+      "objective/train/value_loss": 0.0007588982116430998,
+      "objective/train/value_max": -0.00039196014404296875,
+      "objective/train/value_min": -0.9423828125,
+      "objective/train/value_reward_corr": 0.6866913112731529,
+      "objective/train/value_std": 0.0245361328125,
+      "objective/train/weight_avg": 1.0012253522872925,
+      "objective/train/weighted_lm_loss": 1.6334694623947144,
+      "objective/train/weights_max": 2.1991829872131348,
+      "objective/train/weights_min": 0.3683927655220032,
+      "theoretical_loss": 4.026257656636431,
+      "tokens_seen": 412876800
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008837265286470871,
+      "loss": 0.086,
+      "theoretical_loss": 4.026257656636431,
+      "tokens_seen": 412876800
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008836462847055048,
+      "loss": 0.0873,
+      "theoretical_loss": 4.025975301289906,
+      "tokens_seen": 413138944
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008835660407639223,
+      "loss": 0.0868,
+      "theoretical_loss": 4.025693175174443,
+      "tokens_seen": 413401088
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.00088348579682234,
+      "loss": 0.084,
+      "theoretical_loss": 4.02541127795876,
+      "tokens_seen": 413663232
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008834055528807575,
+      "loss": 0.0864,
+      "theoretical_loss": 4.02512960931226,
+      "tokens_seen": 413925376
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.000883325308939175,
+      "loss": 0.0834,
+      "theoretical_loss": 4.024848168905035,
+      "tokens_seen": 414187520
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008832450649975928,
+      "loss": 0.0868,
+      "theoretical_loss": 4.02456695640786,
+      "tokens_seen": 414449664
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008831648210560103,
+      "loss": 0.0859,
+      "theoretical_loss": 4.0242859714921915,
+      "tokens_seen": 414711808
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008830845771144279,
+      "loss": 0.0889,
+      "theoretical_loss": 4.024005213830171,
+      "tokens_seen": 414973952
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008830043331728455,
+      "loss": 0.0838,
+      "theoretical_loss": 4.023724683094615,
+      "tokens_seen": 415236096
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008829240892312631,
+      "loss": 0.0836,
+      "theoretical_loss": 4.023444378959019,
+      "tokens_seen": 415498240
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008828438452896806,
+      "loss": 0.0853,
+      "theoretical_loss": 4.023164301097555,
+      "tokens_seen": 415760384
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008827636013480982,
+      "loss": 0.0857,
+      "theoretical_loss": 4.0228844491850655,
+      "tokens_seen": 416022528
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": 0.002180003095418215,
+      "objective/train/docs_used": 157543,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5781835317611694,
+      "objective/train/original_loss": 1.5781837701797485,
+      "objective/train/theoretical_loss": 4.022744607858259,
+      "objective/train/tokens_used": 436613600,
+      "objective/train/value_avg": -0.008880615234375,
+      "objective/train/value_loss": 0.00037475841236300766,
+      "objective/train/value_max": -0.0002397298812866211,
+      "objective/train/value_min": -0.37060546875,
+      "objective/train/value_reward_corr": 0.5816138183738614,
+      "objective/train/value_std": 0.0136260986328125,
+      "objective/train/weight_avg": 1.0023459196090698,
+      "objective/train/weighted_lm_loss": 1.5822744369506836,
+      "objective/train/weights_max": 1.4486113786697388,
+      "objective/train/weights_min": 0.37067708373069763,
+      "theoretical_loss": 4.022744607858259,
+      "tokens_seen": 416153600
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008826833574065158,
+      "loss": 0.0879,
+      "theoretical_loss": 4.022604822897068,
+      "tokens_seen": 416284672
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008826031134649333,
+      "loss": 0.0876,
+      "theoretical_loss": 4.0223254219097475,
+      "tokens_seen": 416546816
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008825228695233511,
+      "loss": 0.0878,
+      "theoretical_loss": 4.022046245899958,
+      "tokens_seen": 416808960
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008824426255817686,
+      "loss": 0.0861,
+      "theoretical_loss": 4.021767294545221,
+      "tokens_seen": 417071104
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008823623816401862,
+      "loss": 0.0877,
+      "theoretical_loss": 4.021488567523721,
+      "tokens_seen": 417333248
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008822821376986038,
+      "loss": 0.0875,
+      "theoretical_loss": 4.021210064514305,
+      "tokens_seen": 417595392
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008822018937570213,
+      "loss": 0.0835,
+      "theoretical_loss": 4.020931785196484,
+      "tokens_seen": 417857536
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.000882121649815439,
+      "loss": 0.0852,
+      "theoretical_loss": 4.020653729250424,
+      "tokens_seen": 418119680
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008820414058738565,
+      "loss": 0.0851,
+      "theoretical_loss": 4.020375896356951,
+      "tokens_seen": 418381824
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008819611619322741,
+      "loss": 0.0909,
+      "theoretical_loss": 4.0200982861975465,
+      "tokens_seen": 418643968
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008818809179906917,
+      "loss": 0.0883,
+      "theoretical_loss": 4.019820898454345,
+      "tokens_seen": 418906112
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008818006740491094,
+      "loss": 0.0843,
+      "theoretical_loss": 4.019543732810134,
+      "tokens_seen": 419168256
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": 0.0019043140346184373,
+      "objective/train/docs_used": 158751,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.881985068321228,
+      "objective/train/original_loss": 1.881984829902649,
+      "objective/train/theoretical_loss": 4.019266788948352,
+      "objective/train/tokens_used": 439890400,
+      "objective/train/value_avg": -0.008880615234375,
+      "objective/train/value_loss": 0.00036503959563560784,
+      "objective/train/value_max": -0.0002892017364501953,
+      "objective/train/value_min": -0.80029296875,
+      "objective/train/value_reward_corr": 0.6310464755810599,
+      "objective/train/value_std": 0.01422882080078125,
+      "objective/train/weight_avg": 1.0020649433135986,
+      "objective/train/weighted_lm_loss": 1.8862380981445312,
+      "objective/train/weights_max": 1.4735807180404663,
+      "objective/train/weights_min": 0.3895317018032074,
+      "theoretical_loss": 4.019266788948352,
+      "tokens_seen": 419430400
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008817204301075269,
+      "loss": 0.0876,
+      "theoretical_loss": 4.019266788948352,
+      "tokens_seen": 419430400
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008816401861659445,
+      "loss": 0.0901,
+      "theoretical_loss": 4.0189900665530836,
+      "tokens_seen": 419692544
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008815599422243621,
+      "loss": 0.0881,
+      "theoretical_loss": 4.0187135653090635,
+      "tokens_seen": 419954688
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008814796982827796,
+      "loss": 0.0832,
+      "theoretical_loss": 4.018437284901671,
+      "tokens_seen": 420216832
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008813994543411973,
+      "loss": 0.09,
+      "theoretical_loss": 4.018161225016926,
+      "tokens_seen": 420478976
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008813192103996148,
+      "loss": 0.088,
+      "theoretical_loss": 4.0178853853414935,
+      "tokens_seen": 420741120
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008812389664580324,
+      "loss": 0.0876,
+      "theoretical_loss": 4.017609765562678,
+      "tokens_seen": 421003264
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.00088115872251645,
+      "loss": 0.0874,
+      "theoretical_loss": 4.017334365368422,
+      "tokens_seen": 421265408
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008810784785748677,
+      "loss": 0.0892,
+      "theoretical_loss": 4.017059184447303,
+      "tokens_seen": 421527552
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008809982346332853,
+      "loss": 0.088,
+      "theoretical_loss": 4.0167842224885355,
+      "tokens_seen": 421789696
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008809179906917028,
+      "loss": 0.0832,
+      "theoretical_loss": 4.016509479181968,
+      "tokens_seen": 422051840
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008808377467501204,
+      "loss": 0.0867,
+      "theoretical_loss": 4.016234954218078,
+      "tokens_seen": 422313984
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.000880757502808538,
+      "loss": 0.086,
+      "theoretical_loss": 4.015960647287975,
+      "tokens_seen": 422576128
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": 0.000816687592305243,
+      "objective/train/docs_used": 159824,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6725859642028809,
+      "objective/train/original_loss": 1.6725860834121704,
+      "objective/train/theoretical_loss": 4.015823575489237,
+      "objective/train/tokens_used": 443167200,
+      "objective/train/value_avg": -0.00794219970703125,
+      "objective/train/value_loss": 0.0004758470749948174,
+      "objective/train/value_max": -0.0003101825714111328,
+      "objective/train/value_min": -0.78759765625,
+      "objective/train/value_reward_corr": 0.706160026017736,
+      "objective/train/value_std": 0.0199432373046875,
+      "objective/train/weight_avg": 1.0010310411453247,
+      "objective/train/weighted_lm_loss": 1.6731016635894775,
+      "objective/train/weights_max": 1.929433822631836,
+      "objective/train/weights_min": 0.39345037937164307,
+      "theoretical_loss": 4.015823575489237,
+      "tokens_seen": 422707200
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008806772588669556,
+      "loss": 0.0878,
+      "theoretical_loss": 4.015686558083396,
+      "tokens_seen": 422838272
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008805970149253731,
+      "loss": 0.0882,
+      "theoretical_loss": 4.015412686296704,
+      "tokens_seen": 423100416
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008805167709837908,
+      "loss": 0.0873,
+      "theoretical_loss": 4.0151390316208895,
+      "tokens_seen": 423362560
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008804365270422083,
+      "loss": 0.0878,
+      "theoretical_loss": 4.014865593749563,
+      "tokens_seen": 423624704
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008803562831006258,
+      "loss": 0.0855,
+      "theoretical_loss": 4.014592372376958,
+      "tokens_seen": 423886848
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008802760391590436,
+      "loss": 0.0876,
+      "theoretical_loss": 4.014319367197928,
+      "tokens_seen": 424148992
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008801957952174611,
+      "loss": 0.0907,
+      "theoretical_loss": 4.014046577907946,
+      "tokens_seen": 424411136
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008801155512758787,
+      "loss": 0.089,
+      "theoretical_loss": 4.013774004203099,
+      "tokens_seen": 424673280
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008800353073342963,
+      "loss": 0.0877,
+      "theoretical_loss": 4.013501645780092,
+      "tokens_seen": 424935424
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008799550633927139,
+      "loss": 0.0839,
+      "theoretical_loss": 4.013229502336242,
+      "tokens_seen": 425197568
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008798748194511314,
+      "loss": 0.0868,
+      "theoretical_loss": 4.0129575735694765,
+      "tokens_seen": 425459712
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.000879794575509549,
+      "loss": 0.0894,
+      "theoretical_loss": 4.012685859178337,
+      "tokens_seen": 425721856
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": 0.0018310642335563898,
+      "objective/train/docs_used": 161088,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5738030672073364,
+      "objective/train/original_loss": 1.5738029479980469,
+      "objective/train/theoretical_loss": 4.012414358861969,
+      "objective/train/tokens_used": 446444000,
+      "objective/train/value_avg": -0.006816864013671875,
+      "objective/train/value_loss": 0.00017067580483853817,
+      "objective/train/value_max": -0.0003542900085449219,
+      "objective/train/value_min": -0.53857421875,
+      "objective/train/value_reward_corr": 0.5243874673729859,
+      "objective/train/value_std": 0.009002685546875,
+      "objective/train/weight_avg": 1.0019116401672363,
+      "objective/train/weighted_lm_loss": 1.5763643980026245,
+      "objective/train/weights_max": 1.7135618925094604,
+      "objective/train/weights_min": 0.3707987070083618,
+      "theoretical_loss": 4.012414358861969,
+      "tokens_seen": 425984000
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008797143315679666,
+      "loss": 0.0845,
+      "theoretical_loss": 4.012414358861969,
+      "tokens_seen": 425984000
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008796340876263842,
+      "loss": 0.0872,
+      "theoretical_loss": 4.01214307232013,
+      "tokens_seen": 426246144
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008795538436848019,
+      "loss": 0.0868,
+      "theoretical_loss": 4.011871999253178,
+      "tokens_seen": 426508288
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008794735997432194,
+      "loss": 0.0889,
+      "theoretical_loss": 4.011601139362078,
+      "tokens_seen": 426770432
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008793933558016371,
+      "loss": 0.0858,
+      "theoretical_loss": 4.011330492348397,
+      "tokens_seen": 427032576
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008793131118600546,
+      "loss": 0.0894,
+      "theoretical_loss": 4.0110600579143,
+      "tokens_seen": 427294720
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008792328679184721,
+      "loss": 0.0869,
+      "theoretical_loss": 4.010789835762555,
+      "tokens_seen": 427556864
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008791526239768898,
+      "loss": 0.0891,
+      "theoretical_loss": 4.010519825596525,
+      "tokens_seen": 427819008
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008790723800353073,
+      "loss": 0.0865,
+      "theoretical_loss": 4.010250027120169,
+      "tokens_seen": 428081152
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008789921360937249,
+      "loss": 0.0857,
+      "theoretical_loss": 4.009980440038043,
+      "tokens_seen": 428343296
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008789118921521425,
+      "loss": 0.0893,
+      "theoretical_loss": 4.009711064055291,
+      "tokens_seen": 428605440
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008788316482105602,
+      "loss": 0.0856,
+      "theoretical_loss": 4.009441898877652,
+      "tokens_seen": 428867584
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008787514042689777,
+      "loss": 0.0858,
+      "theoretical_loss": 4.009172944211455,
+      "tokens_seen": 429129728
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": 0.0020093463826924562,
+      "objective/train/docs_used": 162190,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7857974767684937,
+      "objective/train/original_loss": 1.7857975959777832,
+      "objective/train/theoretical_loss": 4.009038545728536,
+      "objective/train/tokens_used": 449720800,
+      "objective/train/value_avg": -0.00783538818359375,
+      "objective/train/value_loss": 0.0001668190088821575,
+      "objective/train/value_max": -0.00030541419982910156,
+      "objective/train/value_min": -0.62890625,
+      "objective/train/value_reward_corr": 0.671483245399306,
+      "objective/train/value_std": 0.01235198974609375,
+      "objective/train/weight_avg": 1.0020873546600342,
+      "objective/train/weighted_lm_loss": 1.7891877889633179,
+      "objective/train/weights_max": 1.2835239171981812,
+      "objective/train/weights_min": 0.3886055648326874,
+      "theoretical_loss": 4.009038545728536,
+      "tokens_seen": 429260800
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008786711603273953,
+      "loss": 0.0874,
+      "theoretical_loss": 4.008904199763615,
+      "tokens_seen": 429391872
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008785909163858129,
+      "loss": 0.0856,
+      "theoretical_loss": 4.008635665241635,
+      "tokens_seen": 429654016
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008785106724442304,
+      "loss": 0.0889,
+      "theoretical_loss": 4.008367340353602,
+      "tokens_seen": 429916160
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008784304285026481,
+      "loss": 0.0869,
+      "theoretical_loss": 4.008099224808188,
+      "tokens_seen": 430178304
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008783501845610656,
+      "loss": 0.0842,
+      "theoretical_loss": 4.007831318314645,
+      "tokens_seen": 430440448
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008782699406194833,
+      "loss": 0.0848,
+      "theoretical_loss": 4.00756362058281,
+      "tokens_seen": 430702592
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008781896966779008,
+      "loss": 0.0841,
+      "theoretical_loss": 4.007296131323094,
+      "tokens_seen": 430964736
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008781094527363185,
+      "loss": 0.087,
+      "theoretical_loss": 4.007028850246487,
+      "tokens_seen": 431226880
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008780292087947361,
+      "loss": 0.0865,
+      "theoretical_loss": 4.006761777064557,
+      "tokens_seen": 431489024
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008779489648531536,
+      "loss": 0.0862,
+      "theoretical_loss": 4.006494911489444,
+      "tokens_seen": 431751168
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008778687209115712,
+      "loss": 0.0868,
+      "theoretical_loss": 4.006228253233864,
+      "tokens_seen": 432013312
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008777884769699888,
+      "loss": 0.0884,
+      "theoretical_loss": 4.0059618020111,
+      "tokens_seen": 432275456
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": 0.0010576657950878143,
+      "objective/train/docs_used": 163281,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.687941551208496,
+      "objective/train/original_loss": 1.687941312789917,
+      "objective/train/theoretical_loss": 4.00569555753501,
+      "objective/train/tokens_used": 452997600,
+      "objective/train/value_avg": -0.006397247314453125,
+      "objective/train/value_loss": 0.0002257569576613605,
+      "objective/train/value_max": -0.0002758502960205078,
+      "objective/train/value_min": -0.3525390625,
+      "objective/train/value_reward_corr": 0.5840597410895295,
+      "objective/train/value_std": 0.01157379150390625,
+      "objective/train/weight_avg": 1.0011610984802246,
+      "objective/train/weighted_lm_loss": 1.6893194913864136,
+      "objective/train/weights_max": 1.2730658054351807,
+      "objective/train/weights_min": 0.3710364103317261,
+      "theoretical_loss": 4.00569555753501,
+      "tokens_seen": 432537600
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008777082330284064,
+      "loss": 0.0873,
+      "theoretical_loss": 4.00569555753501,
+      "tokens_seen": 432537600
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008776279890868239,
+      "loss": 0.0862,
+      "theoretical_loss": 4.0054295195200185,
+      "tokens_seen": 432799744
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008775477451452416,
+      "loss": 0.0855,
+      "theoretical_loss": 4.005163687681116,
+      "tokens_seen": 433061888
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008774675012036591,
+      "loss": 0.0855,
+      "theoretical_loss": 4.00489806173386,
+      "tokens_seen": 433324032
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008773872572620766,
+      "loss": 0.0895,
+      "theoretical_loss": 4.004632641394372,
+      "tokens_seen": 433586176
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008773070133204944,
+      "loss": 0.0874,
+      "theoretical_loss": 4.0043674263793365,
+      "tokens_seen": 433848320
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008772267693789119,
+      "loss": 0.0848,
+      "theoretical_loss": 4.004102416405998,
+      "tokens_seen": 434110464
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008771465254373296,
+      "loss": 0.0857,
+      "theoretical_loss": 4.0038376111921625,
+      "tokens_seen": 434372608
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008770662814957471,
+      "loss": 0.0877,
+      "theoretical_loss": 4.0035730104561935,
+      "tokens_seen": 434634752
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008769860375541647,
+      "loss": 0.0893,
+      "theoretical_loss": 4.003308613917012,
+      "tokens_seen": 434896896
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008769057936125823,
+      "loss": 0.0879,
+      "theoretical_loss": 4.003044421294094,
+      "tokens_seen": 435159040
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008768255496709998,
+      "loss": 0.0856,
+      "theoretical_loss": 4.002780432307468,
+      "tokens_seen": 435421184
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008767453057294174,
+      "loss": 0.0858,
+      "theoretical_loss": 4.0025166466777184,
+      "tokens_seen": 435683328
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": 0.0019057170720770955,
+      "objective/train/docs_used": 164592,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.62677001953125,
+      "objective/train/original_loss": 1.6267703771591187,
+      "objective/train/theoretical_loss": 4.002384830034506,
+      "objective/train/tokens_used": 456274400,
+      "objective/train/value_avg": -0.00865936279296875,
+      "objective/train/value_loss": 0.00031009086524136364,
+      "objective/train/value_max": -0.00029587745666503906,
+      "objective/train/value_min": -0.68408203125,
+      "objective/train/value_reward_corr": 0.6592497390378055,
+      "objective/train/value_std": 0.01568603515625,
+      "objective/train/weight_avg": 1.0020478963851929,
+      "objective/train/weighted_lm_loss": 1.629705548286438,
+      "objective/train/weights_max": 1.3052014112472534,
+      "objective/train/weights_min": 0.3907938003540039,
+      "theoretical_loss": 4.002384830034506,
+      "tokens_seen": 435814400
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.000876665061787835,
+      "loss": 0.0874,
+      "theoretical_loss": 4.00225306412598,
+      "tokens_seen": 435945472
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008765848178462527,
+      "loss": 0.0876,
+      "theoretical_loss": 4.001989684373934,
+      "tokens_seen": 436207616
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008765045739046702,
+      "loss": 0.0865,
+      "theoretical_loss": 4.0017265071438155,
+      "tokens_seen": 436469760
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008764243299630879,
+      "loss": 0.0866,
+      "theoretical_loss": 4.001463532158402,
+      "tokens_seen": 436731904
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008763440860215054,
+      "loss": 0.0868,
+      "theoretical_loss": 4.001200759141019,
+      "tokens_seen": 436994048
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008762638420799229,
+      "loss": 0.0861,
+      "theoretical_loss": 4.000938187815535,
+      "tokens_seen": 437256192
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008761835981383406,
+      "loss": 0.0855,
+      "theoretical_loss": 4.000675817906362,
+      "tokens_seen": 437518336
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008761033541967581,
+      "loss": 0.0857,
+      "theoretical_loss": 4.000413649138453,
+      "tokens_seen": 437780480
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008760231102551757,
+      "loss": 0.0873,
+      "theoretical_loss": 4.000151681237301,
+      "tokens_seen": 438042624
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008759428663135933,
+      "loss": 0.0856,
+      "theoretical_loss": 3.9998899139289392,
+      "tokens_seen": 438304768
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.000875862622372011,
+      "loss": 0.0858,
+      "theoretical_loss": 3.999628346939934,
+      "tokens_seen": 438566912
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008757823784304286,
+      "loss": 0.0886,
+      "theoretical_loss": 3.9993669799973928,
+      "tokens_seen": 438829056
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": 0.0023695742711424828,
+      "objective/train/docs_used": 165818,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.712162733078003,
+      "objective/train/original_loss": 1.7121628522872925,
+      "objective/train/theoretical_loss": 3.9991058128289536,
+      "objective/train/tokens_used": 459551200,
+      "objective/train/value_avg": -0.0120391845703125,
+      "objective/train/value_loss": 0.0006223080563358963,
+      "objective/train/value_max": -0.0002002716064453125,
+      "objective/train/value_min": -0.81982421875,
+      "objective/train/value_reward_corr": 0.7511700502995355,
+      "objective/train/value_std": 0.0264739990234375,
+      "objective/train/weight_avg": 1.0026506185531616,
+      "objective/train/weighted_lm_loss": 1.716010570526123,
+      "objective/train/weights_max": 1.3089244365692139,
+      "objective/train/weights_min": 0.41716745495796204,
+      "theoretical_loss": 3.9991058128289536,
+      "tokens_seen": 439091200
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008757021344888461,
+      "loss": 0.0856,
+      "theoretical_loss": 3.9991058128289536,
+      "tokens_seen": 439091200
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008756218905472637,
+      "loss": 0.0895,
+      "theoretical_loss": 3.998844845162789,
+      "tokens_seen": 439353344
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008755416466056813,
+      "loss": 0.0873,
+      "theoretical_loss": 3.998584076727604,
+      "tokens_seen": 439615488
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008754614026640989,
+      "loss": 0.0886,
+      "theoretical_loss": 3.998323507252633,
+      "tokens_seen": 439877632
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008753811587225164,
+      "loss": 0.0853,
+      "theoretical_loss": 3.998063136467639,
+      "tokens_seen": 440139776
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008753009147809341,
+      "loss": 0.0881,
+      "theoretical_loss": 3.9978029641029154,
+      "tokens_seen": 440401920
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008752206708393516,
+      "loss": 0.0861,
+      "theoretical_loss": 3.9975429898892783,
+      "tokens_seen": 440664064
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008751404268977691,
+      "loss": 0.0871,
+      "theoretical_loss": 3.9972832135580707,
+      "tokens_seen": 440926208
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008750601829561869,
+      "loss": 0.087,
+      "theoretical_loss": 3.9970236348411605,
+      "tokens_seen": 441188352
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008749799390146044,
+      "loss": 0.0868,
+      "theoretical_loss": 3.996764253470935,
+      "tokens_seen": 441450496
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.000874899695073022,
+      "loss": 0.0825,
+      "theoretical_loss": 3.996505069180306,
+      "tokens_seen": 441712640
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008748194511314396,
+      "loss": 0.0872,
+      "theoretical_loss": 3.9962460817027017,
+      "tokens_seen": 441974784
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008747392071898572,
+      "loss": 0.0865,
+      "theoretical_loss": 3.995987290772071,
+      "tokens_seen": 442236928
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": 0.0025604458060115576,
+      "objective/train/docs_used": 166951,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6768556833267212,
+      "objective/train/original_loss": 1.6768556833267212,
+      "objective/train/theoretical_loss": 3.9958579689288705,
+      "objective/train/tokens_used": 462828000,
+      "objective/train/value_avg": -0.0081634521484375,
+      "objective/train/value_loss": 0.00017307909729424864,
+      "objective/train/value_max": -0.0002472400665283203,
+      "objective/train/value_min": -0.494140625,
+      "objective/train/value_reward_corr": 0.5969880927520907,
+      "objective/train/value_std": 0.0106353759765625,
+      "objective/train/weight_avg": 1.0026414394378662,
+      "objective/train/weighted_lm_loss": 1.6814841032028198,
+      "objective/train/weights_max": 1.2068192958831787,
+      "objective/train/weights_min": 0.37579405307769775,
+      "theoretical_loss": 3.9958579689288705,
+      "tokens_seen": 442368000
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008746589632482748,
+      "loss": 0.0859,
+      "theoretical_loss": 3.9957286961228786,
+      "tokens_seen": 442499072
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008745787193066924,
+      "loss": 0.0838,
+      "theoretical_loss": 3.995470297490106,
+      "tokens_seen": 442761216
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008744984753651099,
+      "loss": 0.0841,
+      "theoretical_loss": 3.995212094609249,
+      "tokens_seen": 443023360
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008744182314235275,
+      "loss": 0.0857,
+      "theoretical_loss": 3.994954087216315,
+      "tokens_seen": 443285504
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008743379874819452,
+      "loss": 0.0886,
+      "theoretical_loss": 3.994696275047825,
+      "tokens_seen": 443547648
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008742577435403627,
+      "loss": 0.0862,
+      "theoretical_loss": 3.9944386578408113,
+      "tokens_seen": 443809792
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008741774995987804,
+      "loss": 0.0845,
+      "theoretical_loss": 3.9941812353328126,
+      "tokens_seen": 444071936
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008740972556571979,
+      "loss": 0.0852,
+      "theoretical_loss": 3.993924007261878,
+      "tokens_seen": 444334080
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008740170117156155,
+      "loss": 0.0881,
+      "theoretical_loss": 3.9936669733665617,
+      "tokens_seen": 444596224
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008739367677740331,
+      "loss": 0.0859,
+      "theoretical_loss": 3.9934101333859253,
+      "tokens_seen": 444858368
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008738565238324506,
+      "loss": 0.0882,
+      "theoretical_loss": 3.9931534870595327,
+      "tokens_seen": 445120512
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 0.0008737762798908682,
+      "loss": 0.0857,
+      "theoretical_loss": 3.9928970341274517,
+      "tokens_seen": 445382656
+    },
+    {
+      "epoch": 0.13,
+      "objective/train/advantage_avg": -4.0266091673402116e-05,
+      "objective/train/docs_used": 168097,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8889011144638062,
+      "objective/train/original_loss": 1.8889009952545166,
+      "objective/train/theoretical_loss": 3.992640774330251,
+      "objective/train/tokens_used": 466104800,
+      "objective/train/value_avg": -0.007015228271484375,
+      "objective/train/value_loss": 0.00026242577587254345,
+      "objective/train/value_max": -0.0002415180206298828,
+      "objective/train/value_min": -0.1566162109375,
+      "objective/train/value_reward_corr": 0.5724330601050478,
+      "objective/train/value_std": 0.0086822509765625,
+      "objective/train/weight_avg": 1.0000807046890259,
+      "objective/train/weighted_lm_loss": 1.889707326889038,
+      "objective/train/weights_max": 1.112181305885315,
+      "objective/train/weights_min": 0.38205578923225403,
+      "theoretical_loss": 3.992640774330251,
+      "tokens_seen": 445644800
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008736960359492858,
+      "loss": 0.0861,
+      "theoretical_loss": 3.992640774330251,
+      "tokens_seen": 445644800
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008736157920077035,
+      "loss": 0.084,
+      "theoretical_loss": 3.9923847074090015,
+      "tokens_seen": 445906944
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000873535548066121,
+      "loss": 0.0835,
+      "theoretical_loss": 3.9921288331052702,
+      "tokens_seen": 446169088
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008734553041245387,
+      "loss": 0.0856,
+      "theoretical_loss": 3.991873151161124,
+      "tokens_seen": 446431232
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008733750601829562,
+      "loss": 0.0886,
+      "theoretical_loss": 3.9916176613191263,
+      "tokens_seen": 446693376
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008732948162413738,
+      "loss": 0.0862,
+      "theoretical_loss": 3.9913623633223354,
+      "tokens_seen": 446955520
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008732145722997914,
+      "loss": 0.0831,
+      "theoretical_loss": 3.9911072569143036,
+      "tokens_seen": 447217664
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008731343283582089,
+      "loss": 0.0842,
+      "theoretical_loss": 3.9908523418390764,
+      "tokens_seen": 447479808
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008730540844166266,
+      "loss": 0.0841,
+      "theoretical_loss": 3.990597617841191,
+      "tokens_seen": 447741952
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008729738404750441,
+      "loss": 0.0841,
+      "theoretical_loss": 3.9903430846656742,
+      "tokens_seen": 448004096
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008728935965334618,
+      "loss": 0.0871,
+      "theoretical_loss": 3.990088742058043,
+      "tokens_seen": 448266240
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008728133525918794,
+      "loss": 0.0835,
+      "theoretical_loss": 3.9898345897643024,
+      "tokens_seen": 448528384
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008727331086502969,
+      "loss": 0.0835,
+      "theoretical_loss": 3.989580627530943,
+      "tokens_seen": 448790528
+    },
+    {
+      "epoch": 0.14,
+      "objective/train/advantage_avg": 0.0008596550906077027,
+      "objective/train/docs_used": 169309,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.786888599395752,
+      "objective/train/original_loss": 1.786888599395752,
+      "objective/train/theoretical_loss": 3.9894537176078178,
+      "objective/train/tokens_used": 469381600,
+      "objective/train/value_avg": -0.008880615234375,
+      "objective/train/value_loss": 0.00042329219286330044,
+      "objective/train/value_max": -0.0002359151840209961,
+      "objective/train/value_min": -0.681640625,
+      "objective/train/value_reward_corr": 0.5803828830417395,
+      "objective/train/value_std": 0.014617919921875,
+      "objective/train/weight_avg": 1.0010493993759155,
+      "objective/train/weighted_lm_loss": 1.7893576622009277,
+      "objective/train/weights_max": 1.8888424634933472,
+      "objective/train/weights_min": 0.3687688410282135,
+      "theoretical_loss": 3.9894537176078178,
+      "tokens_seen": 448921600
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008726528647087145,
+      "loss": 0.0836,
+      "theoretical_loss": 3.9893268551049417,
+      "tokens_seen": 449052672
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008725726207671321,
+      "loss": 0.0848,
+      "theoretical_loss": 3.9890732722337594,
+      "tokens_seen": 449314816
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008724923768255497,
+      "loss": 0.0882,
+      "theoretical_loss": 3.988819878665341,
+      "tokens_seen": 449576960
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008724121328839672,
+      "loss": 0.085,
+      "theoretical_loss": 3.988566674148111,
+      "tokens_seen": 449839104
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008723318889423849,
+      "loss": 0.0854,
+      "theoretical_loss": 3.988313658430978,
+      "tokens_seen": 450101248
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008722516450008024,
+      "loss": 0.0872,
+      "theoretical_loss": 3.9880608312633274,
+      "tokens_seen": 450363392
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008721714010592199,
+      "loss": 0.0875,
+      "theoretical_loss": 3.9878081923950237,
+      "tokens_seen": 450625536
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008720911571176377,
+      "loss": 0.0854,
+      "theoretical_loss": 3.9875557415764087,
+      "tokens_seen": 450887680
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008720109131760552,
+      "loss": 0.0834,
+      "theoretical_loss": 3.9873034785582995,
+      "tokens_seen": 451149824
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008719306692344729,
+      "loss": 0.0829,
+      "theoretical_loss": 3.9870514030919884,
+      "tokens_seen": 451411968
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008718504252928904,
+      "loss": 0.0848,
+      "theoretical_loss": 3.986799514929242,
+      "tokens_seen": 451674112
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000871770181351308,
+      "loss": 0.0876,
+      "theoretical_loss": 3.9865478138222974,
+      "tokens_seen": 451936256
+    },
+    {
+      "epoch": 0.14,
+      "objective/train/advantage_avg": 0.0022668084129691124,
+      "objective/train/docs_used": 170598,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.797487735748291,
+      "objective/train/original_loss": 1.797487735748291,
+      "objective/train/theoretical_loss": 3.9862962995238647,
+      "objective/train/tokens_used": 472658400,
+      "objective/train/value_avg": -0.0083465576171875,
+      "objective/train/value_loss": 0.0004899487248621881,
+      "objective/train/value_max": -0.0002779960632324219,
+      "objective/train/value_min": -0.68115234375,
+      "objective/train/value_reward_corr": 0.532976155116499,
+      "objective/train/value_std": 0.0135040283203125,
+      "objective/train/weight_avg": 1.0024774074554443,
+      "objective/train/weighted_lm_loss": 1.8017926216125488,
+      "objective/train/weights_max": 1.5135427713394165,
+      "objective/train/weights_min": 0.3686288893222809,
+      "theoretical_loss": 3.9862962995238647,
+      "tokens_seen": 452198400
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008716899374097256,
+      "loss": 0.086,
+      "theoretical_loss": 3.9862962995238647,
+      "tokens_seen": 452198400
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008716096934681432,
+      "loss": 0.0884,
+      "theoretical_loss": 3.9860449717871234,
+      "tokens_seen": 452460544
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008715294495265607,
+      "loss": 0.0851,
+      "theoretical_loss": 3.9857938303657217,
+      "tokens_seen": 452722688
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008714492055849783,
+      "loss": 0.0885,
+      "theoretical_loss": 3.9855428750137754,
+      "tokens_seen": 452984832
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000871368961643396,
+      "loss": 0.0889,
+      "theoretical_loss": 3.9852921054858665,
+      "tokens_seen": 453246976
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008712887177018135,
+      "loss": 0.0843,
+      "theoretical_loss": 3.9850415215370445,
+      "tokens_seen": 453509120
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008712084737602312,
+      "loss": 0.0819,
+      "theoretical_loss": 3.984791122922821,
+      "tokens_seen": 453771264
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008711282298186487,
+      "loss": 0.0859,
+      "theoretical_loss": 3.98454090939917,
+      "tokens_seen": 454033408
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008710479858770663,
+      "loss": 0.0861,
+      "theoretical_loss": 3.984290880722531,
+      "tokens_seen": 454295552
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008709677419354839,
+      "loss": 0.0834,
+      "theoretical_loss": 3.9840410366498,
+      "tokens_seen": 454557696
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008708874979939014,
+      "loss": 0.0869,
+      "theoretical_loss": 3.983791376938336,
+      "tokens_seen": 454819840
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008708072540523191,
+      "loss": 0.083,
+      "theoretical_loss": 3.983541901345955,
+      "tokens_seen": 455081984
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008707270101107366,
+      "loss": 0.0871,
+      "theoretical_loss": 3.983292609630931,
+      "tokens_seen": 455344128
+    },
+    {
+      "epoch": 0.14,
+      "objective/train/advantage_avg": 6.911411037435755e-05,
+      "objective/train/docs_used": 171735,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7402760982513428,
+      "objective/train/original_loss": 1.7402762174606323,
+      "objective/train/theoretical_loss": 3.983168032652013,
+      "objective/train/tokens_used": 475935200,
+      "objective/train/value_avg": -0.01013946533203125,
+      "objective/train/value_loss": 0.0009821663843467832,
+      "objective/train/value_max": -0.00021147727966308594,
+      "objective/train/value_min": -0.916015625,
+      "objective/train/value_reward_corr": 0.7078605683518372,
+      "objective/train/value_std": 0.0244293212890625,
+      "objective/train/weight_avg": 1.000481367111206,
+      "objective/train/weighted_lm_loss": 1.7407102584838867,
+      "objective/train/weights_max": 2.12113881111145,
+      "objective/train/weights_min": 0.36906227469444275,
+      "theoretical_loss": 3.983168032652013,
+      "tokens_seen": 455475200
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008706467661691543,
+      "loss": 0.0875,
+      "theoretical_loss": 3.9830435015519936,
+      "tokens_seen": 455606272
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008705665222275719,
+      "loss": 0.086,
+      "theoretical_loss": 3.982794576868328,
+      "tokens_seen": 455868416
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008704862782859895,
+      "loss": 0.0847,
+      "theoretical_loss": 3.982545835339573,
+      "tokens_seen": 456130560
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000870406034344407,
+      "loss": 0.0849,
+      "theoretical_loss": 3.982297276725822,
+      "tokens_seen": 456392704
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008703257904028246,
+      "loss": 0.0838,
+      "theoretical_loss": 3.9820489007876176,
+      "tokens_seen": 456654848
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008702455464612422,
+      "loss": 0.0813,
+      "theoretical_loss": 3.981800707285955,
+      "tokens_seen": 456916992
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008701653025196597,
+      "loss": 0.0853,
+      "theoretical_loss": 3.9815526959822787,
+      "tokens_seen": 457179136
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008700850585780774,
+      "loss": 0.0835,
+      "theoretical_loss": 3.981304866638481,
+      "tokens_seen": 457441280
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008700048146364949,
+      "loss": 0.0874,
+      "theoretical_loss": 3.9810572190169027,
+      "tokens_seen": 457703424
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008699245706949126,
+      "loss": 0.0864,
+      "theoretical_loss": 3.9808097528803295,
+      "tokens_seen": 457965568
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008698443267533302,
+      "loss": 0.0861,
+      "theoretical_loss": 3.9805624679919935,
+      "tokens_seen": 458227712
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008697640828117477,
+      "loss": 0.0878,
+      "theoretical_loss": 3.98031536411557,
+      "tokens_seen": 458489856
+    },
+    {
+      "epoch": 0.14,
+      "objective/train/advantage_avg": 0.0018690774450078607,
+      "objective/train/docs_used": 173009,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6653363704681396,
+      "objective/train/original_loss": 1.6653363704681396,
+      "objective/train/theoretical_loss": 3.9800684410151783,
+      "objective/train/tokens_used": 479212000,
+      "objective/train/value_avg": -0.007537841796875,
+      "objective/train/value_loss": 0.00013798951113130897,
+      "objective/train/value_max": -0.0002846717834472656,
+      "objective/train/value_min": -0.277099609375,
+      "objective/train/value_reward_corr": 0.48532692544123834,
+      "objective/train/value_std": 0.00695037841796875,
+      "objective/train/weight_avg": 1.0019335746765137,
+      "objective/train/weighted_lm_loss": 1.6689884662628174,
+      "objective/train/weights_max": 1.0873464345932007,
+      "objective/train/weights_min": 0.3710477352142334,
+      "theoretical_loss": 3.9800684410151783,
+      "tokens_seen": 458752000
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008696838388701653,
+      "loss": 0.0846,
+      "theoretical_loss": 3.9800684410151783,
+      "tokens_seen": 458752000
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008696035949285829,
+      "loss": 0.0826,
+      "theoretical_loss": 3.979821698455379,
+      "tokens_seen": 459014144
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008695233509870005,
+      "loss": 0.0861,
+      "theoretical_loss": 3.9795751362011735,
+      "tokens_seen": 459276288
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008694431070454181,
+      "loss": 0.0848,
+      "theoretical_loss": 3.979328754018004,
+      "tokens_seen": 459538432
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008693628631038357,
+      "loss": 0.085,
+      "theoretical_loss": 3.979082551671749,
+      "tokens_seen": 459800576
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008692826191622532,
+      "loss": 0.0876,
+      "theoretical_loss": 3.9788365289287286,
+      "tokens_seen": 460062720
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008692023752206708,
+      "loss": 0.0879,
+      "theoretical_loss": 3.9785906855556945,
+      "tokens_seen": 460324864
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008691221312790885,
+      "loss": 0.0846,
+      "theoretical_loss": 3.9783450213198384,
+      "tokens_seen": 460587008
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000869041887337506,
+      "loss": 0.0837,
+      "theoretical_loss": 3.9780995359887843,
+      "tokens_seen": 460849152
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008689616433959237,
+      "loss": 0.0868,
+      "theoretical_loss": 3.9778542293305894,
+      "tokens_seen": 461111296
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008688813994543412,
+      "loss": 0.0865,
+      "theoretical_loss": 3.977609101113744,
+      "tokens_seen": 461373440
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008688011555127588,
+      "loss": 0.0809,
+      "theoretical_loss": 3.97736415110717,
+      "tokens_seen": 461635584
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008687209115711764,
+      "loss": 0.0903,
+      "theoretical_loss": 3.977119379080218,
+      "tokens_seen": 461897728
+    },
+    {
+      "epoch": 0.14,
+      "objective/train/advantage_avg": 0.0003697322681546211,
+      "objective/train/docs_used": 174193,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8311926126480103,
+      "objective/train/original_loss": 1.8311924934387207,
+      "objective/train/theoretical_loss": 3.9769970597371405,
+      "objective/train/tokens_used": 482488800,
+      "objective/train/value_avg": -0.0074005126953125,
+      "objective/train/value_loss": 0.00031192455207929015,
+      "objective/train/value_max": -0.00023233890533447266,
+      "objective/train/value_min": -0.40869140625,
+      "objective/train/value_reward_corr": 0.5429328033135243,
+      "objective/train/value_std": 0.011566162109375,
+      "objective/train/weight_avg": 1.0005161762237549,
+      "objective/train/weighted_lm_loss": 1.8320584297180176,
+      "objective/train/weights_max": 1.3934298753738403,
+      "objective/train/weights_min": 0.6126090288162231,
+      "theoretical_loss": 3.9769970597371405,
+      "tokens_seen": 462028800
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000868640667629594,
+      "loss": 0.0886,
+      "theoretical_loss": 3.9768747848026695,
+      "tokens_seen": 462159872
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008685604236880115,
+      "loss": 0.0836,
+      "theoretical_loss": 3.9766303680447335,
+      "tokens_seen": 462422016
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008684801797464291,
+      "loss": 0.0834,
+      "theoretical_loss": 3.9763861285770457,
+      "tokens_seen": 462684160
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008683999358048468,
+      "loss": 0.0862,
+      "theoretical_loss": 3.9761420661706683,
+      "tokens_seen": 462946304
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008683196918632643,
+      "loss": 0.084,
+      "theoretical_loss": 3.975898180597089,
+      "tokens_seen": 463208448
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000868239447921682,
+      "loss": 0.0857,
+      "theoretical_loss": 3.9756544716282187,
+      "tokens_seen": 463470592
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008681592039800995,
+      "loss": 0.0837,
+      "theoretical_loss": 3.975410939036392,
+      "tokens_seen": 463732736
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008680789600385172,
+      "loss": 0.0852,
+      "theoretical_loss": 3.9751675825943638,
+      "tokens_seen": 463994880
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008679987160969347,
+      "loss": 0.0895,
+      "theoretical_loss": 3.974924402075313,
+      "tokens_seen": 464257024
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008679184721553522,
+      "loss": 0.0863,
+      "theoretical_loss": 3.9746813972528354,
+      "tokens_seen": 464519168
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008678382282137699,
+      "loss": 0.0875,
+      "theoretical_loss": 3.9744385679009486,
+      "tokens_seen": 464781312
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008677579842721874,
+      "loss": 0.0833,
+      "theoretical_loss": 3.9741959137940848,
+      "tokens_seen": 465043456
+    },
+    {
+      "epoch": 0.14,
+      "objective/train/advantage_avg": 0.002355672651901841,
+      "objective/train/docs_used": 175471,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4537473917007446,
+      "objective/train/original_loss": 1.4537473917007446,
+      "objective/train/theoretical_loss": 3.973953434707096,
+      "objective/train/tokens_used": 485765600,
+      "objective/train/value_avg": -0.0084075927734375,
+      "objective/train/value_loss": 0.00017681960889603943,
+      "objective/train/value_max": -0.00033664703369140625,
+      "objective/train/value_min": -0.244384765625,
+      "objective/train/value_reward_corr": 0.6067470079442566,
+      "objective/train/value_std": 0.011962890625,
+      "objective/train/weight_avg": 1.0024397373199463,
+      "objective/train/weighted_lm_loss": 1.4569942951202393,
+      "objective/train/weights_max": 1.135114312171936,
+      "objective/train/weights_min": 0.36958563327789307,
+      "theoretical_loss": 3.973953434707096,
+      "tokens_seen": 465305600
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008676777403306051,
+      "loss": 0.0827,
+      "theoretical_loss": 3.973953434707096,
+      "tokens_seen": 465305600
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008675974963890227,
+      "loss": 0.084,
+      "theoretical_loss": 3.973711130415248,
+      "tokens_seen": 465567744
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008675172524474403,
+      "loss": 0.0855,
+      "theoretical_loss": 3.973469000694223,
+      "tokens_seen": 465829888
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008674370085058578,
+      "loss": 0.084,
+      "theoretical_loss": 3.973227045320117,
+      "tokens_seen": 466092032
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008673567645642754,
+      "loss": 0.0852,
+      "theoretical_loss": 3.9729852640694383,
+      "tokens_seen": 466354176
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000867276520622693,
+      "loss": 0.0855,
+      "theoretical_loss": 3.972743656719107,
+      "tokens_seen": 466616320
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008671962766811105,
+      "loss": 0.0819,
+      "theoretical_loss": 3.9725022230464537,
+      "tokens_seen": 466878464
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008671160327395282,
+      "loss": 0.0849,
+      "theoretical_loss": 3.9722609628292207,
+      "tokens_seen": 467140608
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008670357887979457,
+      "loss": 0.084,
+      "theoretical_loss": 3.9720198758455574,
+      "tokens_seen": 467402752
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008669555448563635,
+      "loss": 0.0854,
+      "theoretical_loss": 3.9717789618740227,
+      "tokens_seen": 467664896
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000866875300914781,
+      "loss": 0.0864,
+      "theoretical_loss": 3.9715382206935814,
+      "tokens_seen": 467927040
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008667950569731985,
+      "loss": 0.082,
+      "theoretical_loss": 3.9712976520836043,
+      "tokens_seen": 468189184
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008667148130316162,
+      "loss": 0.0839,
+      "theoretical_loss": 3.971057255823868,
+      "tokens_seen": 468451328
+    },
+    {
+      "epoch": 0.14,
+      "objective/train/advantage_avg": 0.0019131185254082084,
+      "objective/train/docs_used": 176758,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8133913278579712,
+      "objective/train/original_loss": 1.8133914470672607,
+      "objective/train/theoretical_loss": 3.9709371222566308,
+      "objective/train/tokens_used": 489042400,
+      "objective/train/value_avg": -0.00600433349609375,
+      "objective/train/value_loss": 0.00013036759628448635,
+      "objective/train/value_max": -0.00028252601623535156,
+      "objective/train/value_min": -0.2418212890625,
+      "objective/train/value_reward_corr": 0.4864233011946609,
+      "objective/train/value_std": 0.006320953369140625,
+      "objective/train/weight_avg": 1.0019735097885132,
+      "objective/train/weighted_lm_loss": 1.817116618156433,
+      "objective/train/weights_max": 1.2735666036605835,
+      "objective/train/weights_min": 0.3718809485435486,
+      "theoretical_loss": 3.9709371222566308,
+      "tokens_seen": 468582400
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008666345690900337,
+      "loss": 0.0859,
+      "theoretical_loss": 3.9708170316945526,
+      "tokens_seen": 468713472
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008665543251484513,
+      "loss": 0.0853,
+      "theoretical_loss": 3.9705769794762418,
+      "tokens_seen": 468975616
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008664740812068689,
+      "loss": 0.0866,
+      "theoretical_loss": 3.97033709894992,
+      "tokens_seen": 469237760
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008663938372652865,
+      "loss": 0.085,
+      "theoretical_loss": 3.970097389896975,
+      "tokens_seen": 469499904
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000866313593323704,
+      "loss": 0.0853,
+      "theoretical_loss": 3.9698578520991936,
+      "tokens_seen": 469762048
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008662333493821216,
+      "loss": 0.0873,
+      "theoretical_loss": 3.9696184853387617,
+      "tokens_seen": 470024192
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008661531054405393,
+      "loss": 0.0834,
+      "theoretical_loss": 3.9693792893982636,
+      "tokens_seen": 470286336
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008660728614989568,
+      "loss": 0.0861,
+      "theoretical_loss": 3.969140264060681,
+      "tokens_seen": 470548480
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008659926175573745,
+      "loss": 0.0844,
+      "theoretical_loss": 3.9689014091093933,
+      "tokens_seen": 470810624
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000865912373615792,
+      "loss": 0.0843,
+      "theoretical_loss": 3.9686627243281727,
+      "tokens_seen": 471072768
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008658321296742096,
+      "loss": 0.088,
+      "theoretical_loss": 3.9684242095011886,
+      "tokens_seen": 471334912
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008657518857326272,
+      "loss": 0.0841,
+      "theoretical_loss": 3.9681858644130017,
+      "tokens_seen": 471597056
+    },
+    {
+      "epoch": 0.14,
+      "objective/train/advantage_avg": 0.0015276180347427726,
+      "objective/train/docs_used": 177948,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7598378658294678,
+      "objective/train/original_loss": 1.7598377466201782,
+      "objective/train/theoretical_loss": 3.967947688848568,
+      "objective/train/tokens_used": 492319200,
+      "objective/train/value_avg": -0.01081085205078125,
+      "objective/train/value_loss": 0.00034501656773500144,
+      "objective/train/value_max": -0.00039505958557128906,
+      "objective/train/value_min": -0.76025390625,
+      "objective/train/value_reward_corr": 0.7302345418778485,
+      "objective/train/value_std": 0.0194549560546875,
+      "objective/train/weight_avg": 1.0016895532608032,
+      "objective/train/weighted_lm_loss": 1.7620900869369507,
+      "objective/train/weights_max": 1.419206142425537,
+      "objective/train/weights_min": 0.39239516854286194,
+      "theoretical_loss": 3.967947688848568,
+      "tokens_seen": 471859200
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008656716417910447,
+      "loss": 0.0869,
+      "theoretical_loss": 3.967947688848568,
+      "tokens_seen": 471859200
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008655913978494624,
+      "loss": 0.0807,
+      "theoretical_loss": 3.9677096825932328,
+      "tokens_seen": 472121344
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008655111539078799,
+      "loss": 0.0841,
+      "theoretical_loss": 3.9674718454327325,
+      "tokens_seen": 472383488
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008654309099662976,
+      "loss": 0.0864,
+      "theoretical_loss": 3.9672341771531956,
+      "tokens_seen": 472645632
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008653506660247152,
+      "loss": 0.0862,
+      "theoretical_loss": 3.9669966775411365,
+      "tokens_seen": 472907776
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008652704220831328,
+      "loss": 0.0841,
+      "theoretical_loss": 3.96675934638346,
+      "tokens_seen": 473169920
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008651901781415503,
+      "loss": 0.0864,
+      "theoretical_loss": 3.9665221834674558,
+      "tokens_seen": 473432064
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000865109934199968,
+      "loss": 0.0841,
+      "theoretical_loss": 3.9662851885808026,
+      "tokens_seen": 473694208
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008650296902583855,
+      "loss": 0.0848,
+      "theoretical_loss": 3.966048361511562,
+      "tokens_seen": 473956352
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000864949446316803,
+      "loss": 0.0829,
+      "theoretical_loss": 3.96581170204818,
+      "tokens_seen": 474218496
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008648692023752207,
+      "loss": 0.087,
+      "theoretical_loss": 3.9655752099794874,
+      "tokens_seen": 474480640
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008647889584336382,
+      "loss": 0.0888,
+      "theoretical_loss": 3.9653388850946976,
+      "tokens_seen": 474742784
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008647087144920559,
+      "loss": 0.084,
+      "theoretical_loss": 3.965102727183404,
+      "tokens_seen": 475004928
+    },
+    {
+      "epoch": 0.14,
+      "objective/train/advantage_avg": 0.0018348618177697062,
+      "objective/train/docs_used": 179087,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5616334676742554,
+      "objective/train/original_loss": 1.5616333484649658,
+      "objective/train/theoretical_loss": 3.9649847107771707,
+      "objective/train/tokens_used": 495596000,
+      "objective/train/value_avg": -0.0064697265625,
+      "objective/train/value_loss": 0.00020544622384477407,
+      "objective/train/value_max": -0.0003077983856201172,
+      "objective/train/value_min": -0.6181640625,
+      "objective/train/value_reward_corr": 0.5330140721340589,
+      "objective/train/value_std": 0.00867462158203125,
+      "objective/train/weight_avg": 1.0019270181655884,
+      "objective/train/weighted_lm_loss": 1.564939260482788,
+      "objective/train/weights_max": 1.3273746967315674,
+      "objective/train/weights_min": 0.36936861276626587,
+      "theoretical_loss": 3.9649847107771707,
+      "tokens_seen": 475136000
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008646284705504735,
+      "loss": 0.0832,
+      "theoretical_loss": 3.9648667360355816,
+      "tokens_seen": 475267072
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008645482266088911,
+      "loss": 0.0869,
+      "theoretical_loss": 3.9646309114415863,
+      "tokens_seen": 475529216
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008644679826673087,
+      "loss": 0.0838,
+      "theoretical_loss": 3.9643952531921505,
+      "tokens_seen": 475791360
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008643877387257262,
+      "loss": 0.0817,
+      "theoretical_loss": 3.9641597610783874,
+      "tokens_seen": 476053504
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008643074947841438,
+      "loss": 0.0847,
+      "theoretical_loss": 3.9639244348917853,
+      "tokens_seen": 476315648
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008642272508425614,
+      "loss": 0.0826,
+      "theoretical_loss": 3.96368927442421,
+      "tokens_seen": 476577792
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000864147006900979,
+      "loss": 0.0855,
+      "theoretical_loss": 3.9634542794679013,
+      "tokens_seen": 476839936
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008640667629593965,
+      "loss": 0.0832,
+      "theoretical_loss": 3.963219449815475,
+      "tokens_seen": 477102080
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008639865190178143,
+      "loss": 0.0809,
+      "theoretical_loss": 3.9629847852599207,
+      "tokens_seen": 477364224
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008639062750762318,
+      "loss": 0.0871,
+      "theoretical_loss": 3.9627502855945984,
+      "tokens_seen": 477626368
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008638260311346493,
+      "loss": 0.0839,
+      "theoretical_loss": 3.962515950613242,
+      "tokens_seen": 477888512
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.000863745787193067,
+      "loss": 0.0862,
+      "theoretical_loss": 3.962281780109957,
+      "tokens_seen": 478150656
+    },
+    {
+      "epoch": 0.14,
+      "objective/train/advantage_avg": -0.0006487751961685717,
+      "objective/train/docs_used": 180329,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7242850065231323,
+      "objective/train/original_loss": 1.7242848873138428,
+      "objective/train/theoretical_loss": 3.9620477738792164,
+      "objective/train/tokens_used": 498872800,
+      "objective/train/value_avg": -0.005603790283203125,
+      "objective/train/value_loss": 0.00016961862274911255,
+      "objective/train/value_max": -0.00028252601623535156,
+      "objective/train/value_min": -0.17919921875,
+      "objective/train/value_reward_corr": 0.5824427047069172,
+      "objective/train/value_std": 0.0064849853515625,
+      "objective/train/weight_avg": 0.999433696269989,
+      "objective/train/weighted_lm_loss": 1.724251389503479,
+      "objective/train/weights_max": 1.0798728466033936,
+      "objective/train/weights_min": 0.6110895276069641,
+      "theoretical_loss": 3.9620477738792164,
+      "tokens_seen": 478412800
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 0.0008636655432514845,
+      "loss": 0.0843,
+      "theoretical_loss": 3.9620477738792164,
+      "tokens_seen": 478412800
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008635852993099021,
+      "loss": 0.0854,
+      "theoretical_loss": 3.9618139317158647,
+      "tokens_seen": 478674944
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008635050553683197,
+      "loss": 0.0847,
+      "theoretical_loss": 3.961580253415114,
+      "tokens_seen": 478937088
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008634248114267373,
+      "loss": 0.0845,
+      "theoretical_loss": 3.9613467387725434,
+      "tokens_seen": 479199232
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008633445674851548,
+      "loss": 0.0807,
+      "theoretical_loss": 3.9611133875841,
+      "tokens_seen": 479461376
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008632643235435724,
+      "loss": 0.0854,
+      "theoretical_loss": 3.960880199646096,
+      "tokens_seen": 479723520
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008631840796019901,
+      "loss": 0.0858,
+      "theoretical_loss": 3.9606471747552083,
+      "tokens_seen": 479985664
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008631038356604077,
+      "loss": 0.0826,
+      "theoretical_loss": 3.9604143127084774,
+      "tokens_seen": 480247808
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008630235917188253,
+      "loss": 0.0847,
+      "theoretical_loss": 3.960181613303309,
+      "tokens_seen": 480509952
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008629433477772428,
+      "loss": 0.0875,
+      "theoretical_loss": 3.9599490763374687,
+      "tokens_seen": 480772096
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008628631038356605,
+      "loss": 0.0857,
+      "theoretical_loss": 3.9597167016090866,
+      "tokens_seen": 481034240
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000862782859894078,
+      "loss": 0.0827,
+      "theoretical_loss": 3.95948448891665,
+      "tokens_seen": 481296384
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008627026159524955,
+      "loss": 0.0838,
+      "theoretical_loss": 3.959252438059009,
+      "tokens_seen": 481558528
+    },
+    {
+      "epoch": 0.15,
+      "objective/train/advantage_avg": 0.0011554267257452011,
+      "objective/train/docs_used": 181505,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.673024296760559,
+      "objective/train/original_loss": 1.6730244159698486,
+      "objective/train/theoretical_loss": 3.959136473255475,
+      "objective/train/tokens_used": 502149600,
+      "objective/train/value_avg": -0.00830078125,
+      "objective/train/value_loss": 0.0005257127340883017,
+      "objective/train/value_max": -0.0005254745483398438,
+      "objective/train/value_min": -0.478515625,
+      "objective/train/value_reward_corr": 0.5014164299225942,
+      "objective/train/value_std": 0.0123291015625,
+      "objective/train/weight_avg": 1.0013680458068848,
+      "objective/train/weighted_lm_loss": 1.674941062927246,
+      "objective/train/weights_max": 1.4412035942077637,
+      "objective/train/weights_min": 0.1574522852897644,
+      "theoretical_loss": 3.959136473255475,
+      "tokens_seen": 481689600
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008626223720109132,
+      "loss": 0.0841,
+      "theoretical_loss": 3.9590205488353707,
+      "tokens_seen": 481820672
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008625421280693307,
+      "loss": 0.0867,
+      "theoretical_loss": 3.958788821045302,
+      "tokens_seen": 482082816
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008624618841277484,
+      "loss": 0.0848,
+      "theoretical_loss": 3.958557254488727,
+      "tokens_seen": 482344960
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000862381640186166,
+      "loss": 0.084,
+      "theoretical_loss": 3.958325848965925,
+      "tokens_seen": 482607104
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008623013962445836,
+      "loss": 0.0831,
+      "theoretical_loss": 3.958094604277532,
+      "tokens_seen": 482869248
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008622211523030011,
+      "loss": 0.0863,
+      "theoretical_loss": 3.9578635202245387,
+      "tokens_seen": 483131392
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008621409083614188,
+      "loss": 0.0807,
+      "theoretical_loss": 3.95763259660829,
+      "tokens_seen": 483393536
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008620606644198363,
+      "loss": 0.0829,
+      "theoretical_loss": 3.957401833230484,
+      "tokens_seen": 483655680
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008619804204782538,
+      "loss": 0.0833,
+      "theoretical_loss": 3.957171229893171,
+      "tokens_seen": 483917824
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008619001765366715,
+      "loss": 0.0847,
+      "theoretical_loss": 3.956940786398753,
+      "tokens_seen": 484179968
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000861819932595089,
+      "loss": 0.0856,
+      "theoretical_loss": 3.9567105025499827,
+      "tokens_seen": 484442112
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008617396886535068,
+      "loss": 0.0839,
+      "theoretical_loss": 3.9564803781499633,
+      "tokens_seen": 484704256
+    },
+    {
+      "epoch": 0.15,
+      "objective/train/advantage_avg": 0.0023670962546020746,
+      "objective/train/docs_used": 182723,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4888525009155273,
+      "objective/train/original_loss": 1.4888522624969482,
+      "objective/train/theoretical_loss": 3.956250413002146,
+      "objective/train/tokens_used": 505426400,
+      "objective/train/value_avg": -0.01146697998046875,
+      "objective/train/value_loss": 0.0007001186022534966,
+      "objective/train/value_max": -0.0003249645233154297,
+      "objective/train/value_min": -0.8916015625,
+      "objective/train/value_reward_corr": 0.688608831300723,
+      "objective/train/value_std": 0.02484130859375,
+      "objective/train/weight_avg": 1.0026851892471313,
+      "objective/train/weighted_lm_loss": 1.4927496910095215,
+      "objective/train/weights_max": 2.0180838108062744,
+      "objective/train/weights_min": 0.3854755163192749,
+      "theoretical_loss": 3.956250413002146,
+      "tokens_seen": 484966400
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008616594447119243,
+      "loss": 0.0845,
+      "theoretical_loss": 3.956250413002146,
+      "tokens_seen": 484966400
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008615792007703419,
+      "loss": 0.0851,
+      "theoretical_loss": 3.9560206069103314,
+      "tokens_seen": 485228544
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008614989568287595,
+      "loss": 0.086,
+      "theoretical_loss": 3.9557909596786676,
+      "tokens_seen": 485490688
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000861418712887177,
+      "loss": 0.0842,
+      "theoretical_loss": 3.9555614711116487,
+      "tokens_seen": 485752832
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008613384689455946,
+      "loss": 0.0841,
+      "theoretical_loss": 3.9553321410141162,
+      "tokens_seen": 486014976
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008612582250040122,
+      "loss": 0.0831,
+      "theoretical_loss": 3.9551029691912545,
+      "tokens_seen": 486277120
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008611779810624298,
+      "loss": 0.0843,
+      "theoretical_loss": 3.954873955448594,
+      "tokens_seen": 486539264
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008610977371208473,
+      "loss": 0.0846,
+      "theoretical_loss": 3.9546450995920086,
+      "tokens_seen": 486801408
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008610174931792651,
+      "loss": 0.0827,
+      "theoretical_loss": 3.954416401427715,
+      "tokens_seen": 487063552
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008609372492376826,
+      "loss": 0.0827,
+      "theoretical_loss": 3.9541878607622705,
+      "tokens_seen": 487325696
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008608570052961001,
+      "loss": 0.0805,
+      "theoretical_loss": 3.9539594774025755,
+      "tokens_seen": 487587840
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008607767613545178,
+      "loss": 0.084,
+      "theoretical_loss": 3.9537312511558698,
+      "tokens_seen": 487849984
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008606965174129353,
+      "loss": 0.0825,
+      "theoretical_loss": 3.953503181829732,
+      "tokens_seen": 488112128
+    },
+    {
+      "epoch": 0.15,
+      "objective/train/advantage_avg": 0.0009867006447166204,
+      "objective/train/docs_used": 183865,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7441798448562622,
+      "objective/train/original_loss": 1.7441800832748413,
+      "objective/train/theoretical_loss": 3.9533892059518383,
+      "objective/train/tokens_used": 508703200,
+      "objective/train/value_avg": -0.01270294189453125,
+      "objective/train/value_loss": 0.0009546683286316693,
+      "objective/train/value_max": -0.00030303001403808594,
+      "objective/train/value_min": -0.64697265625,
+      "objective/train/value_reward_corr": 0.6314438238999378,
+      "objective/train/value_std": 0.02288818359375,
+      "objective/train/weight_avg": 1.0013914108276367,
+      "objective/train/weighted_lm_loss": 1.7459317445755005,
+      "objective/train/weights_max": 1.6658852100372314,
+      "objective/train/weights_min": 0.3877171576023102,
+      "theoretical_loss": 3.9533892059518383,
+      "tokens_seen": 488243200
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000860616273471353,
+      "loss": 0.0843,
+      "theoretical_loss": 3.9532752692320816,
+      "tokens_seen": 488374272
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008605360295297705,
+      "loss": 0.0811,
+      "theoretical_loss": 3.9530475131711746,
+      "tokens_seen": 488636416
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008604557855881881,
+      "loss": 0.0826,
+      "theoretical_loss": 3.9528199134556044,
+      "tokens_seen": 488898560
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008603755416466057,
+      "loss": 0.0843,
+      "theoretical_loss": 3.9525924698943022,
+      "tokens_seen": 489160704
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008602952977050232,
+      "loss": 0.0833,
+      "theoretical_loss": 3.952365182296533,
+      "tokens_seen": 489422848
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008602150537634409,
+      "loss": 0.0848,
+      "theoretical_loss": 3.9521380504718975,
+      "tokens_seen": 489684992
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008601348098218585,
+      "loss": 0.0852,
+      "theoretical_loss": 3.9519110742303325,
+      "tokens_seen": 489947136
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008600545658802761,
+      "loss": 0.0831,
+      "theoretical_loss": 3.951684253382105,
+      "tokens_seen": 490209280
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008599743219386936,
+      "loss": 0.0832,
+      "theoretical_loss": 3.951457587737817,
+      "tokens_seen": 490471424
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008598940779971113,
+      "loss": 0.0834,
+      "theoretical_loss": 3.9512310771084014,
+      "tokens_seen": 490733568
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008598138340555288,
+      "loss": 0.0848,
+      "theoretical_loss": 3.951004721305123,
+      "tokens_seen": 490995712
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008597335901139463,
+      "loss": 0.0833,
+      "theoretical_loss": 3.950778520139576,
+      "tokens_seen": 491257856
+    },
+    {
+      "epoch": 0.15,
+      "objective/train/advantage_avg": -0.0002496841480024159,
+      "objective/train/docs_used": 185162,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6399861574172974,
+      "objective/train/original_loss": 1.639986276626587,
+      "objective/train/theoretical_loss": 3.9505524734236848,
+      "objective/train/tokens_used": 511980000,
+      "objective/train/value_avg": -0.00838470458984375,
+      "objective/train/value_loss": 0.00021605678193736821,
+      "objective/train/value_max": -0.0002148151397705078,
+      "objective/train/value_min": -0.469970703125,
+      "objective/train/value_reward_corr": 0.7257053344268602,
+      "objective/train/value_std": 0.01364898681640625,
+      "objective/train/weight_avg": 0.9998546242713928,
+      "objective/train/weighted_lm_loss": 1.6394292116165161,
+      "objective/train/weights_max": 1.1676921844482422,
+      "objective/train/weights_min": 0.5068990588188171,
+      "theoretical_loss": 3.9505524734236848,
+      "tokens_seen": 491520000
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000859653346172364,
+      "loss": 0.0843,
+      "theoretical_loss": 3.9505524734236848,
+      "tokens_seen": 491520000
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008595731022307815,
+      "loss": 0.0814,
+      "theoretical_loss": 3.950326580969703,
+      "tokens_seen": 491782144
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008594928582891992,
+      "loss": 0.0857,
+      "theoretical_loss": 3.950100842590212,
+      "tokens_seen": 492044288
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008594126143476168,
+      "loss": 0.085,
+      "theoretical_loss": 3.949875258098121,
+      "tokens_seen": 492306432
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008593323704060344,
+      "loss": 0.0855,
+      "theoretical_loss": 3.949649827306665,
+      "tokens_seen": 492568576
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000859252126464452,
+      "loss": 0.0843,
+      "theoretical_loss": 3.9494245500294047,
+      "tokens_seen": 492830720
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008591718825228696,
+      "loss": 0.084,
+      "theoretical_loss": 3.949199426080228,
+      "tokens_seen": 493092864
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008590916385812871,
+      "loss": 0.0851,
+      "theoretical_loss": 3.9489744552733455,
+      "tokens_seen": 493355008
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008590113946397047,
+      "loss": 0.0891,
+      "theoretical_loss": 3.9487496374232913,
+      "tokens_seen": 493617152
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008589311506981223,
+      "loss": 0.0832,
+      "theoretical_loss": 3.9485249723449236,
+      "tokens_seen": 493879296
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008588509067565398,
+      "loss": 0.0826,
+      "theoretical_loss": 3.9483004598534217,
+      "tokens_seen": 494141440
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008587706628149576,
+      "loss": 0.0859,
+      "theoretical_loss": 3.948076099764288,
+      "tokens_seen": 494403584
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008586904188733751,
+      "loss": 0.0891,
+      "theoretical_loss": 3.947851891893343,
+      "tokens_seen": 494665728
+    },
+    {
+      "epoch": 0.15,
+      "objective/train/advantage_avg": 0.0016816434217616916,
+      "objective/train/docs_used": 186433,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7250248193740845,
+      "objective/train/original_loss": 1.725024700164795,
+      "objective/train/theoretical_loss": 3.947739844982224,
+      "objective/train/tokens_used": 515256800,
+      "objective/train/value_avg": -0.00727081298828125,
+      "objective/train/value_loss": 0.00025146797997877,
+      "objective/train/value_max": -0.0002148151397705078,
+      "objective/train/value_min": -0.45849609375,
+      "objective/train/value_reward_corr": 0.5433043884355142,
+      "objective/train/value_std": 0.01275634765625,
+      "objective/train/weight_avg": 1.0017974376678467,
+      "objective/train/weighted_lm_loss": 1.7270184755325317,
+      "objective/train/weights_max": 1.2517598867416382,
+      "objective/train/weights_min": 0.3689996004104614,
+      "theoretical_loss": 3.947739844982224,
+      "tokens_seen": 494796800
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008586101749317927,
+      "loss": 0.0807,
+      "theoretical_loss": 3.9476278360567303,
+      "tokens_seen": 494927872
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008585299309902103,
+      "loss": 0.0842,
+      "theoretical_loss": 3.9474039320709107,
+      "tokens_seen": 495190016
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008584496870486278,
+      "loss": 0.0869,
+      "theoretical_loss": 3.9471801797526633,
+      "tokens_seen": 495452160
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008583694431070454,
+      "loss": 0.0824,
+      "theoretical_loss": 3.946956578919088,
+      "tokens_seen": 495714304
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000858289199165463,
+      "loss": 0.0844,
+      "theoretical_loss": 3.9467331293875976,
+      "tokens_seen": 495976448
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008582089552238806,
+      "loss": 0.0856,
+      "theoretical_loss": 3.9465098309759252,
+      "tokens_seen": 496238592
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008581287112822982,
+      "loss": 0.0844,
+      "theoretical_loss": 3.9462866835021178,
+      "tokens_seen": 496500736
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008580484673407159,
+      "loss": 0.0841,
+      "theoretical_loss": 3.9460636867845365,
+      "tokens_seen": 496762880
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008579682233991334,
+      "loss": 0.0852,
+      "theoretical_loss": 3.9458408406418584,
+      "tokens_seen": 497025024
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000857887979457551,
+      "loss": 0.0824,
+      "theoretical_loss": 3.945618144893073,
+      "tokens_seen": 497287168
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008578077355159686,
+      "loss": 0.0869,
+      "theoretical_loss": 3.9453955993574845,
+      "tokens_seen": 497549312
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008577274915743861,
+      "loss": 0.0834,
+      "theoretical_loss": 3.945173203854707,
+      "tokens_seen": 497811456
+    },
+    {
+      "epoch": 0.15,
+      "objective/train/advantage_avg": 6.547255179611966e-05,
+      "objective/train/docs_used": 187487,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.78741455078125,
+      "objective/train/original_loss": 1.78741455078125,
+      "objective/train/theoretical_loss": 3.9449509582046662,
+      "objective/train/tokens_used": 518533600,
+      "objective/train/value_avg": -0.00795745849609375,
+      "objective/train/value_loss": 0.0003167542163282633,
+      "objective/train/value_max": -0.00017404556274414062,
+      "objective/train/value_min": -0.7861328125,
+      "objective/train/value_reward_corr": 0.6039141734249502,
+      "objective/train/value_std": 0.0124359130859375,
+      "objective/train/weight_avg": 1.0002011060714722,
+      "objective/train/weighted_lm_loss": 1.7879151105880737,
+      "objective/train/weights_max": 1.272313117980957,
+      "objective/train/weights_min": 0.3683635890483856,
+      "theoretical_loss": 3.9449509582046662,
+      "tokens_seen": 498073600
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008576472476328038,
+      "loss": 0.0857,
+      "theoretical_loss": 3.9449509582046662,
+      "tokens_seen": 498073600
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008575670036912213,
+      "loss": 0.0836,
+      "theoretical_loss": 3.944728862227601,
+      "tokens_seen": 498335744
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008574867597496389,
+      "loss": 0.0814,
+      "theoretical_loss": 3.9445069157440575,
+      "tokens_seen": 498597888
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008574065158080565,
+      "loss": 0.0849,
+      "theoretical_loss": 3.944285118574893,
+      "tokens_seen": 498860032
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000857326271866474,
+      "loss": 0.0845,
+      "theoretical_loss": 3.9440634705412725,
+      "tokens_seen": 499122176
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008572460279248917,
+      "loss": 0.0855,
+      "theoretical_loss": 3.9438419714646695,
+      "tokens_seen": 499384320
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008571657839833093,
+      "loss": 0.0871,
+      "theoretical_loss": 3.9436206211668647,
+      "tokens_seen": 499646464
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008570855400417269,
+      "loss": 0.0809,
+      "theoretical_loss": 3.9433994194699453,
+      "tokens_seen": 499908608
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008570052961001444,
+      "loss": 0.0838,
+      "theoretical_loss": 3.943178366196304,
+      "tokens_seen": 500170752
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008569250521585621,
+      "loss": 0.0851,
+      "theoretical_loss": 3.942957461168639,
+      "tokens_seen": 500432896
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008568448082169796,
+      "loss": 0.0872,
+      "theoretical_loss": 3.9427367042099544,
+      "tokens_seen": 500695040
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008567645642753972,
+      "loss": 0.0811,
+      "theoretical_loss": 3.942516095143555,
+      "tokens_seen": 500957184
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008566843203338148,
+      "loss": 0.0849,
+      "theoretical_loss": 3.9422956337930524,
+      "tokens_seen": 501219328
+    },
+    {
+      "epoch": 0.15,
+      "objective/train/advantage_avg": 0.0007019721670076251,
+      "objective/train/docs_used": 188727,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5103391408920288,
+      "objective/train/original_loss": 1.5103389024734497,
+      "objective/train/theoretical_loss": 3.9421854584562226,
+      "objective/train/tokens_used": 521810400,
+      "objective/train/value_avg": -0.0062103271484375,
+      "objective/train/value_loss": 0.00019669736502692103,
+      "objective/train/value_max": -0.00018966197967529297,
+      "objective/train/value_min": -0.333251953125,
+      "objective/train/value_reward_corr": 0.5123174818364106,
+      "objective/train/value_std": 0.00872802734375,
+      "objective/train/weight_avg": 1.0007885694503784,
+      "objective/train/weighted_lm_loss": 1.5108174085617065,
+      "objective/train/weights_max": 1.1460858583450317,
+      "objective/train/weights_min": 0.3740321099758148,
+      "theoretical_loss": 3.9421854584562226,
+      "tokens_seen": 501350400
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008566040763922323,
+      "loss": 0.0823,
+      "theoretical_loss": 3.942075319982358,
+      "tokens_seen": 501481472
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008565238324506501,
+      "loss": 0.083,
+      "theoretical_loss": 3.941855153535686,
+      "tokens_seen": 501743616
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008564435885090676,
+      "loss": 0.0841,
+      "theoretical_loss": 3.9416351342775524,
+      "tokens_seen": 502005760
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008563633445674852,
+      "loss": 0.0828,
+      "theoretical_loss": 3.9414152620327716,
+      "tokens_seen": 502267904
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008562831006259028,
+      "loss": 0.0801,
+      "theoretical_loss": 3.941195536626461,
+      "tokens_seen": 502530048
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008562028566843203,
+      "loss": 0.084,
+      "theoretical_loss": 3.940975957884034,
+      "tokens_seen": 502792192
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008561226127427379,
+      "loss": 0.0842,
+      "theoretical_loss": 3.9407565256312047,
+      "tokens_seen": 503054336
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008560423688011555,
+      "loss": 0.0861,
+      "theoretical_loss": 3.940537239693983,
+      "tokens_seen": 503316480
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008559621248595731,
+      "loss": 0.0849,
+      "theoretical_loss": 3.9403180998986778,
+      "tokens_seen": 503578624
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008558818809179906,
+      "loss": 0.081,
+      "theoretical_loss": 3.9400991060718935,
+      "tokens_seen": 503840768
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008558016369764084,
+      "loss": 0.087,
+      "theoretical_loss": 3.93988025804053,
+      "tokens_seen": 504102912
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008557213930348259,
+      "loss": 0.0848,
+      "theoretical_loss": 3.9396615556317838,
+      "tokens_seen": 504365056
+    },
+    {
+      "epoch": 0.15,
+      "objective/train/advantage_avg": 0.0015238908817991614,
+      "objective/train/docs_used": 189911,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5270758867263794,
+      "objective/train/original_loss": 1.527076005935669,
+      "objective/train/theoretical_loss": 3.939442998673144,
+      "objective/train/tokens_used": 525087200,
+      "objective/train/value_avg": -0.0095672607421875,
+      "objective/train/value_loss": 0.0004635912482626736,
+      "objective/train/value_max": -0.00027370452880859375,
+      "objective/train/value_min": -0.77197265625,
+      "objective/train/value_reward_corr": 0.581789120363388,
+      "objective/train/value_std": 0.0157928466796875,
+      "objective/train/weight_avg": 1.0017266273498535,
+      "objective/train/weighted_lm_loss": 1.5284429788589478,
+      "objective/train/weights_max": 1.6315034627914429,
+      "objective/train/weights_min": 0.3693629503250122,
+      "theoretical_loss": 3.939442998673144,
+      "tokens_seen": 504627200
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008556411490932436,
+      "loss": 0.0826,
+      "theoretical_loss": 3.939442998673144,
+      "tokens_seen": 504627200
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008555609051516611,
+      "loss": 0.0827,
+      "theoretical_loss": 3.9392245869923954,
+      "tokens_seen": 504889344
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008554806612100786,
+      "loss": 0.087,
+      "theoretical_loss": 3.939006320417614,
+      "tokens_seen": 505151488
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008554004172684963,
+      "loss": 0.0838,
+      "theoretical_loss": 3.9387881987771705,
+      "tokens_seen": 505413632
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008553201733269138,
+      "loss": 0.0852,
+      "theoretical_loss": 3.9385702218997247,
+      "tokens_seen": 505675776
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008552399293853314,
+      "loss": 0.0852,
+      "theoretical_loss": 3.9383523896142316,
+      "tokens_seen": 505937920
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000855159685443749,
+      "loss": 0.0851,
+      "theoretical_loss": 3.9381347017499326,
+      "tokens_seen": 506200064
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008550794415021667,
+      "loss": 0.0852,
+      "theoretical_loss": 3.9379171581363623,
+      "tokens_seen": 506462208
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008549991975605842,
+      "loss": 0.0849,
+      "theoretical_loss": 3.937699758603342,
+      "tokens_seen": 506724352
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008549189536190018,
+      "loss": 0.085,
+      "theoretical_loss": 3.937482502980985,
+      "tokens_seen": 506986496
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008548387096774194,
+      "loss": 0.0878,
+      "theoretical_loss": 3.937265391099688,
+      "tokens_seen": 507248640
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008547584657358369,
+      "loss": 0.0814,
+      "theoretical_loss": 3.9370484227901397,
+      "tokens_seen": 507510784
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008546782217942546,
+      "loss": 0.0822,
+      "theoretical_loss": 3.9368315978833124,
+      "tokens_seen": 507772928
+    },
+    {
+      "epoch": 0.15,
+      "objective/train/advantage_avg": 0.0013619712553918362,
+      "objective/train/docs_used": 191028,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.592982530593872,
+      "objective/train/original_loss": 1.592982530593872,
+      "objective/train/theoretical_loss": 3.936723239153177,
+      "objective/train/tokens_used": 528364000,
+      "objective/train/value_avg": -0.00579071044921875,
+      "objective/train/value_loss": 0.00020233231771271676,
+      "objective/train/value_max": -0.00026535987854003906,
+      "objective/train/value_min": -0.587890625,
+      "objective/train/value_reward_corr": 0.7309359003301154,
+      "objective/train/value_std": 0.0113677978515625,
+      "objective/train/weight_avg": 1.00145423412323,
+      "objective/train/weighted_lm_loss": 1.5958586931228638,
+      "objective/train/weights_max": 1.2670164108276367,
+      "objective/train/weights_min": 0.38817885518074036,
+      "theoretical_loss": 3.936723239153177,
+      "tokens_seen": 507904000
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008545979778526721,
+      "loss": 0.0828,
+      "theoretical_loss": 3.936614916210466,
+      "tokens_seen": 508035072
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008545177339110897,
+      "loss": 0.0816,
+      "theoretical_loss": 3.9363983776031457,
+      "tokens_seen": 508297216
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008544374899695073,
+      "loss": 0.085,
+      "theoretical_loss": 3.936181981893182,
+      "tokens_seen": 508559360
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008543572460279248,
+      "loss": 0.0871,
+      "theoretical_loss": 3.9359657289126875,
+      "tokens_seen": 508821504
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008542770020863426,
+      "loss": 0.0858,
+      "theoretical_loss": 3.935749618494061,
+      "tokens_seen": 509083648
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008541967581447601,
+      "loss": 0.0846,
+      "theoretical_loss": 3.935533650469983,
+      "tokens_seen": 509345792
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008541165142031777,
+      "loss": 0.0831,
+      "theoretical_loss": 3.935317824673417,
+      "tokens_seen": 509607936
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008540362702615953,
+      "loss": 0.0852,
+      "theoretical_loss": 3.935102140937608,
+      "tokens_seen": 509870080
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008539560263200129,
+      "loss": 0.0837,
+      "theoretical_loss": 3.934886599096081,
+      "tokens_seen": 510132224
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008538757823784304,
+      "loss": 0.0828,
+      "theoretical_loss": 3.9346711989826426,
+      "tokens_seen": 510394368
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.000853795538436848,
+      "loss": 0.0842,
+      "theoretical_loss": 3.93445594043138,
+      "tokens_seen": 510656512
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008537152944952656,
+      "loss": 0.0831,
+      "theoretical_loss": 3.9342408232766584,
+      "tokens_seen": 510918656
+    },
+    {
+      "epoch": 0.15,
+      "objective/train/advantage_avg": 0.0022731111384928226,
+      "objective/train/docs_used": 192198,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7514947652816772,
+      "objective/train/original_loss": 1.7514947652816772,
+      "objective/train/theoretical_loss": 3.934025847353122,
+      "objective/train/tokens_used": 531640800,
+      "objective/train/value_avg": -0.006824493408203125,
+      "objective/train/value_loss": 0.00011324948718538508,
+      "objective/train/value_max": -0.0002779960632324219,
+      "objective/train/value_min": -0.3486328125,
+      "objective/train/value_reward_corr": 0.5780862010598011,
+      "objective/train/value_std": 0.007709503173828125,
+      "objective/train/weight_avg": 1.002328634262085,
+      "objective/train/weighted_lm_loss": 1.7554271221160889,
+      "objective/train/weights_max": 1.185212254524231,
+      "objective/train/weights_min": 0.7214187383651733,
+      "theoretical_loss": 3.934025847353122,
+      "tokens_seen": 511180800
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008536350505536831,
+      "loss": 0.0848,
+      "theoretical_loss": 3.934025847353122,
+      "tokens_seen": 511180800
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 0.0008535548066121009,
+      "loss": 0.0827,
+      "theoretical_loss": 3.9338110124956924,
+      "tokens_seen": 511442944
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008534745626705184,
+      "loss": 0.0846,
+      "theoretical_loss": 3.9335963185395713,
+      "tokens_seen": 511705088
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.000853394318728936,
+      "loss": 0.0818,
+      "theoretical_loss": 3.933381765320233,
+      "tokens_seen": 511967232
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008533140747873536,
+      "loss": 0.0821,
+      "theoretical_loss": 3.933167352673432,
+      "tokens_seen": 512229376
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008532338308457711,
+      "loss": 0.0843,
+      "theoretical_loss": 3.9329530804351958,
+      "tokens_seen": 512491520
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008531535869041887,
+      "loss": 0.0862,
+      "theoretical_loss": 3.9327389484418287,
+      "tokens_seen": 512753664
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008530733429626063,
+      "loss": 0.0808,
+      "theoretical_loss": 3.9325249565299076,
+      "tokens_seen": 513015808
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008529930990210239,
+      "loss": 0.085,
+      "theoretical_loss": 3.932311104536285,
+      "tokens_seen": 513277952
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008529128550794415,
+      "loss": 0.0829,
+      "theoretical_loss": 3.9320973922980844,
+      "tokens_seen": 513540096
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008528326111378592,
+      "loss": 0.0849,
+      "theoretical_loss": 3.931883819652705,
+      "tokens_seen": 513802240
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008527523671962767,
+      "loss": 0.0826,
+      "theoretical_loss": 3.9316703864378155,
+      "tokens_seen": 514064384
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008526721232546944,
+      "loss": 0.0825,
+      "theoretical_loss": 3.9314570924913568,
+      "tokens_seen": 514326528
+    },
+    {
+      "epoch": 0.16,
+      "objective/train/advantage_avg": 0.0023866831324994564,
+      "objective/train/docs_used": 193343,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5726442337036133,
+      "objective/train/original_loss": 1.5726442337036133,
+      "objective/train/theoretical_loss": 3.931350497693219,
+      "objective/train/tokens_used": 534917600,
+      "objective/train/value_avg": -0.00785064697265625,
+      "objective/train/value_loss": 0.00021415037917904556,
+      "objective/train/value_max": -0.0003528594970703125,
+      "objective/train/value_min": -0.262451171875,
+      "objective/train/value_reward_corr": 0.5768801997308433,
+      "objective/train/value_std": 0.01004791259765625,
+      "objective/train/weight_avg": 1.002478003501892,
+      "objective/train/weighted_lm_loss": 1.5768412351608276,
+      "objective/train/weights_max": 1.115512490272522,
+      "objective/train/weights_min": 0.36873859167099,
+      "theoretical_loss": 3.931350497693219,
+      "tokens_seen": 514457600
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008525918793131119,
+      "loss": 0.0841,
+      "theoretical_loss": 3.9312439376515407,
+      "tokens_seen": 514588672
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008525116353715294,
+      "loss": 0.0836,
+      "theoretical_loss": 3.9310309217568493,
+      "tokens_seen": 514850816
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008524313914299471,
+      "loss": 0.0851,
+      "theoretical_loss": 3.9308180446460343,
+      "tokens_seen": 515112960
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008523511474883646,
+      "loss": 0.0845,
+      "theoretical_loss": 3.9306053061581165,
+      "tokens_seen": 515375104
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008522709035467822,
+      "loss": 0.0838,
+      "theoretical_loss": 3.930392706132385,
+      "tokens_seen": 515637248
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008521906596051998,
+      "loss": 0.085,
+      "theoretical_loss": 3.9301802444083966,
+      "tokens_seen": 515899392
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008521104156636175,
+      "loss": 0.0851,
+      "theoretical_loss": 3.929967920825977,
+      "tokens_seen": 516161536
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.000852030171722035,
+      "loss": 0.0809,
+      "theoretical_loss": 3.929755735225216,
+      "tokens_seen": 516423680
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008519499277804526,
+      "loss": 0.0836,
+      "theoretical_loss": 3.9295436874464715,
+      "tokens_seen": 516685824
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008518696838388702,
+      "loss": 0.0839,
+      "theoretical_loss": 3.929331777330366,
+      "tokens_seen": 516947968
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008517894398972877,
+      "loss": 0.0802,
+      "theoretical_loss": 3.9291200047177886,
+      "tokens_seen": 517210112
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008517091959557054,
+      "loss": 0.0821,
+      "theoretical_loss": 3.9289083694498905,
+      "tokens_seen": 517472256
+    },
+    {
+      "epoch": 0.16,
+      "objective/train/advantage_avg": 0.00018064206233248115,
+      "objective/train/docs_used": 194581,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6930550336837769,
+      "objective/train/original_loss": 1.6930551528930664,
+      "objective/train/theoretical_loss": 3.9286968713680883,
+      "objective/train/tokens_used": 538194400,
+      "objective/train/value_avg": -0.00974273681640625,
+      "objective/train/value_loss": 0.0005410881130956113,
+      "objective/train/value_max": -0.0002892017364501953,
+      "objective/train/value_min": -0.72412109375,
+      "objective/train/value_reward_corr": 0.7755958663238911,
+      "objective/train/value_std": 0.0196533203125,
+      "objective/train/weight_avg": 1.0004180669784546,
+      "objective/train/weighted_lm_loss": 1.694211721420288,
+      "objective/train/weights_max": 1.331702709197998,
+      "objective/train/weights_min": 0.3937867283821106,
+      "theoretical_loss": 3.9286968713680883,
+      "tokens_seen": 517734400
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008516289520141229,
+      "loss": 0.0811,
+      "theoretical_loss": 3.9286968713680883,
+      "tokens_seen": 517734400
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008515487080725406,
+      "loss": 0.0829,
+      "theoretical_loss": 3.9284855103140615,
+      "tokens_seen": 517996544
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008514684641309581,
+      "loss": 0.0827,
+      "theoretical_loss": 3.9282742861297524,
+      "tokens_seen": 518258688
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008513882201893756,
+      "loss": 0.0793,
+      "theoretical_loss": 3.928063198657365,
+      "tokens_seen": 518520832
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008513079762477934,
+      "loss": 0.0814,
+      "theoretical_loss": 3.9278522477393656,
+      "tokens_seen": 518782976
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008512277323062109,
+      "loss": 0.0829,
+      "theoretical_loss": 3.9276414332184815,
+      "tokens_seen": 519045120
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008511474883646285,
+      "loss": 0.0823,
+      "theoretical_loss": 3.927430754937699,
+      "tokens_seen": 519307264
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008510672444230461,
+      "loss": 0.08,
+      "theoretical_loss": 3.927220212740267,
+      "tokens_seen": 519569408
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008509870004814637,
+      "loss": 0.0799,
+      "theoretical_loss": 3.9270098064696906,
+      "tokens_seen": 519831552
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008509067565398812,
+      "loss": 0.0818,
+      "theoretical_loss": 3.9267995359697356,
+      "tokens_seen": 520093696
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008508265125982988,
+      "loss": 0.0823,
+      "theoretical_loss": 3.926589401084426,
+      "tokens_seen": 520355840
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008507462686567164,
+      "loss": 0.0807,
+      "theoretical_loss": 3.9263794016580427,
+      "tokens_seen": 520617984
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008506660247151339,
+      "loss": 0.0803,
+      "theoretical_loss": 3.9261695375351238,
+      "tokens_seen": 520880128
+    },
+    {
+      "epoch": 0.16,
+      "objective/train/advantage_avg": 0.0009029977954924107,
+      "objective/train/docs_used": 195804,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7540147304534912,
+      "objective/train/original_loss": 1.7540148496627808,
+      "objective/train/theoretical_loss": 3.926064656163952,
+      "objective/train/tokens_used": 541471200,
+      "objective/train/value_avg": -0.01312255859375,
+      "objective/train/value_loss": 0.0006540497415699065,
+      "objective/train/value_max": -0.00019562244415283203,
+      "objective/train/value_min": -0.7529296875,
+      "objective/train/value_reward_corr": 0.6588186474670367,
+      "objective/train/value_std": 0.021728515625,
+      "objective/train/weight_avg": 1.0011897087097168,
+      "objective/train/weighted_lm_loss": 1.7548198699951172,
+      "objective/train/weights_max": 1.322047233581543,
+      "objective/train/weights_min": 0.3694334030151367,
+      "theoretical_loss": 3.926064656163952,
+      "tokens_seen": 521011200
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008505857807735517,
+      "loss": 0.0844,
+      "theoretical_loss": 3.9259598085604646,
+      "tokens_seen": 521142272
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008505055368319692,
+      "loss": 0.0813,
+      "theoretical_loss": 3.925750214579116,
+      "tokens_seen": 521404416
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008504252928903869,
+      "loss": 0.0848,
+      "theoretical_loss": 3.9255407554363835,
+      "tokens_seen": 521666560
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008503450489488044,
+      "loss": 0.0816,
+      "theoretical_loss": 3.92533143097783,
+      "tokens_seen": 521928704
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008502648050072219,
+      "loss": 0.083,
+      "theoretical_loss": 3.9251222410492694,
+      "tokens_seen": 522190848
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008501845610656396,
+      "loss": 0.0834,
+      "theoretical_loss": 3.924913185496772,
+      "tokens_seen": 522452992
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008501043171240571,
+      "loss": 0.0817,
+      "theoretical_loss": 3.924704264166659,
+      "tokens_seen": 522715136
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008500240731824747,
+      "loss": 0.082,
+      "theoretical_loss": 3.9244954769055074,
+      "tokens_seen": 522977280
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008499438292408923,
+      "loss": 0.0844,
+      "theoretical_loss": 3.924286823560144,
+      "tokens_seen": 523239424
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.00084986358529931,
+      "loss": 0.0809,
+      "theoretical_loss": 3.9240783039776472,
+      "tokens_seen": 523501568
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008497833413577275,
+      "loss": 0.0828,
+      "theoretical_loss": 3.9238699180053485,
+      "tokens_seen": 523763712
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008497030974161451,
+      "loss": 0.0837,
+      "theoretical_loss": 3.923661665490828,
+      "tokens_seen": 524025856
+    },
+    {
+      "epoch": 0.16,
+      "objective/train/advantage_avg": 0.0017735909204930067,
+      "objective/train/docs_used": 196964,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5325592756271362,
+      "objective/train/original_loss": 1.5325592756271362,
+      "objective/train/theoretical_loss": 3.9234535462819156,
+      "objective/train/tokens_used": 544748000,
+      "objective/train/value_avg": -0.00875091552734375,
+      "objective/train/value_loss": 0.00024432019563391805,
+      "objective/train/value_max": -0.0002378225326538086,
+      "objective/train/value_min": -0.404052734375,
+      "objective/train/value_reward_corr": 0.6240298333146145,
+      "objective/train/value_std": 0.0148773193359375,
+      "objective/train/weight_avg": 1.0018869638442993,
+      "objective/train/weighted_lm_loss": 1.5350106954574585,
+      "objective/train/weights_max": 1.2263628244400024,
+      "objective/train/weights_min": 0.37659770250320435,
+      "theoretical_loss": 3.9234535462819156,
+      "tokens_seen": 524288000
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008496228534745627,
+      "loss": 0.0818,
+      "theoretical_loss": 3.9234535462819156,
+      "tokens_seen": 524288000
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008495426095329802,
+      "loss": 0.0825,
+      "theoretical_loss": 3.923245560226693,
+      "tokens_seen": 524550144
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008494623655913979,
+      "loss": 0.0794,
+      "theoretical_loss": 3.9230377071734885,
+      "tokens_seen": 524812288
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008493821216498154,
+      "loss": 0.0822,
+      "theoretical_loss": 3.9228299869708794,
+      "tokens_seen": 525074432
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.000849301877708233,
+      "loss": 0.0828,
+      "theoretical_loss": 3.9226223994676923,
+      "tokens_seen": 525336576
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008492216337666506,
+      "loss": 0.0855,
+      "theoretical_loss": 3.9224149445129983,
+      "tokens_seen": 525598720
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008491413898250683,
+      "loss": 0.0816,
+      "theoretical_loss": 3.922207621956119,
+      "tokens_seen": 525860864
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008490611458834859,
+      "loss": 0.0808,
+      "theoretical_loss": 3.9220004316466186,
+      "tokens_seen": 526123008
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008489809019419034,
+      "loss": 0.0842,
+      "theoretical_loss": 3.9217933734343093,
+      "tokens_seen": 526385152
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.000848900658000321,
+      "loss": 0.0824,
+      "theoretical_loss": 3.9215864471692488,
+      "tokens_seen": 526647296
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008488204140587386,
+      "loss": 0.0819,
+      "theoretical_loss": 3.921379652701738,
+      "tokens_seen": 526909440
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008487401701171562,
+      "loss": 0.0812,
+      "theoretical_loss": 3.9211729898823235,
+      "tokens_seen": 527171584
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008486599261755737,
+      "loss": 0.0821,
+      "theoretical_loss": 3.920966458561794,
+      "tokens_seen": 527433728
+    },
+    {
+      "epoch": 0.16,
+      "objective/train/advantage_avg": 0.00034622373641468585,
+      "objective/train/docs_used": 198271,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5029301643371582,
+      "objective/train/original_loss": 1.5029301643371582,
+      "objective/train/theoretical_loss": 3.9208632421670497,
+      "objective/train/tokens_used": 548024800,
+      "objective/train/value_avg": -0.00766754150390625,
+      "objective/train/value_loss": 0.0003754664212465286,
+      "objective/train/value_max": -0.0003077983856201172,
+      "objective/train/value_min": -0.64697265625,
+      "objective/train/value_reward_corr": 0.5987369406329422,
+      "objective/train/value_std": 0.0127410888671875,
+      "objective/train/weight_avg": 1.0005102157592773,
+      "objective/train/weighted_lm_loss": 1.5032236576080322,
+      "objective/train/weights_max": 1.5309104919433594,
+      "objective/train/weights_min": 0.37003564834594727,
+      "theoretical_loss": 3.9208632421670497,
+      "tokens_seen": 527564800
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008485796822339914,
+      "loss": 0.0817,
+      "theoretical_loss": 3.920760058591182,
+      "tokens_seen": 527695872
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008484994382924089,
+      "loss": 0.0827,
+      "theoretical_loss": 3.9205537898217644,
+      "tokens_seen": 527958016
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008484191943508264,
+      "loss": 0.0819,
+      "theoretical_loss": 3.920347652105058,
+      "tokens_seen": 528220160
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008483389504092442,
+      "loss": 0.0831,
+      "theoretical_loss": 3.920141645292821,
+      "tokens_seen": 528482304
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008482587064676617,
+      "loss": 0.0815,
+      "theoretical_loss": 3.919935769237055,
+      "tokens_seen": 528744448
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008481784625260793,
+      "loss": 0.0842,
+      "theoretical_loss": 3.91973002379,
+      "tokens_seen": 529006592
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008480982185844969,
+      "loss": 0.082,
+      "theoretical_loss": 3.919524408804137,
+      "tokens_seen": 529268736
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008480179746429145,
+      "loss": 0.0829,
+      "theoretical_loss": 3.9193189241321873,
+      "tokens_seen": 529530880
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008479377307013321,
+      "loss": 0.0823,
+      "theoretical_loss": 3.9191135696271098,
+      "tokens_seen": 529793024
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008478574867597496,
+      "loss": 0.0852,
+      "theoretical_loss": 3.9189083451421025,
+      "tokens_seen": 530055168
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008477772428181672,
+      "loss": 0.0784,
+      "theoretical_loss": 3.9187032505306023,
+      "tokens_seen": 530317312
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008476969988765848,
+      "loss": 0.0813,
+      "theoretical_loss": 3.918498285646282,
+      "tokens_seen": 530579456
+    },
+    {
+      "epoch": 0.16,
+      "objective/train/advantage_avg": 0.000702679913956672,
+      "objective/train/docs_used": 199409,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6489094495773315,
+      "objective/train/original_loss": 1.648909330368042,
+      "objective/train/theoretical_loss": 3.9182934503430538,
+      "objective/train/tokens_used": 551301600,
+      "objective/train/value_avg": -0.00879669189453125,
+      "objective/train/value_loss": 0.0003273165493737906,
+      "objective/train/value_max": -0.00041413307189941406,
+      "objective/train/value_min": -0.771484375,
+      "objective/train/value_reward_corr": 0.6633614602841378,
+      "objective/train/value_std": 0.0150604248046875,
+      "objective/train/weight_avg": 1.0008490085601807,
+      "objective/train/weighted_lm_loss": 1.6499123573303223,
+      "objective/train/weights_max": 1.3013832569122314,
+      "objective/train/weights_min": 0.38973385095596313,
+      "theoretical_loss": 3.9182934503430538,
+      "tokens_seen": 530841600
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008476167549350025,
+      "loss": 0.0819,
+      "theoretical_loss": 3.9182934503430538,
+      "tokens_seen": 530841600
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.00084753651099342,
+      "loss": 0.0795,
+      "theoretical_loss": 3.918088744475064,
+      "tokens_seen": 531103744
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008474562670518377,
+      "loss": 0.0844,
+      "theoretical_loss": 3.9178841678966956,
+      "tokens_seen": 531365888
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008473760231102552,
+      "loss": 0.0846,
+      "theoretical_loss": 3.9176797204625693,
+      "tokens_seen": 531628032
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008472957791686727,
+      "loss": 0.0833,
+      "theoretical_loss": 3.917475402027537,
+      "tokens_seen": 531890176
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008472155352270904,
+      "loss": 0.0799,
+      "theoretical_loss": 3.917271212446689,
+      "tokens_seen": 532152320
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008471352912855079,
+      "loss": 0.0822,
+      "theoretical_loss": 3.917067151575348,
+      "tokens_seen": 532414464
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008470550473439255,
+      "loss": 0.085,
+      "theoretical_loss": 3.916863219269069,
+      "tokens_seen": 532676608
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008469748034023431,
+      "loss": 0.0816,
+      "theoretical_loss": 3.9166594153836427,
+      "tokens_seen": 532938752
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008468945594607608,
+      "loss": 0.0823,
+      "theoretical_loss": 3.9164557397750897,
+      "tokens_seen": 533200896
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008468143155191783,
+      "loss": 0.0817,
+      "theoretical_loss": 3.916252192299665,
+      "tokens_seen": 533463040
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008467340715775959,
+      "loss": 0.0818,
+      "theoretical_loss": 3.9160487728138538,
+      "tokens_seen": 533725184
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008466538276360135,
+      "loss": 0.0826,
+      "theoretical_loss": 3.9158454811743733,
+      "tokens_seen": 533987328
+    },
+    {
+      "epoch": 0.16,
+      "objective/train/advantage_avg": 0.0011991261271759868,
+      "objective/train/docs_used": 200516,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.67826509475708,
+      "objective/train/original_loss": 1.6782649755477905,
+      "objective/train/theoretical_loss": 3.9157438832522944,
+      "objective/train/tokens_used": 554578400,
+      "objective/train/value_avg": -0.007232666015625,
+      "objective/train/value_loss": 0.00027039533597417176,
+      "objective/train/value_max": -0.00033545494079589844,
+      "objective/train/value_min": -0.525390625,
+      "objective/train/value_reward_corr": 0.46649640516327573,
+      "objective/train/value_std": 0.00963592529296875,
+      "objective/train/weight_avg": 1.0013190507888794,
+      "objective/train/weighted_lm_loss": 1.6796201467514038,
+      "objective/train/weights_max": 1.2117398977279663,
+      "objective/train/weights_min": 0.3694460988044739,
+      "theoretical_loss": 3.9157438832522944,
+      "tokens_seen": 534118400
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008465735836944311,
+      "loss": 0.0848,
+      "theoretical_loss": 3.915642317238171,
+      "tokens_seen": 534249472
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008464933397528487,
+      "loss": 0.084,
+      "theoretical_loss": 3.915439280862423,
+      "tokens_seen": 534511616
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008464130958112662,
+      "loss": 0.0814,
+      "theoretical_loss": 3.915236371904539,
+      "tokens_seen": 534773760
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008463328518696839,
+      "loss": 0.0826,
+      "theoretical_loss": 3.915033590222153,
+      "tokens_seen": 535035904
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008462526079281014,
+      "loss": 0.0826,
+      "theoretical_loss": 3.914830935673132,
+      "tokens_seen": 535298048
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008461723639865191,
+      "loss": 0.0795,
+      "theoretical_loss": 3.914628408115569,
+      "tokens_seen": 535560192
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008460921200449367,
+      "loss": 0.0816,
+      "theoretical_loss": 3.9144260074077843,
+      "tokens_seen": 535822336
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008460118761033542,
+      "loss": 0.0843,
+      "theoretical_loss": 3.9142237334083276,
+      "tokens_seen": 536084480
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008459316321617718,
+      "loss": 0.0818,
+      "theoretical_loss": 3.914021585975973,
+      "tokens_seen": 536346624
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008458513882201894,
+      "loss": 0.0808,
+      "theoretical_loss": 3.9138195649697227,
+      "tokens_seen": 536608768
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.000845771144278607,
+      "loss": 0.0827,
+      "theoretical_loss": 3.9136176702488044,
+      "tokens_seen": 536870912
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008456909003370245,
+      "loss": 0.0794,
+      "theoretical_loss": 3.91341590167267,
+      "tokens_seen": 537133056
+    },
+    {
+      "epoch": 0.16,
+      "objective/train/advantage_avg": 0.0008067527669481933,
+      "objective/train/docs_used": 201641,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5511680841445923,
+      "objective/train/original_loss": 1.5511682033538818,
+      "objective/train/theoretical_loss": 3.9132142591009975,
+      "objective/train/tokens_used": 557855200,
+      "objective/train/value_avg": -0.00760650634765625,
+      "objective/train/value_loss": 0.00028524218942038715,
+      "objective/train/value_max": -0.00030541419982910156,
+      "objective/train/value_min": -0.60986328125,
+      "objective/train/value_reward_corr": 0.643433561791619,
+      "objective/train/value_std": 0.0146026611328125,
+      "objective/train/weight_avg": 1.0009351968765259,
+      "objective/train/weighted_lm_loss": 1.551741361618042,
+      "objective/train/weights_max": 1.4019429683685303,
+      "objective/train/weights_min": 0.3742033541202545,
+      "theoretical_loss": 3.9132142591009975,
+      "tokens_seen": 537395200
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008456106563954422,
+      "loss": 0.0807,
+      "theoretical_loss": 3.9132142591009975,
+      "tokens_seen": 537395200
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008455304124538597,
+      "loss": 0.0813,
+      "theoretical_loss": 3.9130127423936907,
+      "tokens_seen": 537657344
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008454501685122772,
+      "loss": 0.0839,
+      "theoretical_loss": 3.9128113514108733,
+      "tokens_seen": 537919488
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.000845369924570695,
+      "loss": 0.0821,
+      "theoretical_loss": 3.9126100860128963,
+      "tokens_seen": 538181632
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008452896806291125,
+      "loss": 0.0836,
+      "theoretical_loss": 3.9124089460603324,
+      "tokens_seen": 538443776
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008452094366875302,
+      "loss": 0.0828,
+      "theoretical_loss": 3.9122079314139766,
+      "tokens_seen": 538705920
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008451291927459477,
+      "loss": 0.0828,
+      "theoretical_loss": 3.9120070419348463,
+      "tokens_seen": 538968064
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008450489488043653,
+      "loss": 0.0799,
+      "theoretical_loss": 3.9118062774841804,
+      "tokens_seen": 539230208
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008449687048627829,
+      "loss": 0.084,
+      "theoretical_loss": 3.91160563792344,
+      "tokens_seen": 539492352
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008448884609212004,
+      "loss": 0.0829,
+      "theoretical_loss": 3.911405123114305,
+      "tokens_seen": 539754496
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.000844808216979618,
+      "loss": 0.0847,
+      "theoretical_loss": 3.9112047329186783,
+      "tokens_seen": 540016640
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008447279730380356,
+      "loss": 0.0833,
+      "theoretical_loss": 3.911004467198679,
+      "tokens_seen": 540278784
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008446477290964533,
+      "loss": 0.0842,
+      "theoretical_loss": 3.9108043258166485,
+      "tokens_seen": 540540928
+    },
+    {
+      "epoch": 0.16,
+      "objective/train/advantage_avg": 0.0016820263117551804,
+      "objective/train/docs_used": 202818,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5808069705963135,
+      "objective/train/original_loss": 1.5808069705963135,
+      "objective/train/theoretical_loss": 3.910704301709413,
+      "objective/train/tokens_used": 561132000,
+      "objective/train/value_avg": -0.00955963134765625,
+      "objective/train/value_loss": 0.0003617853799369186,
+      "objective/train/value_max": -0.0003407001495361328,
+      "objective/train/value_min": -0.6962890625,
+      "objective/train/value_reward_corr": 0.5872441373349445,
+      "objective/train/value_std": 0.016387939453125,
+      "objective/train/weight_avg": 1.0018519163131714,
+      "objective/train/weighted_lm_loss": 1.5838106870651245,
+      "objective/train/weights_max": 1.8153696060180664,
+      "objective/train/weights_min": 0.3881314992904663,
+      "theoretical_loss": 3.910704301709413,
+      "tokens_seen": 540672000
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008445674851548708,
+      "loss": 0.0816,
+      "theoretical_loss": 3.910604308635146,
+      "tokens_seen": 540803072
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008444872412132885,
+      "loss": 0.0799,
+      "theoretical_loss": 3.9104044155169495,
+      "tokens_seen": 541065216
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.000844406997271706,
+      "loss": 0.0854,
+      "theoretical_loss": 3.910204646325055,
+      "tokens_seen": 541327360
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008443267533301235,
+      "loss": 0.0843,
+      "theoretical_loss": 3.9100050009226752,
+      "tokens_seen": 541589504
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008442465093885412,
+      "loss": 0.0833,
+      "theoretical_loss": 3.9098054791732406,
+      "tokens_seen": 541851648
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008441662654469587,
+      "loss": 0.0837,
+      "theoretical_loss": 3.909606080940399,
+      "tokens_seen": 542113792
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008440860215053764,
+      "loss": 0.0824,
+      "theoretical_loss": 3.909406806088013,
+      "tokens_seen": 542375936
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008440057775637939,
+      "loss": 0.0841,
+      "theoretical_loss": 3.909207654480162,
+      "tokens_seen": 542638080
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008439255336222116,
+      "loss": 0.082,
+      "theoretical_loss": 3.9090086259811403,
+      "tokens_seen": 542900224
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008438452896806292,
+      "loss": 0.0825,
+      "theoretical_loss": 3.908809720455457,
+      "tokens_seen": 543162368
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008437650457390467,
+      "loss": 0.0824,
+      "theoretical_loss": 3.908610937767836,
+      "tokens_seen": 543424512
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008436848017974643,
+      "loss": 0.083,
+      "theoretical_loss": 3.9084122777832144,
+      "tokens_seen": 543686656
+    },
+    {
+      "epoch": 0.16,
+      "objective/train/advantage_avg": -0.0008586941985413432,
+      "objective/train/docs_used": 204061,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6755656003952026,
+      "objective/train/original_loss": 1.675565481185913,
+      "objective/train/theoretical_loss": 3.908213740366744,
+      "objective/train/tokens_used": 564408800,
+      "objective/train/value_avg": -0.00843048095703125,
+      "objective/train/value_loss": 0.0002858492953237146,
+      "objective/train/value_max": -0.00023055076599121094,
+      "objective/train/value_min": -0.22607421875,
+      "objective/train/value_reward_corr": 0.6344704655141316,
+      "objective/train/value_std": 0.01132965087890625,
+      "objective/train/weight_avg": 0.9992753863334656,
+      "objective/train/weighted_lm_loss": 1.6743135452270508,
+      "objective/train/weights_max": 1.2291204929351807,
+      "objective/train/weights_min": 0.37150946259498596,
+      "theoretical_loss": 3.908213740366744,
+      "tokens_seen": 543948800
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008436045578558819,
+      "loss": 0.0836,
+      "theoretical_loss": 3.908213740366744,
+      "tokens_seen": 543948800
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.0008435243139142995,
+      "loss": 0.083,
+      "theoretical_loss": 3.908015325383788,
+      "tokens_seen": 544210944
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 0.000843444069972717,
+      "loss": 0.0823,
+      "theoretical_loss": 3.907817032699924,
+      "tokens_seen": 544473088
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008433638260311347,
+      "loss": 0.0813,
+      "theoretical_loss": 3.9076188621809416,
+      "tokens_seen": 544735232
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008432835820895522,
+      "loss": 0.0842,
+      "theoretical_loss": 3.9074208136928408,
+      "tokens_seen": 544997376
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008432033381479699,
+      "loss": 0.0833,
+      "theoretical_loss": 3.907222887101834,
+      "tokens_seen": 545259520
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008431230942063875,
+      "loss": 0.0858,
+      "theoretical_loss": 3.9070250822743446,
+      "tokens_seen": 545521664
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000843042850264805,
+      "loss": 0.0846,
+      "theoretical_loss": 3.906827399077006,
+      "tokens_seen": 545783808
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008429626063232226,
+      "loss": 0.083,
+      "theoretical_loss": 3.9066298373766615,
+      "tokens_seen": 546045952
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008428823623816402,
+      "loss": 0.0835,
+      "theoretical_loss": 3.9064323970403656,
+      "tokens_seen": 546308096
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008428021184400578,
+      "loss": 0.0835,
+      "theoretical_loss": 3.9062350779353787,
+      "tokens_seen": 546570240
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008427218744984754,
+      "loss": 0.083,
+      "theoretical_loss": 3.906037879929174,
+      "tokens_seen": 546832384
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000842641630556893,
+      "loss": 0.0816,
+      "theoretical_loss": 3.90584080288943,
+      "tokens_seen": 547094528
+    },
+    {
+      "epoch": 0.17,
+      "objective/train/advantage_avg": 0.00023532530758529902,
+      "objective/train/docs_used": 204823,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5397506952285767,
+      "objective/train/original_loss": 1.5397508144378662,
+      "objective/train/theoretical_loss": 3.9057423096906874,
+      "objective/train/tokens_used": 567685600,
+      "objective/train/value_avg": -0.00724029541015625,
+      "objective/train/value_loss": 0.0003734648635145277,
+      "objective/train/value_max": -0.00015354156494140625,
+      "objective/train/value_min": -0.76318359375,
+      "objective/train/value_reward_corr": 0.8074329039979607,
+      "objective/train/value_std": 0.0235595703125,
+      "objective/train/weight_avg": 1.0004050731658936,
+      "objective/train/weighted_lm_loss": 1.539077639579773,
+      "objective/train/weights_max": 1.4787062406539917,
+      "objective/train/weights_min": 0.3683115839958191,
+      "theoretical_loss": 3.9057423096906874,
+      "tokens_seen": 547225600
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008425613866153105,
+      "loss": 0.0832,
+      "theoretical_loss": 3.905643846684034,
+      "tokens_seen": 547356672
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008424811426737282,
+      "loss": 0.0801,
+      "theoretical_loss": 3.9054470111810815,
+      "tokens_seen": 547618816
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008424008987321458,
+      "loss": 0.0815,
+      "theoretical_loss": 3.9052502962488735,
+      "tokens_seen": 547880960
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008423206547905633,
+      "loss": 0.083,
+      "theoretical_loss": 3.9050537017559197,
+      "tokens_seen": 548143104
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000842240410848981,
+      "loss": 0.0831,
+      "theoretical_loss": 3.904857227570934,
+      "tokens_seen": 548405248
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008421601669073985,
+      "loss": 0.0789,
+      "theoretical_loss": 3.904660873562837,
+      "tokens_seen": 548667392
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008420799229658161,
+      "loss": 0.0817,
+      "theoretical_loss": 3.9044646396007545,
+      "tokens_seen": 548929536
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008419996790242337,
+      "loss": 0.0823,
+      "theoretical_loss": 3.9042685255540177,
+      "tokens_seen": 549191680
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008419194350826512,
+      "loss": 0.0831,
+      "theoretical_loss": 3.9040725312921616,
+      "tokens_seen": 549453824
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008418391911410688,
+      "loss": 0.0827,
+      "theoretical_loss": 3.9038766566849263,
+      "tokens_seen": 549715968
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008417589471994864,
+      "loss": 0.0825,
+      "theoretical_loss": 3.903680901602254,
+      "tokens_seen": 549978112
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008416787032579041,
+      "loss": 0.0814,
+      "theoretical_loss": 3.9034852659142913,
+      "tokens_seen": 550240256
+    },
+    {
+      "epoch": 0.17,
+      "objective/train/advantage_avg": 0.0012200219789519906,
+      "objective/train/docs_used": 206347,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7369581460952759,
+      "objective/train/original_loss": 1.7369580268859863,
+      "objective/train/theoretical_loss": 3.9032897494913876,
+      "objective/train/tokens_used": 570962400,
+      "objective/train/value_avg": -0.00876617431640625,
+      "objective/train/value_loss": 0.0002591983356978744,
+      "objective/train/value_max": -0.0002359151840209961,
+      "objective/train/value_min": -0.483154296875,
+      "objective/train/value_reward_corr": 0.6927815042926995,
+      "objective/train/value_std": 0.0133209228515625,
+      "objective/train/weight_avg": 1.0013376474380493,
+      "objective/train/weighted_lm_loss": 1.73981511592865,
+      "objective/train/weights_max": 1.3604191541671753,
+      "objective/train/weights_min": 0.38016000390052795,
+      "theoretical_loss": 3.9032897494913876,
+      "tokens_seen": 550502400
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008415984593163216,
+      "loss": 0.0813,
+      "theoretical_loss": 3.9032897494913876,
+      "tokens_seen": 550502400
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008415182153747393,
+      "loss": 0.0812,
+      "theoretical_loss": 3.9030943522040946,
+      "tokens_seen": 550764544
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008414379714331568,
+      "loss": 0.0848,
+      "theoretical_loss": 3.902899073923166,
+      "tokens_seen": 551026688
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008413577274915744,
+      "loss": 0.0828,
+      "theoretical_loss": 3.902703914519557,
+      "tokens_seen": 551288832
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000841277483549992,
+      "loss": 0.083,
+      "theoretical_loss": 3.9025088738644236,
+      "tokens_seen": 551550976
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008411972396084095,
+      "loss": 0.0801,
+      "theoretical_loss": 3.9023139518291243,
+      "tokens_seen": 551813120
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008411169956668272,
+      "loss": 0.0809,
+      "theoretical_loss": 3.902119148285216,
+      "tokens_seen": 552075264
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008410367517252447,
+      "loss": 0.0828,
+      "theoretical_loss": 3.9019244631044563,
+      "tokens_seen": 552337408
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008409565077836624,
+      "loss": 0.0801,
+      "theoretical_loss": 3.9017298961588027,
+      "tokens_seen": 552599552
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.00084087626384208,
+      "loss": 0.0819,
+      "theoretical_loss": 3.901535447320412,
+      "tokens_seen": 552861696
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008407960199004975,
+      "loss": 0.0812,
+      "theoretical_loss": 3.901341116461639,
+      "tokens_seen": 553123840
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008407157759589151,
+      "loss": 0.0807,
+      "theoretical_loss": 3.9011469034550372,
+      "tokens_seen": 553385984
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008406355320173327,
+      "loss": 0.0819,
+      "theoretical_loss": 3.900952808173358,
+      "tokens_seen": 553648128
+    },
+    {
+      "epoch": 0.17,
+      "objective/train/advantage_avg": -0.00015300983795896173,
+      "objective/train/docs_used": 207589,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8608996868133545,
+      "objective/train/original_loss": 1.8608994483947754,
+      "objective/train/theoretical_loss": 3.9008558046396535,
+      "objective/train/tokens_used": 574239200,
+      "objective/train/value_avg": -0.0084075927734375,
+      "objective/train/value_loss": 0.0002866761351469904,
+      "objective/train/value_max": -0.00022876262664794922,
+      "objective/train/value_min": -0.2509765625,
+      "objective/train/value_reward_corr": 0.6109745044269907,
+      "objective/train/value_std": 0.01207733154296875,
+      "objective/train/weight_avg": 0.999976634979248,
+      "objective/train/weighted_lm_loss": 1.8602499961853027,
+      "objective/train/weights_max": 1.1194891929626465,
+      "objective/train/weights_min": 0.36891940236091614,
+      "theoretical_loss": 3.9008558046396535,
+      "tokens_seen": 553779200
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008405552880757503,
+      "loss": 0.0839,
+      "theoretical_loss": 3.900758830489551,
+      "tokens_seen": 553910272
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008404750441341678,
+      "loss": 0.083,
+      "theoretical_loss": 3.900564970276762,
+      "tokens_seen": 554172416
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008403948001925855,
+      "loss": 0.0812,
+      "theoretical_loss": 3.9003712274083346,
+      "tokens_seen": 554434560
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000840314556251003,
+      "loss": 0.0826,
+      "theoretical_loss": 3.9001776017578074,
+      "tokens_seen": 554696704
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008402343123094207,
+      "loss": 0.0817,
+      "theoretical_loss": 3.899984093198916,
+      "tokens_seen": 554958848
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008401540683678383,
+      "loss": 0.0802,
+      "theoretical_loss": 3.899790701605592,
+      "tokens_seen": 555220992
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008400738244262558,
+      "loss": 0.081,
+      "theoretical_loss": 3.899597426851961,
+      "tokens_seen": 555483136
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008399935804846735,
+      "loss": 0.0781,
+      "theoretical_loss": 3.899404268812343,
+      "tokens_seen": 555745280
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000839913336543091,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8992112273612545,
+      "tokens_seen": 556007424
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008398330926015086,
+      "loss": 0.0824,
+      "theoretical_loss": 3.8990183023734044,
+      "tokens_seen": 556269568
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008397528486599262,
+      "loss": 0.0818,
+      "theoretical_loss": 3.8988254937236952,
+      "tokens_seen": 556531712
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008396726047183438,
+      "loss": 0.0819,
+      "theoretical_loss": 3.8986328012872233,
+      "tokens_seen": 556793856
+    },
+    {
+      "epoch": 0.17,
+      "objective/train/advantage_avg": 0.0018912331433966756,
+      "objective/train/docs_used": 208740,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6763869524002075,
+      "objective/train/original_loss": 1.6763869524002075,
+      "objective/train/theoretical_loss": 3.8984402249392778,
+      "objective/train/tokens_used": 577516000,
+      "objective/train/value_avg": -0.00763702392578125,
+      "objective/train/value_loss": 0.00018121585890185088,
+      "objective/train/value_max": -0.00034880638122558594,
+      "objective/train/value_min": -0.187255859375,
+      "objective/train/value_reward_corr": 0.5036202731375501,
+      "objective/train/value_std": 0.00873565673828125,
+      "objective/train/weight_avg": 1.0019742250442505,
+      "objective/train/weighted_lm_loss": 1.6807695627212524,
+      "objective/train/weights_max": 1.0873464345932007,
+      "objective/train/weights_min": 0.368623286485672,
+      "theoretical_loss": 3.8984402249392778,
+      "tokens_seen": 557056000
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008395923607767613,
+      "loss": 0.0794,
+      "theoretical_loss": 3.8984402249392778,
+      "tokens_seen": 557056000
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000839512116835179,
+      "loss": 0.0784,
+      "theoretical_loss": 3.8982477645553395,
+      "tokens_seen": 557318144
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008394318728935966,
+      "loss": 0.0799,
+      "theoretical_loss": 3.898055420011082,
+      "tokens_seen": 557580288
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008393516289520141,
+      "loss": 0.0782,
+      "theoretical_loss": 3.8978631911823705,
+      "tokens_seen": 557842432
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008392713850104318,
+      "loss": 0.0802,
+      "theoretical_loss": 3.8976710779452612,
+      "tokens_seen": 558104576
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008391911410688493,
+      "loss": 0.0815,
+      "theoretical_loss": 3.8974790801760015,
+      "tokens_seen": 558366720
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008391108971272669,
+      "loss": 0.0841,
+      "theoretical_loss": 3.897287197751029,
+      "tokens_seen": 558628864
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008390306531856845,
+      "loss": 0.0831,
+      "theoretical_loss": 3.897095430546971,
+      "tokens_seen": 558891008
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000838950409244102,
+      "loss": 0.0834,
+      "theoretical_loss": 3.896903778440646,
+      "tokens_seen": 559153152
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008388701653025197,
+      "loss": 0.0823,
+      "theoretical_loss": 3.896712241309061,
+      "tokens_seen": 559415296
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008387899213609372,
+      "loss": 0.0833,
+      "theoretical_loss": 3.896520819029411,
+      "tokens_seen": 559677440
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008387096774193549,
+      "loss": 0.082,
+      "theoretical_loss": 3.896329511479082,
+      "tokens_seen": 559939584
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008386294334777725,
+      "loss": 0.0792,
+      "theoretical_loss": 3.8961383185356455,
+      "tokens_seen": 560201728
+    },
+    {
+      "epoch": 0.17,
+      "objective/train/advantage_avg": 0.0007566718268208206,
+      "objective/train/docs_used": 209977,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.573256254196167,
+      "objective/train/original_loss": 1.573256015777588,
+      "objective/train/theoretical_loss": 3.8960427650033047,
+      "objective/train/tokens_used": 580792800,
+      "objective/train/value_avg": -0.00795745849609375,
+      "objective/train/value_loss": 0.00029120329418219626,
+      "objective/train/value_max": -0.00023055076599121094,
+      "objective/train/value_min": -0.6328125,
+      "objective/train/value_reward_corr": 0.6374020027425222,
+      "objective/train/value_std": 0.012054443359375,
+      "objective/train/weight_avg": 1.0008906126022339,
+      "objective/train/weighted_lm_loss": 1.5738105773925781,
+      "objective/train/weights_max": 1.2740330696105957,
+      "objective/train/weights_min": 0.3893653154373169,
+      "theoretical_loss": 3.8960427650033047,
+      "tokens_seen": 560332800
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008385491895361901,
+      "loss": 0.0794,
+      "theoretical_loss": 3.895947240076862,
+      "tokens_seen": 560463872
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008384689455946076,
+      "loss": 0.0784,
+      "theoretical_loss": 3.895756275980681,
+      "tokens_seen": 560726016
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008383887016530252,
+      "loss": 0.0817,
+      "theoretical_loss": 3.895565426125237,
+      "tokens_seen": 560988160
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008383084577114428,
+      "loss": 0.0812,
+      "theoretical_loss": 3.8953746903888513,
+      "tokens_seen": 561250304
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008382282137698603,
+      "loss": 0.0787,
+      "theoretical_loss": 3.895184068650033,
+      "tokens_seen": 561512448
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000838147969828278,
+      "loss": 0.0817,
+      "theoretical_loss": 3.8949935607874764,
+      "tokens_seen": 561774592
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008380677258866955,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8948031666800613,
+      "tokens_seen": 562036736
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008379874819451132,
+      "loss": 0.0838,
+      "theoretical_loss": 3.8946128862068528,
+      "tokens_seen": 562298880
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008379072380035308,
+      "loss": 0.0804,
+      "theoretical_loss": 3.8944227192471006,
+      "tokens_seen": 562561024
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008378269940619483,
+      "loss": 0.082,
+      "theoretical_loss": 3.8942326656802395,
+      "tokens_seen": 562823168
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000837746750120366,
+      "loss": 0.0829,
+      "theoretical_loss": 3.894042725385888,
+      "tokens_seen": 563085312
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008376665061787835,
+      "loss": 0.0837,
+      "theoretical_loss": 3.893852898243849,
+      "tokens_seen": 563347456
+    },
+    {
+      "epoch": 0.17,
+      "objective/train/advantage_avg": 0.0011895080097019672,
+      "objective/train/docs_used": 211161,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.739345908164978,
+      "objective/train/original_loss": 1.7393460273742676,
+      "objective/train/theoretical_loss": 3.8936631841341076,
+      "objective/train/tokens_used": 584069600,
+      "objective/train/value_avg": -0.009521484375,
+      "objective/train/value_loss": 0.000342073617503047,
+      "objective/train/value_max": -0.00018525123596191406,
+      "objective/train/value_min": -0.6494140625,
+      "objective/train/value_reward_corr": 0.6713985854388784,
+      "objective/train/value_std": 0.0169830322265625,
+      "objective/train/weight_avg": 1.0013439655303955,
+      "objective/train/weighted_lm_loss": 1.7415797710418701,
+      "objective/train/weights_max": 1.3866688013076782,
+      "objective/train/weights_min": 0.3763507008552551,
+      "theoretical_loss": 3.8936631841341076,
+      "tokens_seen": 563609600
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008375862622372011,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8936631841341076,
+      "tokens_seen": 563609600
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008375060182956187,
+      "loss": 0.0814,
+      "theoretical_loss": 3.893473582936833,
+      "tokens_seen": 563871744
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008374257743540363,
+      "loss": 0.083,
+      "theoretical_loss": 3.8932840945323774,
+      "tokens_seen": 564133888
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008373455304124538,
+      "loss": 0.0809,
+      "theoretical_loss": 3.8930947188012737,
+      "tokens_seen": 564396032
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008372652864708715,
+      "loss": 0.0784,
+      "theoretical_loss": 3.8929054556242377,
+      "tokens_seen": 564658176
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008371850425292891,
+      "loss": 0.0789,
+      "theoretical_loss": 3.892716304882167,
+      "tokens_seen": 564920320
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008371047985877066,
+      "loss": 0.0815,
+      "theoretical_loss": 3.892527266456141,
+      "tokens_seen": 565182464
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008370245546461243,
+      "loss": 0.0829,
+      "theoretical_loss": 3.8923383402274174,
+      "tokens_seen": 565444608
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008369443107045418,
+      "loss": 0.0822,
+      "theoretical_loss": 3.8921495260774375,
+      "tokens_seen": 565706752
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008368640667629594,
+      "loss": 0.0811,
+      "theoretical_loss": 3.8919608238878216,
+      "tokens_seen": 565968896
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000836783822821377,
+      "loss": 0.0774,
+      "theoretical_loss": 3.891772233540369,
+      "tokens_seen": 566231040
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008367035788797946,
+      "loss": 0.08,
+      "theoretical_loss": 3.8915837549170584,
+      "tokens_seen": 566493184
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008366233349382121,
+      "loss": 0.0794,
+      "theoretical_loss": 3.89139538790005,
+      "tokens_seen": 566755328
+    },
+    {
+      "epoch": 0.17,
+      "objective/train/advantage_avg": 0.0005871827597729862,
+      "objective/train/docs_used": 212374,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6017706394195557,
+      "objective/train/original_loss": 1.6017706394195557,
+      "objective/train/theoretical_loss": 3.8913012462071315,
+      "objective/train/tokens_used": 587346400,
+      "objective/train/value_avg": -0.00801849365234375,
+      "objective/train/value_loss": 0.00017772662977222353,
+      "objective/train/value_max": -0.00031757354736328125,
+      "objective/train/value_min": -0.368408203125,
+      "objective/train/value_reward_corr": 0.6835943969490854,
+      "objective/train/value_std": 0.0125732421875,
+      "objective/train/weight_avg": 1.0006738901138306,
+      "objective/train/weighted_lm_loss": 1.6030828952789307,
+      "objective/train/weights_max": 1.1462258100509644,
+      "objective/train/weights_min": 0.7185884118080139,
+      "theoretical_loss": 3.8913012462071315,
+      "tokens_seen": 566886400
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008365430909966297,
+      "loss": 0.0827,
+      "theoretical_loss": 3.8912071323716795,
+      "tokens_seen": 567017472
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008364628470550474,
+      "loss": 0.0814,
+      "theoretical_loss": 3.8910189882144626,
+      "tokens_seen": 567279616
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000836382603113465,
+      "loss": 0.0837,
+      "theoretical_loss": 3.8908309553110936,
+      "tokens_seen": 567541760
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008363023591718826,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8906430335444426,
+      "tokens_seen": 567803904
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008362221152303001,
+      "loss": 0.0813,
+      "theoretical_loss": 3.8904552227975593,
+      "tokens_seen": 568066048
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008361418712887178,
+      "loss": 0.082,
+      "theoretical_loss": 3.8902675229536685,
+      "tokens_seen": 568328192
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008360616273471353,
+      "loss": 0.0825,
+      "theoretical_loss": 3.8900799338961725,
+      "tokens_seen": 568590336
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008359813834055528,
+      "loss": 0.0832,
+      "theoretical_loss": 3.8898924555086496,
+      "tokens_seen": 568852480
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008359011394639705,
+      "loss": 0.0811,
+      "theoretical_loss": 3.8897050876748542,
+      "tokens_seen": 569114624
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000835820895522388,
+      "loss": 0.0825,
+      "theoretical_loss": 3.8895178302787166,
+      "tokens_seen": 569376768
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008357406515808057,
+      "loss": 0.0817,
+      "theoretical_loss": 3.8893306832043404,
+      "tokens_seen": 569638912
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008356604076392233,
+      "loss": 0.0835,
+      "theoretical_loss": 3.8891436463360076,
+      "tokens_seen": 569901056
+    },
+    {
+      "epoch": 0.17,
+      "objective/train/advantage_avg": -0.0002838138898368925,
+      "objective/train/docs_used": 213119,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5427865982055664,
+      "objective/train/original_loss": 1.5427868366241455,
+      "objective/train/theoretical_loss": 3.8889567195581716,
+      "objective/train/tokens_used": 590623200,
+      "objective/train/value_avg": -0.007099151611328125,
+      "objective/train/value_loss": 0.0001902274671010673,
+      "objective/train/value_max": -0.00019109249114990234,
+      "objective/train/value_min": -0.2220458984375,
+      "objective/train/value_reward_corr": 0.6440366057374965,
+      "objective/train/value_std": 0.00963592529296875,
+      "objective/train/weight_avg": 0.9998065829277039,
+      "objective/train/weighted_lm_loss": 1.543387532234192,
+      "objective/train/weights_max": 1.1371427774429321,
+      "objective/train/weights_min": 0.6790033578872681,
+      "theoretical_loss": 3.8889567195581716,
+      "tokens_seen": 570163200
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008355801636976409,
+      "loss": 0.082,
+      "theoretical_loss": 3.8889567195581716,
+      "tokens_seen": 570163200
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008354999197560584,
+      "loss": 0.0789,
+      "theoretical_loss": 3.8887699027554614,
+      "tokens_seen": 570425344
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000835419675814476,
+      "loss": 0.0805,
+      "theoretical_loss": 3.8885831958126786,
+      "tokens_seen": 570687488
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008353394318728936,
+      "loss": 0.0826,
+      "theoretical_loss": 3.8883965986148015,
+      "tokens_seen": 570949632
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008352591879313111,
+      "loss": 0.0792,
+      "theoretical_loss": 3.888210111046978,
+      "tokens_seen": 571211776
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008351789439897288,
+      "loss": 0.0824,
+      "theoretical_loss": 3.8880237329945295,
+      "tokens_seen": 571473920
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008350987000481463,
+      "loss": 0.0839,
+      "theoretical_loss": 3.887837464342952,
+      "tokens_seen": 571736064
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008350184561065641,
+      "loss": 0.0831,
+      "theoretical_loss": 3.8876513049779113,
+      "tokens_seen": 571998208
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008349382121649816,
+      "loss": 0.083,
+      "theoretical_loss": 3.887465254785246,
+      "tokens_seen": 572260352
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008348579682233991,
+      "loss": 0.0838,
+      "theoretical_loss": 3.887279313650967,
+      "tokens_seen": 572522496
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008347777242818168,
+      "loss": 0.0834,
+      "theoretical_loss": 3.8870934814612546,
+      "tokens_seen": 572784640
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008346974803402343,
+      "loss": 0.0786,
+      "theoretical_loss": 3.886907758102461,
+      "tokens_seen": 573046784
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008346172363986519,
+      "loss": 0.0831,
+      "theoretical_loss": 3.8867221434611094,
+      "tokens_seen": 573308928
+    },
+    {
+      "epoch": 0.17,
+      "objective/train/advantage_avg": 0.00023172213695943356,
+      "objective/train/docs_used": 214228,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.640936255455017,
+      "objective/train/original_loss": 1.640936255455017,
+      "objective/train/theoretical_loss": 3.8866293768740587,
+      "objective/train/tokens_used": 593900000,
+      "objective/train/value_avg": -0.00962066650390625,
+      "objective/train/value_loss": 0.00022791478841099888,
+      "objective/train/value_max": -0.0002434253692626953,
+      "objective/train/value_min": -0.2286376953125,
+      "objective/train/value_reward_corr": 0.7870548857924071,
+      "objective/train/value_std": 0.0165863037109375,
+      "objective/train/weight_avg": 1.0003434419631958,
+      "objective/train/weighted_lm_loss": 1.641239881515503,
+      "objective/train/weights_max": 1.2273142337799072,
+      "objective/train/weights_min": 0.6072003245353699,
+      "theoretical_loss": 3.8866293768740587,
+      "tokens_seen": 573440000
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008345369924570695,
+      "loss": 0.0818,
+      "theoretical_loss": 3.8865366374238914,
+      "tokens_seen": 573571072
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008344567485154871,
+      "loss": 0.0821,
+      "theoretical_loss": 3.88635123987767,
+      "tokens_seen": 573833216
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008343765045739046,
+      "loss": 0.0825,
+      "theoretical_loss": 3.8861659507094766,
+      "tokens_seen": 574095360
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008342962606323223,
+      "loss": 0.0816,
+      "theoretical_loss": 3.885980769806513,
+      "tokens_seen": 574357504
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008342160166907399,
+      "loss": 0.0846,
+      "theoretical_loss": 3.8857956970561487,
+      "tokens_seen": 574619648
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008341357727491574,
+      "loss": 0.0788,
+      "theoretical_loss": 3.8856107323459215,
+      "tokens_seen": 574881792
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008340555288075751,
+      "loss": 0.0822,
+      "theoretical_loss": 3.8854258755635387,
+      "tokens_seen": 575143936
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008339752848659926,
+      "loss": 0.0831,
+      "theoretical_loss": 3.885241126596874,
+      "tokens_seen": 575406080
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008338950409244103,
+      "loss": 0.0803,
+      "theoretical_loss": 3.885056485333969,
+      "tokens_seen": 575668224
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008338147969828278,
+      "loss": 0.083,
+      "theoretical_loss": 3.884871951663034,
+      "tokens_seen": 575930368
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008337345530412454,
+      "loss": 0.078,
+      "theoretical_loss": 3.8846875254724442,
+      "tokens_seen": 576192512
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.000833654309099663,
+      "loss": 0.08,
+      "theoretical_loss": 3.8845032066507414,
+      "tokens_seen": 576454656
+    },
+    {
+      "epoch": 0.17,
+      "objective/train/advantage_avg": 0.0014360037166625261,
+      "objective/train/docs_used": 215374,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6849347352981567,
+      "objective/train/original_loss": 1.6849347352981567,
+      "objective/train/theoretical_loss": 3.884318995086635,
+      "objective/train/tokens_used": 597176800,
+      "objective/train/value_avg": -0.00965118408203125,
+      "objective/train/value_loss": 0.00028802509768866,
+      "objective/train/value_max": -0.0003101825714111328,
+      "objective/train/value_min": -0.41748046875,
+      "objective/train/value_reward_corr": 0.6993208086391764,
+      "objective/train/value_std": 0.015655517578125,
+      "objective/train/weight_avg": 1.0015720129013062,
+      "objective/train/weighted_lm_loss": 1.6866198778152466,
+      "objective/train/weights_max": 1.1339787244796753,
+      "objective/train/weights_min": 0.41069987416267395,
+      "theoretical_loss": 3.884318995086635,
+      "tokens_seen": 576716800
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008335740651580805,
+      "loss": 0.0794,
+      "theoretical_loss": 3.884318995086635,
+      "tokens_seen": 576716800
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008334938212164982,
+      "loss": 0.0816,
+      "theoretical_loss": 3.8841348906689985,
+      "tokens_seen": 576978944
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 0.0008334135772749158,
+      "loss": 0.0831,
+      "theoretical_loss": 3.8839508932868725,
+      "tokens_seen": 577241088
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008333333333333334,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8837670028294626,
+      "tokens_seen": 577503232
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008332530893917509,
+      "loss": 0.0822,
+      "theoretical_loss": 3.883583219186138,
+      "tokens_seen": 577765376
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008331728454501686,
+      "loss": 0.0824,
+      "theoretical_loss": 3.8833995422464342,
+      "tokens_seen": 578027520
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008330926015085861,
+      "loss": 0.0821,
+      "theoretical_loss": 3.88321597190005,
+      "tokens_seen": 578289664
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008330123575670036,
+      "loss": 0.0824,
+      "theoretical_loss": 3.883032508036848,
+      "tokens_seen": 578551808
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008329321136254213,
+      "loss": 0.0797,
+      "theoretical_loss": 3.882849150546856,
+      "tokens_seen": 578813952
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008328518696838388,
+      "loss": 0.0833,
+      "theoretical_loss": 3.8826658993202625,
+      "tokens_seen": 579076096
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008327716257422565,
+      "loss": 0.0817,
+      "theoretical_loss": 3.8824827542474214,
+      "tokens_seen": 579338240
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008326913818006741,
+      "loss": 0.0822,
+      "theoretical_loss": 3.882299715218848,
+      "tokens_seen": 579600384
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008326111378590917,
+      "loss": 0.0817,
+      "theoretical_loss": 3.8821167821252196,
+      "tokens_seen": 579862528
+    },
+    {
+      "epoch": 0.18,
+      "objective/train/advantage_avg": 0.0010556569322943687,
+      "objective/train/docs_used": 216668,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7441470623016357,
+      "objective/train/original_loss": 1.7441470623016357,
+      "objective/train/theoretical_loss": 3.8820253552698913,
+      "objective/train/tokens_used": 600453600,
+      "objective/train/value_avg": -0.01125335693359375,
+      "objective/train/value_loss": 0.0008476045331917703,
+      "objective/train/value_max": -0.00021660327911376953,
+      "objective/train/value_min": -0.91845703125,
+      "objective/train/value_reward_corr": 0.5798144182694726,
+      "objective/train/value_std": 0.020416259765625,
+      "objective/train/weight_avg": 1.0014194250106812,
+      "objective/train/weighted_lm_loss": 1.7468537092208862,
+      "objective/train/weights_max": 1.7236318588256836,
+      "objective/train/weights_min": 0.3710477352142334,
+      "theoretical_loss": 3.8820253552698913,
+      "tokens_seen": 579993600
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008325308939175093,
+      "loss": 0.083,
+      "theoretical_loss": 3.8819339548573772,
+      "tokens_seen": 580124672
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008324506499759268,
+      "loss": 0.0819,
+      "theoretical_loss": 3.881751233306322,
+      "tokens_seen": 580386816
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008323704060343444,
+      "loss": 0.0818,
+      "theoretical_loss": 3.881568617363218,
+      "tokens_seen": 580648960
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000832290162092762,
+      "loss": 0.0827,
+      "theoretical_loss": 3.881386106919389,
+      "tokens_seen": 580911104
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008322099181511796,
+      "loss": 0.083,
+      "theoretical_loss": 3.88120370186632,
+      "tokens_seen": 581173248
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008321296742095971,
+      "loss": 0.0806,
+      "theoretical_loss": 3.881021402095657,
+      "tokens_seen": 581435392
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008320494302680149,
+      "loss": 0.081,
+      "theoretical_loss": 3.880839207499205,
+      "tokens_seen": 581697536
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008319691863264324,
+      "loss": 0.0823,
+      "theoretical_loss": 3.880657117968931,
+      "tokens_seen": 581959680
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008318889423848499,
+      "loss": 0.0846,
+      "theoretical_loss": 3.880475133396959,
+      "tokens_seen": 582221824
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008318086984432676,
+      "loss": 0.0817,
+      "theoretical_loss": 3.8802932536755748,
+      "tokens_seen": 582483968
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008317284545016851,
+      "loss": 0.0835,
+      "theoretical_loss": 3.880111478697221,
+      "tokens_seen": 582746112
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008316482105601027,
+      "loss": 0.0851,
+      "theoretical_loss": 3.8799298083545004,
+      "tokens_seen": 583008256
+    },
+    {
+      "epoch": 0.18,
+      "objective/train/advantage_avg": 0.0006918111466802657,
+      "objective/train/docs_used": 217803,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7163397073745728,
+      "objective/train/original_loss": 1.7163399457931519,
+      "objective/train/theoretical_loss": 3.879748242540173,
+      "objective/train/tokens_used": 603730400,
+      "objective/train/value_avg": -0.0118408203125,
+      "objective/train/value_loss": 0.00036437122616916895,
+      "objective/train/value_max": -0.0002892017364501953,
+      "objective/train/value_min": -0.70654296875,
+      "objective/train/value_reward_corr": 0.6686834210191256,
+      "objective/train/value_std": 0.0202178955078125,
+      "objective/train/weight_avg": 1.0008691549301147,
+      "objective/train/weighted_lm_loss": 1.7178725004196167,
+      "objective/train/weights_max": 1.8649041652679443,
+      "objective/train/weights_min": 0.3742033541202545,
+      "theoretical_loss": 3.879748242540173,
+      "tokens_seen": 583270400
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008315679666185203,
+      "loss": 0.0817,
+      "theoretical_loss": 3.879748242540173,
+      "tokens_seen": 583270400
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008314877226769379,
+      "loss": 0.084,
+      "theoretical_loss": 3.8795667811471573,
+      "tokens_seen": 583532544
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008314074787353555,
+      "loss": 0.0824,
+      "theoretical_loss": 3.8793854240685306,
+      "tokens_seen": 583794688
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000831327234793773,
+      "loss": 0.0821,
+      "theoretical_loss": 3.879204171197525,
+      "tokens_seen": 584056832
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008312469908521907,
+      "loss": 0.0813,
+      "theoretical_loss": 3.879023022427533,
+      "tokens_seen": 584318976
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008311667469106083,
+      "loss": 0.0818,
+      "theoretical_loss": 3.878841977652101,
+      "tokens_seen": 584581120
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008310865029690259,
+      "loss": 0.0819,
+      "theoretical_loss": 3.8786610367649343,
+      "tokens_seen": 584843264
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008310062590274434,
+      "loss": 0.0829,
+      "theoretical_loss": 3.8784801996598928,
+      "tokens_seen": 585105408
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008309260150858611,
+      "loss": 0.0824,
+      "theoretical_loss": 3.878299466230992,
+      "tokens_seen": 585367552
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008308457711442786,
+      "loss": 0.0813,
+      "theoretical_loss": 3.8781188363724057,
+      "tokens_seen": 585629696
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008307655272026961,
+      "loss": 0.0813,
+      "theoretical_loss": 3.87793830997846,
+      "tokens_seen": 585891840
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008306852832611138,
+      "loss": 0.0807,
+      "theoretical_loss": 3.8777578869436384,
+      "tokens_seen": 586153984
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008306050393195313,
+      "loss": 0.0822,
+      "theoretical_loss": 3.8775775671625765,
+      "tokens_seen": 586416128
+    },
+    {
+      "epoch": 0.18,
+      "objective/train/advantage_avg": 0.0013700941344723105,
+      "objective/train/docs_used": 218927,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.789874792098999,
+      "objective/train/original_loss": 1.789874792098999,
+      "objective/train/theoretical_loss": 3.8774874459593223,
+      "objective/train/tokens_used": 607007200,
+      "objective/train/value_avg": -0.00925445556640625,
+      "objective/train/value_loss": 0.0004925947869196534,
+      "objective/train/value_max": -0.0002779960632324219,
+      "objective/train/value_min": -0.70556640625,
+      "objective/train/value_reward_corr": 0.5972247492646405,
+      "objective/train/value_std": 0.0181732177734375,
+      "objective/train/weight_avg": 1.0015876293182373,
+      "objective/train/weighted_lm_loss": 1.7908934354782104,
+      "objective/train/weights_max": 1.9563560485839844,
+      "objective/train/weights_min": 0.3702559173107147,
+      "theoretical_loss": 3.8774874459593223,
+      "tokens_seen": 586547200
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000830524795377949,
+      "loss": 0.0824,
+      "theoretical_loss": 3.8773973505300674,
+      "tokens_seen": 586678272
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008304445514363666,
+      "loss": 0.0803,
+      "theoretical_loss": 3.877217236941055,
+      "tokens_seen": 586940416
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008303643074947842,
+      "loss": 0.0836,
+      "theoretical_loss": 3.877037226290641,
+      "tokens_seen": 587202560
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008302840635532017,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8768573184740767,
+      "tokens_seen": 587464704
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008302038196116194,
+      "loss": 0.0806,
+      "theoretical_loss": 3.87667751338677,
+      "tokens_seen": 587726848
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008301235756700369,
+      "loss": 0.0803,
+      "theoretical_loss": 3.8764978109242794,
+      "tokens_seen": 587988992
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008300433317284545,
+      "loss": 0.0822,
+      "theoretical_loss": 3.8763182109823173,
+      "tokens_seen": 588251136
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008299630877868721,
+      "loss": 0.0763,
+      "theoretical_loss": 3.8761387134567475,
+      "tokens_seen": 588513280
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008298828438452896,
+      "loss": 0.0813,
+      "theoretical_loss": 3.8759593182435874,
+      "tokens_seen": 588775424
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008298025999037074,
+      "loss": 0.0849,
+      "theoretical_loss": 3.875780025239005,
+      "tokens_seen": 589037568
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008297223559621249,
+      "loss": 0.0797,
+      "theoretical_loss": 3.8756008343393202,
+      "tokens_seen": 589299712
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008296421120205425,
+      "loss": 0.0835,
+      "theoretical_loss": 3.8754217454410043,
+      "tokens_seen": 589561856
+    },
+    {
+      "epoch": 0.18,
+      "objective/train/advantage_avg": -6.183000368764624e-05,
+      "objective/train/docs_used": 220165,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6896483898162842,
+      "objective/train/original_loss": 1.6896482706069946,
+      "objective/train/theoretical_loss": 3.875242758440679,
+      "objective/train/tokens_used": 610284000,
+      "objective/train/value_avg": -0.0089569091796875,
+      "objective/train/value_loss": 0.00046085796202532947,
+      "objective/train/value_max": -0.0002415180206298828,
+      "objective/train/value_min": -0.673828125,
+      "objective/train/value_reward_corr": 0.5928429396910613,
+      "objective/train/value_std": 0.015838623046875,
+      "objective/train/weight_avg": 1.0001388788223267,
+      "objective/train/weighted_lm_loss": 1.6889548301696777,
+      "objective/train/weights_max": 1.2902297973632812,
+      "objective/train/weights_min": 0.36886030435562134,
+      "theoretical_loss": 3.875242758440679,
+      "tokens_seen": 589824000
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008295618680789601,
+      "loss": 0.0821,
+      "theoretical_loss": 3.875242758440679,
+      "tokens_seen": 589824000
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008294816241373776,
+      "loss": 0.0809,
+      "theoretical_loss": 3.875063873235117,
+      "tokens_seen": 590086144
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008294013801957952,
+      "loss": 0.0854,
+      "theoretical_loss": 3.874885089721242,
+      "tokens_seen": 590348288
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008293211362542128,
+      "loss": 0.0782,
+      "theoretical_loss": 3.8747064077961264,
+      "tokens_seen": 590610432
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008292408923126304,
+      "loss": 0.079,
+      "theoretical_loss": 3.874527827356994,
+      "tokens_seen": 590872576
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008291606483710479,
+      "loss": 0.0832,
+      "theoretical_loss": 3.8743493483012172,
+      "tokens_seen": 591134720
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008290804044294657,
+      "loss": 0.0806,
+      "theoretical_loss": 3.874170970526317,
+      "tokens_seen": 591396864
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008290001604878832,
+      "loss": 0.0826,
+      "theoretical_loss": 3.873992693929965,
+      "tokens_seen": 591659008
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008289199165463007,
+      "loss": 0.0782,
+      "theoretical_loss": 3.8738145184099797,
+      "tokens_seen": 591921152
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008288396726047184,
+      "loss": 0.0798,
+      "theoretical_loss": 3.8736364438643296,
+      "tokens_seen": 592183296
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008287594286631359,
+      "loss": 0.0833,
+      "theoretical_loss": 3.87345847019113,
+      "tokens_seen": 592445440
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008286791847215536,
+      "loss": 0.079,
+      "theoretical_loss": 3.8732805972886446,
+      "tokens_seen": 592707584
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008285989407799711,
+      "loss": 0.0794,
+      "theoretical_loss": 3.873102825055285,
+      "tokens_seen": 592969728
+    },
+    {
+      "epoch": 0.18,
+      "objective/train/advantage_avg": 0.0022287839092314243,
+      "objective/train/docs_used": 221319,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6529507637023926,
+      "objective/train/original_loss": 1.6529508829116821,
+      "objective/train/theoretical_loss": 3.873013976657821,
+      "objective/train/tokens_used": 613560800,
+      "objective/train/value_avg": -0.01081085205078125,
+      "objective/train/value_loss": 0.0005324392695911229,
+      "objective/train/value_max": -0.00033926963806152344,
+      "objective/train/value_min": -0.43798828125,
+      "objective/train/value_reward_corr": 0.6031346357258855,
+      "objective/train/value_std": 0.0146331787109375,
+      "objective/train/weight_avg": 1.0024420022964478,
+      "objective/train/weighted_lm_loss": 1.6559507846832275,
+      "objective/train/weights_max": 1.158463716506958,
+      "objective/train/weights_min": 0.2581728398799896,
+      "theoretical_loss": 3.873013976657821,
+      "tokens_seen": 593100800
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008285186968383887,
+      "loss": 0.0797,
+      "theoretical_loss": 3.87292515338961,
+      "tokens_seen": 593231872
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008284384528968063,
+      "loss": 0.0804,
+      "theoretical_loss": 3.872747582190324,
+      "tokens_seen": 593494016
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008283582089552239,
+      "loss": 0.0818,
+      "theoretical_loss": 3.8725701113562794,
+      "tokens_seen": 593756160
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008282779650136415,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8723927407864758,
+      "tokens_seen": 594018304
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008281977210720591,
+      "loss": 0.0818,
+      "theoretical_loss": 3.8722154703800573,
+      "tokens_seen": 594280448
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008281174771304767,
+      "loss": 0.0825,
+      "theoretical_loss": 3.8720383000363148,
+      "tokens_seen": 594542592
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008280372331888942,
+      "loss": 0.081,
+      "theoretical_loss": 3.871861229654684,
+      "tokens_seen": 594804736
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008279569892473119,
+      "loss": 0.0779,
+      "theoretical_loss": 3.8716842591347476,
+      "tokens_seen": 595066880
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008278767453057294,
+      "loss": 0.0822,
+      "theoretical_loss": 3.871507388376231,
+      "tokens_seen": 595329024
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008277965013641469,
+      "loss": 0.0796,
+      "theoretical_loss": 3.871330617279006,
+      "tokens_seen": 595591168
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008277162574225646,
+      "loss": 0.0815,
+      "theoretical_loss": 3.8711539457430897,
+      "tokens_seen": 595853312
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008276360134809821,
+      "loss": 0.0797,
+      "theoretical_loss": 3.87097737366864,
+      "tokens_seen": 596115456
+    },
+    {
+      "epoch": 0.18,
+      "objective/train/advantage_avg": 0.00167937483638525,
+      "objective/train/docs_used": 222574,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.542468547821045,
+      "objective/train/original_loss": 1.5424683094024658,
+      "objective/train/theoretical_loss": 3.870800900955963,
+      "objective/train/tokens_used": 616837600,
+      "objective/train/value_avg": -0.0085296630859375,
+      "objective/train/value_loss": 0.0005001907702535391,
+      "objective/train/value_max": -0.00029587745666503906,
+      "objective/train/value_min": -0.86962890625,
+      "objective/train/value_reward_corr": 0.6588531620632475,
+      "objective/train/value_std": 0.0174713134765625,
+      "objective/train/weight_avg": 1.0018889904022217,
+      "objective/train/weighted_lm_loss": 1.5452805757522583,
+      "objective/train/weights_max": 1.3651901483535767,
+      "objective/train/weights_min": 0.37030255794525146,
+      "theoretical_loss": 3.870800900955963,
+      "tokens_seen": 596377600
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008275557695393999,
+      "loss": 0.0794,
+      "theoretical_loss": 3.870800900955963,
+      "tokens_seen": 596377600
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008274755255978174,
+      "loss": 0.0816,
+      "theoretical_loss": 3.8706245275055062,
+      "tokens_seen": 596639744
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000827395281656235,
+      "loss": 0.08,
+      "theoretical_loss": 3.8704482532178606,
+      "tokens_seen": 596901888
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008273150377146526,
+      "loss": 0.0839,
+      "theoretical_loss": 3.8702720779937607,
+      "tokens_seen": 597164032
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008272347937730702,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8700960017340833,
+      "tokens_seen": 597426176
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008271545498314877,
+      "loss": 0.08,
+      "theoretical_loss": 3.8699200243398493,
+      "tokens_seen": 597688320
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008270743058899053,
+      "loss": 0.0848,
+      "theoretical_loss": 3.8697441457122204,
+      "tokens_seen": 597950464
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008269940619483229,
+      "loss": 0.0823,
+      "theoretical_loss": 3.8695683657525013,
+      "tokens_seen": 598212608
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008269138180067404,
+      "loss": 0.0826,
+      "theoretical_loss": 3.8693926843621376,
+      "tokens_seen": 598474752
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008268335740651582,
+      "loss": 0.0805,
+      "theoretical_loss": 3.8692171014427177,
+      "tokens_seen": 598736896
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008267533301235757,
+      "loss": 0.0796,
+      "theoretical_loss": 3.86904161689597,
+      "tokens_seen": 598999040
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008266730861819933,
+      "loss": 0.0832,
+      "theoretical_loss": 3.868866230623766,
+      "tokens_seen": 599261184
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008265928422404109,
+      "loss": 0.0809,
+      "theoretical_loss": 3.8686909425281146,
+      "tokens_seen": 599523328
+    },
+    {
+      "epoch": 0.18,
+      "objective/train/advantage_avg": 0.0010342653840780258,
+      "objective/train/docs_used": 223680,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6975395679473877,
+      "objective/train/original_loss": 1.6975395679473877,
+      "objective/train/theoretical_loss": 3.868603335265913,
+      "objective/train/tokens_used": 620114400,
+      "objective/train/value_avg": -0.010101318359375,
+      "objective/train/value_loss": 0.00020030724408570677,
+      "objective/train/value_max": -0.00017404556274414062,
+      "objective/train/value_min": -0.347900390625,
+      "objective/train/value_reward_corr": 0.8020624486643968,
+      "objective/train/value_std": 0.0178985595703125,
+      "objective/train/weight_avg": 1.00112783908844,
+      "objective/train/weighted_lm_loss": 1.6995559930801392,
+      "objective/train/weights_max": 1.3582996129989624,
+      "objective/train/weights_min": 0.39774826169013977,
+      "theoretical_loss": 3.868603335265913,
+      "tokens_seen": 599654400
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008265125982988284,
+      "loss": 0.0814,
+      "theoretical_loss": 3.8685157525111684,
+      "tokens_seen": 599785472
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000826432354357246,
+      "loss": 0.0816,
+      "theoretical_loss": 3.8683406604752184,
+      "tokens_seen": 600047616
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008263521104156636,
+      "loss": 0.0829,
+      "theoretical_loss": 3.868165666322696,
+      "tokens_seen": 600309760
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008262718664740812,
+      "loss": 0.0828,
+      "theoretical_loss": 3.8679907699561733,
+      "tokens_seen": 600571904
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008261916225324988,
+      "loss": 0.0822,
+      "theoretical_loss": 3.86781597127836,
+      "tokens_seen": 600834048
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008261113785909165,
+      "loss": 0.0826,
+      "theoretical_loss": 3.867641270192107,
+      "tokens_seen": 601096192
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000826031134649334,
+      "loss": 0.0834,
+      "theoretical_loss": 3.867466666600402,
+      "tokens_seen": 601358336
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008259508907077516,
+      "loss": 0.0819,
+      "theoretical_loss": 3.867292160406373,
+      "tokens_seen": 601620480
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008258706467661692,
+      "loss": 0.0798,
+      "theoretical_loss": 3.8671177515132857,
+      "tokens_seen": 601882624
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008257904028245867,
+      "loss": 0.0824,
+      "theoretical_loss": 3.866943439824545,
+      "tokens_seen": 602144768
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008257101588830044,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8667692252436914,
+      "tokens_seen": 602406912
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008256299149414219,
+      "loss": 0.0822,
+      "theoretical_loss": 3.8665951076744056,
+      "tokens_seen": 602669056
+    },
+    {
+      "epoch": 0.18,
+      "objective/train/advantage_avg": -8.796821930445731e-05,
+      "objective/train/docs_used": 224736,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5687521696090698,
+      "objective/train/original_loss": 1.5687522888183594,
+      "objective/train/theoretical_loss": 3.866421087020504,
+      "objective/train/tokens_used": 623391200,
+      "objective/train/value_avg": -0.00591278076171875,
+      "objective/train/value_loss": 0.00038589219911955297,
+      "objective/train/value_max": -0.00026535987854003906,
+      "objective/train/value_min": -0.270263671875,
+      "objective/train/value_reward_corr": 0.5731155725865233,
+      "objective/train/value_std": 0.00867462158203125,
+      "objective/train/weight_avg": 1.0000841617584229,
+      "objective/train/weighted_lm_loss": 1.569719672203064,
+      "objective/train/weights_max": 1.1083186864852905,
+      "objective/train/weights_min": 0.36869922280311584,
+      "theoretical_loss": 3.866421087020504,
+      "tokens_seen": 602931200
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008255496709998395,
+      "loss": 0.0794,
+      "theoretical_loss": 3.866421087020504,
+      "tokens_seen": 602931200
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008254694270582571,
+      "loss": 0.0768,
+      "theoretical_loss": 3.8662471631859407,
+      "tokens_seen": 603193344
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008253891831166747,
+      "loss": 0.0821,
+      "theoretical_loss": 3.866073336074807,
+      "tokens_seen": 603455488
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008253089391750923,
+      "loss": 0.0787,
+      "theoretical_loss": 3.8658996055913297,
+      "tokens_seen": 603717632
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008252286952335099,
+      "loss": 0.0838,
+      "theoretical_loss": 3.8657259716398737,
+      "tokens_seen": 603979776
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008251484512919275,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8655524341249388,
+      "tokens_seen": 604241920
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000825068207350345,
+      "loss": 0.0827,
+      "theoretical_loss": 3.865378992951161,
+      "tokens_seen": 604504064
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008249879634087627,
+      "loss": 0.0812,
+      "theoretical_loss": 3.865205648023311,
+      "tokens_seen": 604766208
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008249077194671802,
+      "loss": 0.0821,
+      "theoretical_loss": 3.8650323992462963,
+      "tokens_seen": 605028352
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008248274755255978,
+      "loss": 0.0807,
+      "theoretical_loss": 3.8648592465251586,
+      "tokens_seen": 605290496
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008247472315840154,
+      "loss": 0.0817,
+      "theoretical_loss": 3.864686189765075,
+      "tokens_seen": 605552640
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000824666987642433,
+      "loss": 0.082,
+      "theoretical_loss": 3.864513228871357,
+      "tokens_seen": 605814784
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008245867437008507,
+      "loss": 0.0809,
+      "theoretical_loss": 3.8643403637494504,
+      "tokens_seen": 606076928
+    },
+    {
+      "epoch": 0.18,
+      "objective/train/advantage_avg": 0.0005140507710166276,
+      "objective/train/docs_used": 225863,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.601678729057312,
+      "objective/train/original_loss": 1.6016786098480225,
+      "objective/train/theoretical_loss": 3.8642539670734144,
+      "objective/train/tokens_used": 626668000,
+      "objective/train/value_avg": -0.006038665771484375,
+      "objective/train/value_loss": 9.395569941261783e-05,
+      "objective/train/value_max": -0.00026535987854003906,
+      "objective/train/value_min": -0.2125244140625,
+      "objective/train/value_reward_corr": 0.690265895238366,
+      "objective/train/value_std": 0.0079193115234375,
+      "objective/train/weight_avg": 1.0005604028701782,
+      "objective/train/weighted_lm_loss": 1.6026065349578857,
+      "objective/train/weights_max": 1.0788187980651855,
+      "objective/train/weights_min": 0.8205270171165466,
+      "theoretical_loss": 3.8642539670734144,
+      "tokens_seen": 606208000
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008245064997592682,
+      "loss": 0.0797,
+      "theoretical_loss": 3.8641675943049343,
+      "tokens_seen": 606339072
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008244262558176858,
+      "loss": 0.0831,
+      "theoretical_loss": 3.863994920443523,
+      "tokens_seen": 606601216
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008243460118761034,
+      "loss": 0.0815,
+      "theoretical_loss": 3.8638223420710647,
+      "tokens_seen": 606863360
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000824265767934521,
+      "loss": 0.079,
+      "theoretical_loss": 3.863649859093538,
+      "tokens_seen": 607125504
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008241855239929385,
+      "loss": 0.0795,
+      "theoretical_loss": 3.863477471417059,
+      "tokens_seen": 607387648
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008241052800513561,
+      "loss": 0.0821,
+      "theoretical_loss": 3.8633051789478734,
+      "tokens_seen": 607649792
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008240250361097737,
+      "loss": 0.0841,
+      "theoretical_loss": 3.8631329815923605,
+      "tokens_seen": 607911936
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008239447921681912,
+      "loss": 0.079,
+      "theoretical_loss": 3.862960879257032,
+      "tokens_seen": 608174080
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000823864548226609,
+      "loss": 0.0832,
+      "theoretical_loss": 3.8627888718485313,
+      "tokens_seen": 608436224
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008237843042850265,
+      "loss": 0.0798,
+      "theoretical_loss": 3.862616959273635,
+      "tokens_seen": 608698368
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008237040603434442,
+      "loss": 0.0803,
+      "theoretical_loss": 3.8624451414392498,
+      "tokens_seen": 608960512
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008236238164018617,
+      "loss": 0.0785,
+      "theoretical_loss": 3.8622734182524154,
+      "tokens_seen": 609222656
+    },
+    {
+      "epoch": 0.18,
+      "objective/train/advantage_avg": 0.00217598513700068,
+      "objective/train/docs_used": 227108,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5498684644699097,
+      "objective/train/original_loss": 1.5498683452606201,
+      "objective/train/theoretical_loss": 3.8621017896203007,
+      "objective/train/tokens_used": 629944800,
+      "objective/train/value_avg": -0.00588226318359375,
+      "objective/train/value_loss": 8.349979179911315e-05,
+      "objective/train/value_max": -0.0002779960632324219,
+      "objective/train/value_min": -0.254150390625,
+      "objective/train/value_reward_corr": 0.5373123562529432,
+      "objective/train/value_std": 0.00652313232421875,
+      "objective/train/weight_avg": 1.0022169351577759,
+      "objective/train/weighted_lm_loss": 1.5533596277236938,
+      "objective/train/weights_max": 1.130799412727356,
+      "objective/train/weights_min": 0.6347857117652893,
+      "theoretical_loss": 3.8621017896203007,
+      "tokens_seen": 609484800
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008235435724602792,
+      "loss": 0.0803,
+      "theoretical_loss": 3.8621017896203007,
+      "tokens_seen": 609484800
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008234633285186969,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8619302554502077,
+      "tokens_seen": 609746944
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.0008233830845771144,
+      "loss": 0.0787,
+      "theoretical_loss": 3.8617588156495666,
+      "tokens_seen": 610009088
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 0.000823302840635532,
+      "loss": 0.0829,
+      "theoretical_loss": 3.861587470125941,
+      "tokens_seen": 610271232
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008232225966939496,
+      "loss": 0.081,
+      "theoretical_loss": 3.8614162187870216,
+      "tokens_seen": 610533376
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008231423527523673,
+      "loss": 0.0804,
+      "theoretical_loss": 3.861245061540631,
+      "tokens_seen": 610795520
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008230621088107848,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8610739982947218,
+      "tokens_seen": 611057664
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008229818648692024,
+      "loss": 0.0773,
+      "theoretical_loss": 3.8609030289573747,
+      "tokens_seen": 611319808
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.00082290162092762,
+      "loss": 0.0826,
+      "theoretical_loss": 3.8607321534368007,
+      "tokens_seen": 611581952
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008228213769860375,
+      "loss": 0.0772,
+      "theoretical_loss": 3.8605613716413396,
+      "tokens_seen": 611844096
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008227411330444552,
+      "loss": 0.0822,
+      "theoretical_loss": 3.860390683479459,
+      "tokens_seen": 612106240
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008226608891028727,
+      "loss": 0.0778,
+      "theoretical_loss": 3.860220088859757,
+      "tokens_seen": 612368384
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008225806451612903,
+      "loss": 0.0774,
+      "theoretical_loss": 3.860049587690958,
+      "tokens_seen": 612630528
+    },
+    {
+      "epoch": 0.19,
+      "objective/train/advantage_avg": 0.0016880445182323456,
+      "objective/train/docs_used": 228263,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.764600396156311,
+      "objective/train/original_loss": 1.7646005153656006,
+      "objective/train/theoretical_loss": 3.8599643721221586,
+      "objective/train/tokens_used": 633221600,
+      "objective/train/value_avg": -0.006122589111328125,
+      "objective/train/value_loss": 0.00013401157048065215,
+      "objective/train/value_max": -0.00021660327911376953,
+      "objective/train/value_min": -0.330322265625,
+      "objective/train/value_reward_corr": 0.6502286444986164,
+      "objective/train/value_std": 0.01097869873046875,
+      "objective/train/weight_avg": 1.0017540454864502,
+      "objective/train/weighted_lm_loss": 1.767932415008545,
+      "objective/train/weights_max": 1.193005919456482,
+      "objective/train/weights_min": 0.8255929350852966,
+      "theoretical_loss": 3.8599643721221586,
+      "tokens_seen": 612761600
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008225004012197079,
+      "loss": 0.0799,
+      "theoretical_loss": 3.8598791798819154,
+      "tokens_seen": 612892672
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008224201572781254,
+      "loss": 0.0771,
+      "theoretical_loss": 3.859708865341611,
+      "tokens_seen": 613154816
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008223399133365432,
+      "loss": 0.0819,
+      "theoretical_loss": 3.8595386439791532,
+      "tokens_seen": 613416960
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008222596693949607,
+      "loss": 0.0794,
+      "theoretical_loss": 3.859368515703778,
+      "tokens_seen": 613679104
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008221794254533783,
+      "loss": 0.0779,
+      "theoretical_loss": 3.859198480424849,
+      "tokens_seen": 613941248
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008220991815117959,
+      "loss": 0.0787,
+      "theoretical_loss": 3.859028538051856,
+      "tokens_seen": 614203392
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008220189375702135,
+      "loss": 0.0799,
+      "theoretical_loss": 3.858858688494416,
+      "tokens_seen": 614465536
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000821938693628631,
+      "loss": 0.0789,
+      "theoretical_loss": 3.8586889316622726,
+      "tokens_seen": 614727680
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008218584496870486,
+      "loss": 0.076,
+      "theoretical_loss": 3.8585192674652955,
+      "tokens_seen": 614989824
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008217782057454662,
+      "loss": 0.0788,
+      "theoretical_loss": 3.8583496958134793,
+      "tokens_seen": 615251968
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008216979618038837,
+      "loss": 0.0768,
+      "theoretical_loss": 3.8581802166169457,
+      "tokens_seen": 615514112
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008216177178623015,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8580108297859415,
+      "tokens_seen": 615776256
+    },
+    {
+      "epoch": 0.19,
+      "objective/train/advantage_avg": 0.001968114171177149,
+      "objective/train/docs_used": 229528,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.539759874343872,
+      "objective/train/original_loss": 1.5397597551345825,
+      "objective/train/theoretical_loss": 3.857841535230839,
+      "objective/train/tokens_used": 636498400,
+      "objective/train/value_avg": -0.0091400146484375,
+      "objective/train/value_loss": 0.00024838329409249127,
+      "objective/train/value_max": -0.00023412704467773438,
+      "objective/train/value_min": -0.63916015625,
+      "objective/train/value_reward_corr": 0.6929620425353863,
+      "objective/train/value_std": 0.01442718505859375,
+      "objective/train/weight_avg": 1.0020780563354492,
+      "objective/train/weighted_lm_loss": 1.543168306350708,
+      "objective/train/weights_max": 1.582837700843811,
+      "objective/train/weights_min": 0.23275621235370636,
+      "theoretical_loss": 3.857841535230839,
+      "tokens_seen": 616038400
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000821537473920719,
+      "loss": 0.0777,
+      "theoretical_loss": 3.857841535230839,
+      "tokens_seen": 616038400
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008214572299791366,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8576723328621347,
+      "tokens_seen": 616300544
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008213769860375542,
+      "loss": 0.0794,
+      "theoretical_loss": 3.8575032225904513,
+      "tokens_seen": 616562688
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008212967420959717,
+      "loss": 0.0775,
+      "theoretical_loss": 3.8573342043265346,
+      "tokens_seen": 616824832
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008212164981543894,
+      "loss": 0.0832,
+      "theoretical_loss": 3.857165277981256,
+      "tokens_seen": 617086976
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008211362542128069,
+      "loss": 0.0811,
+      "theoretical_loss": 3.8569964434656105,
+      "tokens_seen": 617349120
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008210560102712245,
+      "loss": 0.0811,
+      "theoretical_loss": 3.856827700690718,
+      "tokens_seen": 617611264
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008209757663296422,
+      "loss": 0.0795,
+      "theoretical_loss": 3.8566590495678192,
+      "tokens_seen": 617873408
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008208955223880598,
+      "loss": 0.0817,
+      "theoretical_loss": 3.856490490008282,
+      "tokens_seen": 618135552
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008208152784464773,
+      "loss": 0.0809,
+      "theoretical_loss": 3.856322021923595,
+      "tokens_seen": 618397696
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000820735034504895,
+      "loss": 0.08,
+      "theoretical_loss": 3.8561536452253713,
+      "tokens_seen": 618659840
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008206547905633125,
+      "loss": 0.0773,
+      "theoretical_loss": 3.855985359825346,
+      "tokens_seen": 618921984
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.00082057454662173,
+      "loss": 0.0827,
+      "theoretical_loss": 3.855817165635377,
+      "tokens_seen": 619184128
+    },
+    {
+      "epoch": 0.19,
+      "objective/train/advantage_avg": 0.0019921851344406605,
+      "objective/train/docs_used": 230719,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.756550669670105,
+      "objective/train/original_loss": 1.756550669670105,
+      "objective/train/theoretical_loss": 3.855733102716653,
+      "objective/train/tokens_used": 639775200,
+      "objective/train/value_avg": -0.006992340087890625,
+      "objective/train/value_loss": 0.0001579941454110667,
+      "objective/train/value_max": -0.0001926422119140625,
+      "objective/train/value_min": -0.21337890625,
+      "objective/train/value_reward_corr": 0.45218047642197834,
+      "objective/train/value_std": 0.00893402099609375,
+      "objective/train/weight_avg": 1.002066969871521,
+      "objective/train/weighted_lm_loss": 1.7593868970870972,
+      "objective/train/weights_max": 1.1907228231430054,
+      "objective/train/weights_min": 0.36894404888153076,
+      "theoretical_loss": 3.855733102716653,
+      "tokens_seen": 619315200
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008204943026801477,
+      "loss": 0.082,
+      "theoretical_loss": 3.8556490625674447,
+      "tokens_seen": 619446272
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008204140587385652,
+      "loss": 0.0815,
+      "theoretical_loss": 3.855481050533651,
+      "tokens_seen": 619708416
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008203338147969828,
+      "loss": 0.0814,
+      "theoretical_loss": 3.8553131294462206,
+      "tokens_seen": 619970560
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008202535708554004,
+      "loss": 0.0814,
+      "theoretical_loss": 3.8551452992175,
+      "tokens_seen": 620232704
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008201733269138181,
+      "loss": 0.0813,
+      "theoretical_loss": 3.8549775597599556,
+      "tokens_seen": 620494848
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008200930829722356,
+      "loss": 0.0788,
+      "theoretical_loss": 3.8548099109861775,
+      "tokens_seen": 620756992
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008200128390306532,
+      "loss": 0.0811,
+      "theoretical_loss": 3.8546423528088747,
+      "tokens_seen": 621019136
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008199325950890708,
+      "loss": 0.0804,
+      "theoretical_loss": 3.8544748851408777,
+      "tokens_seen": 621281280
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008198523511474884,
+      "loss": 0.0784,
+      "theoretical_loss": 3.8543075078951388,
+      "tokens_seen": 621543424
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000819772107205906,
+      "loss": 0.0784,
+      "theoretical_loss": 3.8541402209847284,
+      "tokens_seen": 621805568
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008196918632643235,
+      "loss": 0.0819,
+      "theoretical_loss": 3.8539730243228387,
+      "tokens_seen": 622067712
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008196116193227412,
+      "loss": 0.0778,
+      "theoretical_loss": 3.8538059178227817,
+      "tokens_seen": 622329856
+    },
+    {
+      "epoch": 0.19,
+      "objective/train/advantage_avg": 0.0007625448633916676,
+      "objective/train/docs_used": 231894,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6909435987472534,
+      "objective/train/original_loss": 1.690943717956543,
+      "objective/train/theoretical_loss": 3.8536389013979893,
+      "objective/train/tokens_used": 643052000,
+      "objective/train/value_avg": -0.0119781494140625,
+      "objective/train/value_loss": 0.00041994385537691414,
+      "objective/train/value_max": -0.00027370452880859375,
+      "objective/train/value_min": -0.638671875,
+      "objective/train/value_reward_corr": 0.6989086451961908,
+      "objective/train/value_std": 0.0189056396484375,
+      "objective/train/weight_avg": 1.0009546279907227,
+      "objective/train/weighted_lm_loss": 1.6919503211975098,
+      "objective/train/weights_max": 1.6490602493286133,
+      "objective/train/weights_min": 0.37530699372291565,
+      "theoretical_loss": 3.8536389013979893,
+      "tokens_seen": 622592000
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008195313753811587,
+      "loss": 0.0836,
+      "theoretical_loss": 3.8536389013979893,
+      "tokens_seen": 622592000
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008194511314395762,
+      "loss": 0.0795,
+      "theoretical_loss": 3.8534719749620114,
+      "tokens_seen": 622854144
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000819370887497994,
+      "loss": 0.0788,
+      "theoretical_loss": 3.8533051384285195,
+      "tokens_seen": 623116288
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008192906435564115,
+      "loss": 0.0814,
+      "theoretical_loss": 3.853138391711303,
+      "tokens_seen": 623378432
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008192103996148291,
+      "loss": 0.0795,
+      "theoretical_loss": 3.852971734724269,
+      "tokens_seen": 623640576
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008191301556732467,
+      "loss": 0.0787,
+      "theoretical_loss": 3.8528051673814456,
+      "tokens_seen": 623902720
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008190499117316643,
+      "loss": 0.081,
+      "theoretical_loss": 3.8526386895969775,
+      "tokens_seen": 624164864
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008189696677900818,
+      "loss": 0.0826,
+      "theoretical_loss": 3.8524723012851294,
+      "tokens_seen": 624427008
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008188894238484994,
+      "loss": 0.081,
+      "theoretical_loss": 3.852306002360282,
+      "tokens_seen": 624689152
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000818809179906917,
+      "loss": 0.0816,
+      "theoretical_loss": 3.852139792736936,
+      "tokens_seen": 624951296
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008187289359653345,
+      "loss": 0.0787,
+      "theoretical_loss": 3.8519736723297067,
+      "tokens_seen": 625213440
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008186486920237523,
+      "loss": 0.0791,
+      "theoretical_loss": 3.8518076410533304,
+      "tokens_seen": 625475584
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008185684480821698,
+      "loss": 0.0789,
+      "theoretical_loss": 3.8516416988226574,
+      "tokens_seen": 625737728
+    },
+    {
+      "epoch": 0.19,
+      "objective/train/advantage_avg": 0.0014086907031014562,
+      "objective/train/docs_used": 233149,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4563045501708984,
+      "objective/train/original_loss": 1.4563043117523193,
+      "objective/train/theoretical_loss": 3.8515587610728836,
+      "objective/train/tokens_used": 646328800,
+      "objective/train/value_avg": -0.00971221923828125,
+      "objective/train/value_loss": 0.0002561133005656302,
+      "objective/train/value_max": -0.00016605854034423828,
+      "objective/train/value_min": -0.36669921875,
+      "objective/train/value_reward_corr": 0.7151985996360277,
+      "objective/train/value_std": 0.01548004150390625,
+      "objective/train/weight_avg": 1.0015289783477783,
+      "objective/train/weighted_lm_loss": 1.457317590713501,
+      "objective/train/weights_max": 1.146925687789917,
+      "objective/train/weights_min": 0.3961734175682068,
+      "theoretical_loss": 3.8515587610728836,
+      "tokens_seen": 625868800
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008184882041405875,
+      "loss": 0.0795,
+      "theoretical_loss": 3.851475845552658,
+      "tokens_seen": 625999872
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000818407960199005,
+      "loss": 0.0821,
+      "theoretical_loss": 3.8513100811584158,
+      "tokens_seen": 626262016
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008183277162574225,
+      "loss": 0.0818,
+      "theoretical_loss": 3.851144405555134,
+      "tokens_seen": 626524160
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008182474723158402,
+      "loss": 0.0788,
+      "theoretical_loss": 3.85097881865813,
+      "tokens_seen": 626786304
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008181672283742577,
+      "loss": 0.0809,
+      "theoretical_loss": 3.850813320382839,
+      "tokens_seen": 627048448
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008180869844326753,
+      "loss": 0.0812,
+      "theoretical_loss": 3.8506479106448115,
+      "tokens_seen": 627310592
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000818006740491093,
+      "loss": 0.0824,
+      "theoretical_loss": 3.8504825893597134,
+      "tokens_seen": 627572736
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008179264965495106,
+      "loss": 0.0803,
+      "theoretical_loss": 3.850317356443326,
+      "tokens_seen": 627834880
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008178462526079281,
+      "loss": 0.079,
+      "theoretical_loss": 3.8501522118115465,
+      "tokens_seen": 628097024
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008177660086663458,
+      "loss": 0.0827,
+      "theoretical_loss": 3.8499871553803873,
+      "tokens_seen": 628359168
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008176857647247633,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8498221870659743,
+      "tokens_seen": 628621312
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008176055207831808,
+      "loss": 0.0763,
+      "theoretical_loss": 3.8496573067845503,
+      "tokens_seen": 628883456
+    },
+    {
+      "epoch": 0.19,
+      "objective/train/advantage_avg": 0.0006253430037759244,
+      "objective/train/docs_used": 234355,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4915151596069336,
+      "objective/train/original_loss": 1.4915152788162231,
+      "objective/train/theoretical_loss": 3.8494925144524705,
+      "objective/train/tokens_used": 649605600,
+      "objective/train/value_avg": -0.00966644287109375,
+      "objective/train/value_loss": 0.0007618270465172827,
+      "objective/train/value_max": -0.00021660327911376953,
+      "objective/train/value_min": -0.759765625,
+      "objective/train/value_reward_corr": 0.5889753777352726,
+      "objective/train/value_std": 0.0186614990234375,
+      "objective/train/weight_avg": 1.0009468793869019,
+      "objective/train/weighted_lm_loss": 1.4915693998336792,
+      "objective/train/weights_max": 1.9064472913742065,
+      "objective/train/weights_min": 0.3697463870048523,
+      "theoretical_loss": 3.8494925144524705,
+      "tokens_seen": 629145600
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008175252768415985,
+      "loss": 0.0813,
+      "theoretical_loss": 3.8494925144524705,
+      "tokens_seen": 629145600
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000817445032900016,
+      "loss": 0.0806,
+      "theoretical_loss": 3.849327809986206,
+      "tokens_seen": 629407744
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008173647889584337,
+      "loss": 0.0772,
+      "theoretical_loss": 3.8491631933023407,
+      "tokens_seen": 629669888
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008172845450168512,
+      "loss": 0.0817,
+      "theoretical_loss": 3.8489986643175733,
+      "tokens_seen": 629932032
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008172043010752689,
+      "loss": 0.0805,
+      "theoretical_loss": 3.8488342229487156,
+      "tokens_seen": 630194176
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008171240571336865,
+      "loss": 0.0822,
+      "theoretical_loss": 3.8486698691126935,
+      "tokens_seen": 630456320
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000817043813192104,
+      "loss": 0.081,
+      "theoretical_loss": 3.8485056027265454,
+      "tokens_seen": 630718464
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008169635692505216,
+      "loss": 0.0802,
+      "theoretical_loss": 3.848341423707423,
+      "tokens_seen": 630980608
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008168833253089392,
+      "loss": 0.0791,
+      "theoretical_loss": 3.8481773319725914,
+      "tokens_seen": 631242752
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008168030813673568,
+      "loss": 0.0834,
+      "theoretical_loss": 3.8480133274394275,
+      "tokens_seen": 631504896
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008167228374257743,
+      "loss": 0.0764,
+      "theoretical_loss": 3.8478494100254217,
+      "tokens_seen": 631767040
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000816642593484192,
+      "loss": 0.0828,
+      "theoretical_loss": 3.847685579648176,
+      "tokens_seen": 632029184
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008165623495426095,
+      "loss": 0.0826,
+      "theoretical_loss": 3.847521836225404,
+      "tokens_seen": 632291328
+    },
+    {
+      "epoch": 0.19,
+      "objective/train/advantage_avg": 0.0012705049011856318,
+      "objective/train/docs_used": 235545,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.8972015380859375,
+      "objective/train/original_loss": 1.8972015380859375,
+      "objective/train/theoretical_loss": 3.847439997096262,
+      "objective/train/tokens_used": 652882400,
+      "objective/train/value_avg": -0.0090484619140625,
+      "objective/train/value_loss": 0.0002362952072871849,
+      "objective/train/value_max": -0.0002065896987915039,
+      "objective/train/value_min": -0.40380859375,
+      "objective/train/value_reward_corr": 0.6654254308226828,
+      "objective/train/value_std": 0.0135650634765625,
+      "objective/train/weight_avg": 1.0013830661773682,
+      "objective/train/weighted_lm_loss": 1.8993679285049438,
+      "objective/train/weights_max": 1.2981646060943604,
+      "objective/train/weights_min": 0.36874493956565857,
+      "theoretical_loss": 3.847439997096262,
+      "tokens_seen": 632422400
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000816482105601027,
+      "loss": 0.0829,
+      "theoretical_loss": 3.8473581796749317,
+      "tokens_seen": 632553472
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008164018616594448,
+      "loss": 0.0774,
+      "theoretical_loss": 3.8471946099146983,
+      "tokens_seen": 632815616
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008163216177178623,
+      "loss": 0.0793,
+      "theoretical_loss": 3.847031126862751,
+      "tokens_seen": 633077760
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008162413737762799,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8468677304372507,
+      "tokens_seen": 633339904
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008161611298346975,
+      "loss": 0.0781,
+      "theoretical_loss": 3.8467044205564704,
+      "tokens_seen": 633602048
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008160808858931151,
+      "loss": 0.0804,
+      "theoretical_loss": 3.846541197138791,
+      "tokens_seen": 633864192
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008160006419515327,
+      "loss": 0.0839,
+      "theoretical_loss": 3.8463780601027056,
+      "tokens_seen": 634126336
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008159203980099502,
+      "loss": 0.0786,
+      "theoretical_loss": 3.846215009366819,
+      "tokens_seen": 634388480
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008158401540683678,
+      "loss": 0.0812,
+      "theoretical_loss": 3.846052044849843,
+      "tokens_seen": 634650624
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008157599101267855,
+      "loss": 0.0796,
+      "theoretical_loss": 3.845889166470604,
+      "tokens_seen": 634912768
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008156796661852031,
+      "loss": 0.0804,
+      "theoretical_loss": 3.8457263741480343,
+      "tokens_seen": 635174912
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008155994222436206,
+      "loss": 0.0791,
+      "theoretical_loss": 3.845563667801178,
+      "tokens_seen": 635437056
+    },
+    {
+      "epoch": 0.19,
+      "objective/train/advantage_avg": 0.0004186597652733326,
+      "objective/train/docs_used": 236526,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6329374313354492,
+      "objective/train/original_loss": 1.6329374313354492,
+      "objective/train/theoretical_loss": 3.8454010473491884,
+      "objective/train/tokens_used": 656159200,
+      "objective/train/value_avg": -0.007476806640625,
+      "objective/train/value_loss": 0.00015600294864270836,
+      "objective/train/value_max": -0.0002512931823730469,
+      "objective/train/value_min": -0.230712890625,
+      "objective/train/value_reward_corr": 0.6553133134938699,
+      "objective/train/value_std": 0.01104736328125,
+      "objective/train/weight_avg": 1.0004956722259521,
+      "objective/train/weighted_lm_loss": 1.6333415508270264,
+      "objective/train/weights_max": 1.2097692489624023,
+      "objective/train/weights_min": 0.7830726504325867,
+      "theoretical_loss": 3.8454010473491884,
+      "tokens_seen": 635699200
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008155191783020383,
+      "loss": 0.081,
+      "theoretical_loss": 3.8454010473491884,
+      "tokens_seen": 635699200
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008154389343604558,
+      "loss": 0.0795,
+      "theoretical_loss": 3.845238512711327,
+      "tokens_seen": 635961344
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008153586904188733,
+      "loss": 0.0817,
+      "theoretical_loss": 3.845076063806966,
+      "tokens_seen": 636223488
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000815278446477291,
+      "loss": 0.0833,
+      "theoretical_loss": 3.844913700555586,
+      "tokens_seen": 636485632
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008151982025357085,
+      "loss": 0.0782,
+      "theoretical_loss": 3.8447514228767763,
+      "tokens_seen": 636747776
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008151179585941261,
+      "loss": 0.0831,
+      "theoretical_loss": 3.844589230690234,
+      "tokens_seen": 637009920
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008150377146525437,
+      "loss": 0.0797,
+      "theoretical_loss": 3.844427123915766,
+      "tokens_seen": 637272064
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008149574707109614,
+      "loss": 0.0781,
+      "theoretical_loss": 3.8442651024732863,
+      "tokens_seen": 637534208
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000814877226769379,
+      "loss": 0.0784,
+      "theoretical_loss": 3.8441031662828173,
+      "tokens_seen": 637796352
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008147969828277966,
+      "loss": 0.0799,
+      "theoretical_loss": 3.843941315264489,
+      "tokens_seen": 638058496
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008147167388862141,
+      "loss": 0.0788,
+      "theoretical_loss": 3.8437795493385387,
+      "tokens_seen": 638320640
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008146364949446317,
+      "loss": 0.0813,
+      "theoretical_loss": 3.8436178684253126,
+      "tokens_seen": 638582784
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008145562510030493,
+      "loss": 0.0819,
+      "theoretical_loss": 3.843456272445262,
+      "tokens_seen": 638844928
+    },
+    {
+      "epoch": 0.19,
+      "objective/train/advantage_avg": 0.0014796998584643006,
+      "objective/train/docs_used": 237858,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4043687582015991,
+      "objective/train/original_loss": 1.4043687582015991,
+      "objective/train/theoretical_loss": 3.8433755062803483,
+      "objective/train/tokens_used": 659436000,
+      "objective/train/value_avg": -0.0091400146484375,
+      "objective/train/value_loss": 0.0002844578993972391,
+      "objective/train/value_max": -0.0001881122589111328,
+      "objective/train/value_min": -0.2335205078125,
+      "objective/train/value_reward_corr": 0.6038268655262016,
+      "objective/train/value_std": 0.0128326416015625,
+      "objective/train/weight_avg": 1.0016050338745117,
+      "objective/train/weighted_lm_loss": 1.4062108993530273,
+      "objective/train/weights_max": 1.1530903577804565,
+      "objective/train/weights_min": 0.4062868654727936,
+      "theoretical_loss": 3.8433755062803483,
+      "tokens_seen": 638976000
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008144760070614668,
+      "loss": 0.078,
+      "theoretical_loss": 3.8432947613189468,
+      "tokens_seen": 639107072
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008143957631198845,
+      "loss": 0.0825,
+      "theoretical_loss": 3.8431333349670336,
+      "tokens_seen": 639369216
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000814315519178302,
+      "loss": 0.0788,
+      "theoretical_loss": 3.842971993310294,
+      "tokens_seen": 639631360
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008142352752367197,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8428107362696085,
+      "tokens_seen": 639893504
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008141550312951373,
+      "loss": 0.0782,
+      "theoretical_loss": 3.842649563765962,
+      "tokens_seen": 640155648
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008140747873535548,
+      "loss": 0.0818,
+      "theoretical_loss": 3.8424884757204474,
+      "tokens_seen": 640417792
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008139945434119724,
+      "loss": 0.08,
+      "theoretical_loss": 3.8423274720542606,
+      "tokens_seen": 640679936
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.00081391429947039,
+      "loss": 0.0786,
+      "theoretical_loss": 3.842166552688706,
+      "tokens_seen": 640942080
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008138340555288076,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8420057175451934,
+      "tokens_seen": 641204224
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008137538115872251,
+      "loss": 0.0812,
+      "theoretical_loss": 3.841844966545236,
+      "tokens_seen": 641466368
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008136735676456428,
+      "loss": 0.0802,
+      "theoretical_loss": 3.841684299610453,
+      "tokens_seen": 641728512
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008135933237040603,
+      "loss": 0.079,
+      "theoretical_loss": 3.8415237166625698,
+      "tokens_seen": 641990656
+    },
+    {
+      "epoch": 0.19,
+      "objective/train/advantage_avg": -2.3364991648122668e-05,
+      "objective/train/docs_used": 239059,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5975748300552368,
+      "objective/train/original_loss": 1.5975749492645264,
+      "objective/train/theoretical_loss": 3.8413632176234147,
+      "objective/train/tokens_used": 662712800,
+      "objective/train/value_avg": -0.006862640380859375,
+      "objective/train/value_loss": 0.0001523960818303749,
+      "objective/train/value_max": -0.0002033710479736328,
+      "objective/train/value_min": -0.264892578125,
+      "objective/train/value_reward_corr": 0.6269222691924341,
+      "objective/train/value_std": 0.00958251953125,
+      "objective/train/weight_avg": 1.0000513792037964,
+      "objective/train/weighted_lm_loss": 1.597486138343811,
+      "objective/train/weights_max": 1.1600202322006226,
+      "objective/train/weights_min": 0.6130765676498413,
+      "theoretical_loss": 3.8413632176234147,
+      "tokens_seen": 642252800
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.000813513079762478,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8413632176234147,
+      "tokens_seen": 642252800
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008134328358208956,
+      "loss": 0.0834,
+      "theoretical_loss": 3.8412028024149225,
+      "tokens_seen": 642514944
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008133525918793131,
+      "loss": 0.079,
+      "theoretical_loss": 3.841042470959131,
+      "tokens_seen": 642777088
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008132723479377308,
+      "loss": 0.0791,
+      "theoretical_loss": 3.840882223178183,
+      "tokens_seen": 643039232
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 0.0008131921039961483,
+      "loss": 0.0811,
+      "theoretical_loss": 3.8407220589943254,
+      "tokens_seen": 643301376
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008131118600545659,
+      "loss": 0.0778,
+      "theoretical_loss": 3.8405619783299083,
+      "tokens_seen": 643563520
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008130316161129835,
+      "loss": 0.0784,
+      "theoretical_loss": 3.8404019811073864,
+      "tokens_seen": 643825664
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.000812951372171401,
+      "loss": 0.0787,
+      "theoretical_loss": 3.840242067249317,
+      "tokens_seen": 644087808
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008128711282298186,
+      "loss": 0.0783,
+      "theoretical_loss": 3.840082236678362,
+      "tokens_seen": 644349952
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008127908842882363,
+      "loss": 0.08,
+      "theoretical_loss": 3.8399224893172854,
+      "tokens_seen": 644612096
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008127106403466539,
+      "loss": 0.0794,
+      "theoretical_loss": 3.839762825088955,
+      "tokens_seen": 644874240
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008126303964050714,
+      "loss": 0.0833,
+      "theoretical_loss": 3.8396032439163394,
+      "tokens_seen": 645136384
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008125501524634891,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8394437457225132,
+      "tokens_seen": 645398528
+    },
+    {
+      "epoch": 0.2,
+      "objective/train/advantage_avg": -3.4955566661665216e-05,
+      "objective/train/docs_used": 240316,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6006709337234497,
+      "objective/train/original_loss": 1.6006708145141602,
+      "objective/train/theoretical_loss": 3.8393640277186343,
+      "objective/train/tokens_used": 665989600,
+      "objective/train/value_avg": -0.01154327392578125,
+      "objective/train/value_loss": 0.00029328398522920907,
+      "objective/train/value_max": -0.00016224384307861328,
+      "objective/train/value_min": -0.419189453125,
+      "objective/train/value_reward_corr": 0.8165028368002758,
+      "objective/train/value_std": 0.0216522216796875,
+      "objective/train/weight_avg": 1.0001044273376465,
+      "objective/train/weighted_lm_loss": 1.5991564989089966,
+      "objective/train/weights_max": 1.134090781211853,
+      "objective/train/weights_min": 0.3825574517250061,
+      "theoretical_loss": 3.8393640277186343,
+      "tokens_seen": 645529600
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008124699085219066,
+      "loss": 0.0832,
+      "theoretical_loss": 3.839284330430651,
+      "tokens_seen": 645660672
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008123896645803241,
+      "loss": 0.081,
+      "theoretical_loss": 3.8391249979640305,
+      "tokens_seen": 645922816
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008123094206387418,
+      "loss": 0.0802,
+      "theoretical_loss": 3.8389657482460313,
+      "tokens_seen": 646184960
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008122291766971593,
+      "loss": 0.0774,
+      "theoretical_loss": 3.838806581200134,
+      "tokens_seen": 646447104
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.000812148932755577,
+      "loss": 0.0808,
+      "theoretical_loss": 3.838647496749924,
+      "tokens_seen": 646709248
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008120686888139945,
+      "loss": 0.0818,
+      "theoretical_loss": 3.8384884948190847,
+      "tokens_seen": 646971392
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008119884448724122,
+      "loss": 0.0809,
+      "theoretical_loss": 3.838329575331403,
+      "tokens_seen": 647233536
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008119082009308298,
+      "loss": 0.0798,
+      "theoretical_loss": 3.8381707382107657,
+      "tokens_seen": 647495680
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008118279569892473,
+      "loss": 0.0836,
+      "theoretical_loss": 3.838011983381162,
+      "tokens_seen": 647757824
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008117477130476649,
+      "loss": 0.0792,
+      "theoretical_loss": 3.8378533107666817,
+      "tokens_seen": 648019968
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008116674691060825,
+      "loss": 0.0833,
+      "theoretical_loss": 3.8376947202915144,
+      "tokens_seen": 648282112
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008115872251645001,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8375362118799505,
+      "tokens_seen": 648544256
+    },
+    {
+      "epoch": 0.2,
+      "objective/train/advantage_avg": 0.002151393797248602,
+      "objective/train/docs_used": 241544,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5193239450454712,
+      "objective/train/original_loss": 1.5193241834640503,
+      "objective/train/theoretical_loss": 3.8373777854563813,
+      "objective/train/tokens_used": 669266400,
+      "objective/train/value_avg": -0.007175445556640625,
+      "objective/train/value_loss": 0.0002748131228145212,
+      "objective/train/value_max": -0.00010722875595092773,
+      "objective/train/value_min": -0.6533203125,
+      "objective/train/value_reward_corr": 0.5608935796152028,
+      "objective/train/value_std": 0.01296234130859375,
+      "objective/train/weight_avg": 1.002267599105835,
+      "objective/train/weighted_lm_loss": 1.5228184461593628,
+      "objective/train/weights_max": 1.1310755014419556,
+      "objective/train/weights_min": 0.36963358521461487,
+      "theoretical_loss": 3.8373777854563813,
+      "tokens_seen": 648806400
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008115069812229176,
+      "loss": 0.0798,
+      "theoretical_loss": 3.8373777854563813,
+      "tokens_seen": 648806400
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008114267372813353,
+      "loss": 0.0816,
+      "theoretical_loss": 3.837219440945298,
+      "tokens_seen": 649068544
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008113464933397528,
+      "loss": 0.08,
+      "theoretical_loss": 3.8370611782712922,
+      "tokens_seen": 649330688
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008112662493981705,
+      "loss": 0.0823,
+      "theoretical_loss": 3.8369029973590543,
+      "tokens_seen": 649592832
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008111860054565881,
+      "loss": 0.0796,
+      "theoretical_loss": 3.836744898133376,
+      "tokens_seen": 649854976
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008111057615150056,
+      "loss": 0.079,
+      "theoretical_loss": 3.8365868805191456,
+      "tokens_seen": 650117120
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008110255175734233,
+      "loss": 0.0801,
+      "theoretical_loss": 3.836428944441354,
+      "tokens_seen": 650379264
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008109452736318408,
+      "loss": 0.0804,
+      "theoretical_loss": 3.8362710898250896,
+      "tokens_seen": 650641408
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008108650296902584,
+      "loss": 0.0819,
+      "theoretical_loss": 3.83611331659554,
+      "tokens_seen": 650903552
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.000810784785748676,
+      "loss": 0.0765,
+      "theoretical_loss": 3.8359556246779913,
+      "tokens_seen": 651165696
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008107045418070936,
+      "loss": 0.0828,
+      "theoretical_loss": 3.8357980139978283,
+      "tokens_seen": 651427840
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008106242978655111,
+      "loss": 0.0827,
+      "theoretical_loss": 3.8356404844805354,
+      "tokens_seen": 651689984
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008105440539239288,
+      "loss": 0.0817,
+      "theoretical_loss": 3.835483036051694,
+      "tokens_seen": 651952128
+    },
+    {
+      "epoch": 0.2,
+      "objective/train/advantage_avg": 0.0021593719720840454,
+      "objective/train/docs_used": 242832,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6270049810409546,
+      "objective/train/original_loss": 1.6270049810409546,
+      "objective/train/theoretical_loss": 3.8354043422222133,
+      "objective/train/tokens_used": 672543200,
+      "objective/train/value_avg": -0.00695037841796875,
+      "objective/train/value_loss": 0.0001394545251969248,
+      "objective/train/value_max": -0.00016736984252929688,
+      "objective/train/value_min": -0.36376953125,
+      "objective/train/value_reward_corr": 0.674807627723227,
+      "objective/train/value_std": 0.0125274658203125,
+      "objective/train/weight_avg": 1.0022249221801758,
+      "objective/train/weighted_lm_loss": 1.6309560537338257,
+      "objective/train/weights_max": 1.1779428720474243,
+      "objective/train/weights_min": 0.3801831901073456,
+      "theoretical_loss": 3.8354043422222133,
+      "tokens_seen": 652083200
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008104638099823464,
+      "loss": 0.0802,
+      "theoretical_loss": 3.835325668636983,
+      "tokens_seen": 652214272
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008103835660407639,
+      "loss": 0.0816,
+      "theoretical_loss": 3.8351683821621814,
+      "tokens_seen": 652476416
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008103033220991816,
+      "loss": 0.0819,
+      "theoretical_loss": 3.8350111765531647,
+      "tokens_seen": 652738560
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008102230781575991,
+      "loss": 0.0797,
+      "theoretical_loss": 3.834854051735906,
+      "tokens_seen": 653000704
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008101428342160167,
+      "loss": 0.0824,
+      "theoretical_loss": 3.8346970076364757,
+      "tokens_seen": 653262848
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008100625902744343,
+      "loss": 0.0798,
+      "theoretical_loss": 3.8345400441810424,
+      "tokens_seen": 653524992
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008099823463328518,
+      "loss": 0.0818,
+      "theoretical_loss": 3.8343831612958703,
+      "tokens_seen": 653787136
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008099021023912694,
+      "loss": 0.0817,
+      "theoretical_loss": 3.834226358907322,
+      "tokens_seen": 654049280
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.000809821858449687,
+      "loss": 0.0817,
+      "theoretical_loss": 3.8340696369418565,
+      "tokens_seen": 654311424
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008097416145081047,
+      "loss": 0.0785,
+      "theoretical_loss": 3.833912995326029,
+      "tokens_seen": 654573568
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008096613705665223,
+      "loss": 0.0791,
+      "theoretical_loss": 3.8337564339864914,
+      "tokens_seen": 654835712
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008095811266249399,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8335999528499913,
+      "tokens_seen": 655097856
+    },
+    {
+      "debugging/Compilability": 0.9230769230769231,
+      "debugging/distinct-1-grams": 0.7560193640250044,
+      "debugging/entropy-1-grams": 5.2534808493234895,
+      "debugging/length": 422.2307692307692,
+      "debugging/num_segments": 13,
+      "debugging/raw_token_scores_avg": 0.00857908371835947,
+      "debugging/raw_token_scores_std": 0.02154596894979477,
+      "debugging/score": 0.01638830111902192,
+      "debugging/score_std": 0.019821935735086043,
+      "epoch": 0.2,
+      "objective/train/advantage_avg": 0.002418318996205926,
+      "objective/train/docs_used": 243976,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5416268110275269,
+      "objective/train/original_loss": 1.5416266918182373,
+      "objective/train/theoretical_loss": 3.833443551843374,
+      "objective/train/tokens_used": 675820000,
+      "objective/train/value_avg": -0.01099395751953125,
+      "objective/train/value_loss": 0.00022288458421826363,
+      "objective/train/value_max": -0.00031256675720214844,
+      "objective/train/value_min": -0.326904296875,
+      "objective/train/value_reward_corr": 0.7334904683634422,
+      "objective/train/value_std": 0.0173797607421875,
+      "objective/train/weight_avg": 1.0025254487991333,
+      "objective/train/weighted_lm_loss": 1.5436171293258667,
+      "objective/train/weights_max": 1.2229843139648438,
+      "objective/train/weights_min": 0.37925034761428833,
+      "theoretical_loss": 3.833443551843374,
+      "tokens_seen": 655360000
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008095008826833574,
+      "loss": 0.0824,
+      "theoretical_loss": 3.833443551843374,
+      "tokens_seen": 655360000
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.000809420638741775,
+      "loss": 0.0802,
+      "theoretical_loss": 3.8332872308935793,
+      "tokens_seen": 655622144
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008093403948001926,
+      "loss": 0.081,
+      "theoretical_loss": 3.833130989927643,
+      "tokens_seen": 655884288
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008092601508586101,
+      "loss": 0.0797,
+      "theoretical_loss": 3.8329748288726972,
+      "tokens_seen": 656146432
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008091799069170278,
+      "loss": 0.0838,
+      "theoretical_loss": 3.8328187476559687,
+      "tokens_seen": 656408576
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008090996629754453,
+      "loss": 0.082,
+      "theoretical_loss": 3.83266274620478,
+      "tokens_seen": 656670720
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.000809019419033863,
+      "loss": 0.0783,
+      "theoretical_loss": 3.832506824446549,
+      "tokens_seen": 656932864
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008089391750922806,
+      "loss": 0.0795,
+      "theoretical_loss": 3.832350982308788,
+      "tokens_seen": 657195008
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008088589311506981,
+      "loss": 0.0784,
+      "theoretical_loss": 3.8321952197191043,
+      "tokens_seen": 657457152
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008087786872091157,
+      "loss": 0.0816,
+      "theoretical_loss": 3.8320395366052,
+      "tokens_seen": 657719296
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008086984432675333,
+      "loss": 0.0795,
+      "theoretical_loss": 3.8318839328948715,
+      "tokens_seen": 657981440
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008086181993259509,
+      "loss": 0.0807,
+      "theoretical_loss": 3.8317284085160095,
+      "tokens_seen": 658243584
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008085379553843684,
+      "loss": 0.0805,
+      "theoretical_loss": 3.8315729633965994,
+      "tokens_seen": 658505728
+    },
+    {
+      "epoch": 0.2,
+      "objective/train/advantage_avg": -0.001669611083343625,
+      "objective/train/docs_used": 245288,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5030797719955444,
+      "objective/train/original_loss": 1.503079891204834,
+      "objective/train/theoretical_loss": 3.83149527053671,
+      "objective/train/tokens_used": 679096800,
+      "objective/train/value_avg": -0.00933837890625,
+      "objective/train/value_loss": 0.0003506782522890717,
+      "objective/train/value_max": -0.0002532005310058594,
+      "objective/train/value_min": -0.263427734375,
+      "objective/train/value_reward_corr": 0.5930356277614546,
+      "objective/train/value_std": 0.01061248779296875,
+      "objective/train/weight_avg": 0.9984805583953857,
+      "objective/train/weighted_lm_loss": 1.5003520250320435,
+      "objective/train/weights_max": 1.164748191833496,
+      "objective/train/weights_min": 0.22747893631458282,
+      "theoretical_loss": 3.83149527053671,
+      "tokens_seen": 658636800
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008084577114427861,
+      "loss": 0.0786,
+      "theoretical_loss": 3.8314175974647195,
+      "tokens_seen": 658767872
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008083774675012036,
+      "loss": 0.0808,
+      "theoretical_loss": 3.831262310648544,
+      "tokens_seen": 659030016
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008082972235596214,
+      "loss": 0.0816,
+      "theoretical_loss": 3.831107102876338,
+      "tokens_seen": 659292160
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008082169796180389,
+      "loss": 0.0829,
+      "theoretical_loss": 3.830951974076463,
+      "tokens_seen": 659554304
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008081367356764564,
+      "loss": 0.0825,
+      "theoretical_loss": 3.830796924177371,
+      "tokens_seen": 659816448
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008080564917348741,
+      "loss": 0.0792,
+      "theoretical_loss": 3.830641953107609,
+      "tokens_seen": 660078592
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008079762477932916,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8304870607958175,
+      "tokens_seen": 660340736
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008078960038517092,
+      "loss": 0.0807,
+      "theoretical_loss": 3.8303322471707286,
+      "tokens_seen": 660602880
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008078157599101268,
+      "loss": 0.0784,
+      "theoretical_loss": 3.830177512161167,
+      "tokens_seen": 660865024
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008077355159685444,
+      "loss": 0.0818,
+      "theoretical_loss": 3.8300228556960523,
+      "tokens_seen": 661127168
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008076552720269619,
+      "loss": 0.0778,
+      "theoretical_loss": 3.829868277704393,
+      "tokens_seen": 661389312
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008075750280853796,
+      "loss": 0.0808,
+      "theoretical_loss": 3.829713778115293,
+      "tokens_seen": 661651456
+    },
+    {
+      "epoch": 0.2,
+      "objective/train/advantage_avg": 0.0014472852926701307,
+      "objective/train/docs_used": 246530,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6413122415542603,
+      "objective/train/original_loss": 1.6413123607635498,
+      "objective/train/theoretical_loss": 3.8295593568579465,
+      "objective/train/tokens_used": 682373600,
+      "objective/train/value_avg": -0.0119781494140625,
+      "objective/train/value_loss": 0.000297568243695423,
+      "objective/train/value_max": -0.00022172927856445312,
+      "objective/train/value_min": -0.2919921875,
+      "objective/train/value_reward_corr": 0.7385408733331538,
+      "objective/train/value_std": 0.019012451171875,
+      "objective/train/weight_avg": 1.0015898942947388,
+      "objective/train/weighted_lm_loss": 1.6430256366729736,
+      "objective/train/weights_max": 1.2300209999084473,
+      "objective/train/weights_min": 0.36995241045951843,
+      "theoretical_loss": 3.8295593568579465,
+      "tokens_seen": 661913600
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008074947841437972,
+      "loss": 0.0816,
+      "theoretical_loss": 3.8295593568579465,
+      "tokens_seen": 661913600
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008074145402022147,
+      "loss": 0.0813,
+      "theoretical_loss": 3.8294050138616402,
+      "tokens_seen": 662175744
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008073342962606324,
+      "loss": 0.081,
+      "theoretical_loss": 3.8292507490557526,
+      "tokens_seen": 662437888
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008072540523190499,
+      "loss": 0.0768,
+      "theoretical_loss": 3.8290965623697537,
+      "tokens_seen": 662700032
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008071738083774676,
+      "loss": 0.0783,
+      "theoretical_loss": 3.8289424537332053,
+      "tokens_seen": 662962176
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008070935644358851,
+      "loss": 0.0828,
+      "theoretical_loss": 3.82878842307576,
+      "tokens_seen": 663224320
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008070133204943026,
+      "loss": 0.0797,
+      "theoretical_loss": 3.828634470327162,
+      "tokens_seen": 663486464
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008069330765527203,
+      "loss": 0.0808,
+      "theoretical_loss": 3.8284805954172474,
+      "tokens_seen": 663748608
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008068528326111379,
+      "loss": 0.0812,
+      "theoretical_loss": 3.828326798275941,
+      "tokens_seen": 664010752
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008067725886695555,
+      "loss": 0.0797,
+      "theoretical_loss": 3.82817307883326,
+      "tokens_seen": 664272896
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008066923447279731,
+      "loss": 0.0797,
+      "theoretical_loss": 3.8280194370193112,
+      "tokens_seen": 664535040
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008066121007863907,
+      "loss": 0.0799,
+      "theoretical_loss": 3.827865872764293,
+      "tokens_seen": 664797184
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008065318568448082,
+      "loss": 0.0802,
+      "theoretical_loss": 3.8277123859984936,
+      "tokens_seen": 665059328
+    },
+    {
+      "epoch": 0.2,
+      "objective/train/advantage_avg": 0.00026472590980120003,
+      "objective/train/docs_used": 247583,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7137538194656372,
+      "objective/train/original_loss": 1.713753581047058,
+      "objective/train/theoretical_loss": 3.82763567165229,
+      "objective/train/tokens_used": 685650400,
+      "objective/train/value_avg": -0.00846099853515625,
+      "objective/train/value_loss": 0.00032397088943980634,
+      "objective/train/value_max": -0.00016605854034423828,
+      "objective/train/value_min": -0.64111328125,
+      "objective/train/value_reward_corr": 0.6828980449967452,
+      "objective/train/value_std": 0.01425933837890625,
+      "objective/train/weight_avg": 1.00041663646698,
+      "objective/train/weighted_lm_loss": 1.7142980098724365,
+      "objective/train/weights_max": 1.409537672996521,
+      "objective/train/weights_min": 0.5323238372802734,
+      "theoretical_loss": 3.82763567165229,
+      "tokens_seen": 665190400
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008064516129032258,
+      "loss": 0.0815,
+      "theoretical_loss": 3.8275589766522895,
+      "tokens_seen": 665321472
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008063713689616434,
+      "loss": 0.0795,
+      "theoretical_loss": 3.8274056446561504,
+      "tokens_seen": 665583616
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008062911250200609,
+      "loss": 0.077,
+      "theoretical_loss": 3.827252389940633,
+      "tokens_seen": 665845760
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008062108810784786,
+      "loss": 0.079,
+      "theoretical_loss": 3.827099212436386,
+      "tokens_seen": 666107904
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008061306371368961,
+      "loss": 0.0802,
+      "theoretical_loss": 3.826946112074145,
+      "tokens_seen": 666370048
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008060503931953138,
+      "loss": 0.0791,
+      "theoretical_loss": 3.826793088784737,
+      "tokens_seen": 666632192
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008059701492537314,
+      "loss": 0.0817,
+      "theoretical_loss": 3.826640142499077,
+      "tokens_seen": 666894336
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008058899053121489,
+      "loss": 0.0786,
+      "theoretical_loss": 3.8264872731481705,
+      "tokens_seen": 667156480
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008058096613705666,
+      "loss": 0.0786,
+      "theoretical_loss": 3.8263344806631103,
+      "tokens_seen": 667418624
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008057294174289841,
+      "loss": 0.0804,
+      "theoretical_loss": 3.8261817649750784,
+      "tokens_seen": 667680768
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008056491734874017,
+      "loss": 0.0781,
+      "theoretical_loss": 3.8260291260153463,
+      "tokens_seen": 667942912
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008055689295458193,
+      "loss": 0.0798,
+      "theoretical_loss": 3.8258765637152727,
+      "tokens_seen": 668205056
+    },
+    {
+      "epoch": 0.2,
+      "objective/train/advantage_avg": -0.00018435930542182177,
+      "objective/train/docs_used": 248744,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.638197660446167,
+      "objective/train/original_loss": 1.638197660446167,
+      "objective/train/theoretical_loss": 3.8257240780063055,
+      "objective/train/tokens_used": 688927200,
+      "objective/train/value_avg": -0.006977081298828125,
+      "objective/train/value_loss": 0.0003241746162530035,
+      "objective/train/value_max": -0.00019109249114990234,
+      "objective/train/value_min": -0.227783203125,
+      "objective/train/value_reward_corr": 0.7174840285003992,
+      "objective/train/value_std": 0.0107421875,
+      "objective/train/weight_avg": 0.9999673962593079,
+      "objective/train/weighted_lm_loss": 1.6396377086639404,
+      "objective/train/weights_max": 1.1206990480422974,
+      "objective/train/weights_min": 0.39359450340270996,
+      "theoretical_loss": 3.8257240780063055,
+      "tokens_seen": 668467200
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008054886856042369,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8257240780063055,
+      "tokens_seen": 668467200
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008054084416626544,
+      "loss": 0.078,
+      "theoretical_loss": 3.8255716688199803,
+      "tokens_seen": 668729344
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008053281977210722,
+      "loss": 0.079,
+      "theoretical_loss": 3.825419336087921,
+      "tokens_seen": 668991488
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008052479537794897,
+      "loss": 0.0788,
+      "theoretical_loss": 3.825267079741839,
+      "tokens_seen": 669253632
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008051677098379072,
+      "loss": 0.0796,
+      "theoretical_loss": 3.825114899713533,
+      "tokens_seen": 669515776
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008050874658963249,
+      "loss": 0.0808,
+      "theoretical_loss": 3.8249627959348915,
+      "tokens_seen": 669777920
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008050072219547424,
+      "loss": 0.0791,
+      "theoretical_loss": 3.824810768337887,
+      "tokens_seen": 670040064
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.00080492697801316,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8246588168545816,
+      "tokens_seen": 670302208
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008048467340715776,
+      "loss": 0.0747,
+      "theoretical_loss": 3.824506941417125,
+      "tokens_seen": 670564352
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008047664901299952,
+      "loss": 0.0809,
+      "theoretical_loss": 3.824355141957752,
+      "tokens_seen": 670826496
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008046862461884128,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8242034184087847,
+      "tokens_seen": 671088640
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008046060022468304,
+      "loss": 0.0774,
+      "theoretical_loss": 3.824051770702633,
+      "tokens_seen": 671350784
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.000804525758305248,
+      "loss": 0.0821,
+      "theoretical_loss": 3.823900198771792,
+      "tokens_seen": 671612928
+    },
+    {
+      "epoch": 0.2,
+      "objective/train/advantage_avg": 0.0005832649767398834,
+      "objective/train/docs_used": 249880,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4792834520339966,
+      "objective/train/original_loss": 1.4792835712432861,
+      "objective/train/theoretical_loss": 3.823824441201042,
+      "objective/train/tokens_used": 692204000,
+      "objective/train/value_avg": -0.00798797607421875,
+      "objective/train/value_loss": 0.0001945794647326693,
+      "objective/train/value_max": -0.00016224384307861328,
+      "objective/train/value_min": -0.30322265625,
+      "objective/train/value_reward_corr": 0.7662795001045949,
+      "objective/train/value_std": 0.0170745849609375,
+      "objective/train/weight_avg": 1.0006768703460693,
+      "objective/train/weighted_lm_loss": 1.479491949081421,
+      "objective/train/weights_max": 1.267970323562622,
+      "objective/train/weights_min": 0.5421772599220276,
+      "theoretical_loss": 3.823824441201042,
+      "tokens_seen": 671744000
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008044455143636656,
+      "loss": 0.0767,
+      "theoretical_loss": 3.823748702548845,
+      "tokens_seen": 671875072
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008043652704220832,
+      "loss": 0.0775,
+      "theoretical_loss": 3.823597281966459,
+      "tokens_seen": 672137216
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008042850264805007,
+      "loss": 0.0787,
+      "theoretical_loss": 3.8234459369573894,
+      "tokens_seen": 672399360
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008042047825389184,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8232946674544763,
+      "tokens_seen": 672661504
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008041245385973359,
+      "loss": 0.0805,
+      "theoretical_loss": 3.8231434733906458,
+      "tokens_seen": 672923648
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008040442946557534,
+      "loss": 0.0803,
+      "theoretical_loss": 3.822992354698911,
+      "tokens_seen": 673185792
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008039640507141711,
+      "loss": 0.0784,
+      "theoretical_loss": 3.822841311312368,
+      "tokens_seen": 673447936
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008038838067725887,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8226903431642008,
+      "tokens_seen": 673710080
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008038035628310063,
+      "loss": 0.0789,
+      "theoretical_loss": 3.8225394501876764,
+      "tokens_seen": 673972224
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008037233188894239,
+      "loss": 0.0802,
+      "theoretical_loss": 3.822388632316149,
+      "tokens_seen": 674234368
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008036430749478415,
+      "loss": 0.0779,
+      "theoretical_loss": 3.822237889483057,
+      "tokens_seen": 674496512
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.000803562831006259,
+      "loss": 0.0794,
+      "theoretical_loss": 3.822087221621923,
+      "tokens_seen": 674758656
+    },
+    {
+      "epoch": 0.2,
+      "objective/train/advantage_avg": -0.00032604276202619076,
+      "objective/train/docs_used": 251058,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5687280893325806,
+      "objective/train/original_loss": 1.5687280893325806,
+      "objective/train/theoretical_loss": 3.821936628666355,
+      "objective/train/tokens_used": 695480800,
+      "objective/train/value_avg": -0.006984710693359375,
+      "objective/train/value_loss": 0.00032798174652270973,
+      "objective/train/value_max": -0.0001926422119140625,
+      "objective/train/value_min": -0.90283203125,
+      "objective/train/value_reward_corr": 0.6570254406817982,
+      "objective/train/value_std": 0.01174163818359375,
+      "objective/train/weight_avg": 0.9998152256011963,
+      "objective/train/weighted_lm_loss": 1.568171501159668,
+      "objective/train/weights_max": 1.1402249336242676,
+      "objective/train/weights_min": 0.2320682257413864,
+      "theoretical_loss": 3.821936628666355,
+      "tokens_seen": 675020800
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008034825870646766,
+      "loss": 0.0753,
+      "theoretical_loss": 3.821936628666355,
+      "tokens_seen": 675020800
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008034023431230942,
+      "loss": 0.0777,
+      "theoretical_loss": 3.821786110550045,
+      "tokens_seen": 675282944
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008033220991815118,
+      "loss": 0.0768,
+      "theoretical_loss": 3.8216356672067704,
+      "tokens_seen": 675545088
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008032418552399294,
+      "loss": 0.0794,
+      "theoretical_loss": 3.821485298570391,
+      "tokens_seen": 675807232
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.000803161611298347,
+      "loss": 0.0774,
+      "theoretical_loss": 3.8213350045748533,
+      "tokens_seen": 676069376
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 0.0008030813673567647,
+      "loss": 0.0787,
+      "theoretical_loss": 3.821184785154186,
+      "tokens_seen": 676331520
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008030011234151822,
+      "loss": 0.0792,
+      "theoretical_loss": 3.8210346402425013,
+      "tokens_seen": 676593664
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008029208794735997,
+      "loss": 0.0791,
+      "theoretical_loss": 3.820884569773997,
+      "tokens_seen": 676855808
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008028406355320174,
+      "loss": 0.0771,
+      "theoretical_loss": 3.8207345736829526,
+      "tokens_seen": 677117952
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008027603915904349,
+      "loss": 0.078,
+      "theoretical_loss": 3.820584651903732,
+      "tokens_seen": 677380096
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008026801476488525,
+      "loss": 0.0784,
+      "theoretical_loss": 3.820434804370782,
+      "tokens_seen": 677642240
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008025999037072701,
+      "loss": 0.0781,
+      "theoretical_loss": 3.820285031018633,
+      "tokens_seen": 677904384
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008025196597656877,
+      "loss": 0.0788,
+      "theoretical_loss": 3.8201353317818985,
+      "tokens_seen": 678166528
+    },
+    {
+      "epoch": 0.21,
+      "objective/train/advantage_avg": 0.0017591392388567328,
+      "objective/train/docs_used": 252282,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.499388575553894,
+      "objective/train/original_loss": 1.4993884563446045,
+      "objective/train/theoretical_loss": 3.8200605099364005,
+      "objective/train/tokens_used": 698757600,
+      "objective/train/value_avg": -0.0076446533203125,
+      "objective/train/value_loss": 0.00020787572429981083,
+      "objective/train/value_max": -0.00022876262664794922,
+      "objective/train/value_min": -0.267333984375,
+      "objective/train/value_reward_corr": 0.6231772208041639,
+      "objective/train/value_std": 0.0118560791015625,
+      "objective/train/weight_avg": 1.001856803894043,
+      "objective/train/weighted_lm_loss": 1.5024080276489258,
+      "objective/train/weights_max": 1.1745542287826538,
+      "objective/train/weights_min": 0.3710364103317261,
+      "theoretical_loss": 3.8200605099364005,
+      "tokens_seen": 678297600
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008024394158241052,
+      "loss": 0.0767,
+      "theoretical_loss": 3.819985706595274,
+      "tokens_seen": 678428672
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008023591718825229,
+      "loss": 0.0816,
+      "theoretical_loss": 3.8198361553935385,
+      "tokens_seen": 678690816
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008022789279409405,
+      "loss": 0.0814,
+      "theoretical_loss": 3.8196866781115526,
+      "tokens_seen": 678952960
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000802198683999358,
+      "loss": 0.0831,
+      "theoretical_loss": 3.8195372746842615,
+      "tokens_seen": 679215104
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008021184400577757,
+      "loss": 0.0767,
+      "theoretical_loss": 3.8193879450466905,
+      "tokens_seen": 679477248
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008020381961161932,
+      "loss": 0.0798,
+      "theoretical_loss": 3.819238689133948,
+      "tokens_seen": 679739392
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008019579521746109,
+      "loss": 0.0795,
+      "theoretical_loss": 3.819089506881225,
+      "tokens_seen": 680001536
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008018777082330284,
+      "loss": 0.079,
+      "theoretical_loss": 3.8189403982237935,
+      "tokens_seen": 680263680
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000801797464291446,
+      "loss": 0.0773,
+      "theoretical_loss": 3.818791363097008,
+      "tokens_seen": 680525824
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008017172203498636,
+      "loss": 0.0789,
+      "theoretical_loss": 3.818642401436304,
+      "tokens_seen": 680787968
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008016369764082812,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8184935131771987,
+      "tokens_seen": 681050112
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008015567324666988,
+      "loss": 0.0787,
+      "theoretical_loss": 3.8183446982552915,
+      "tokens_seen": 681312256
+    },
+    {
+      "epoch": 0.21,
+      "objective/train/advantage_avg": 0.0013181634712964296,
+      "objective/train/docs_used": 253498,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5834970474243164,
+      "objective/train/original_loss": 1.5834970474243164,
+      "objective/train/theoretical_loss": 3.818195956606262,
+      "objective/train/tokens_used": 702034400,
+      "objective/train/value_avg": -0.00580596923828125,
+      "objective/train/value_loss": 0.00019775994587689638,
+      "objective/train/value_max": -0.000263214111328125,
+      "objective/train/value_min": -0.3359375,
+      "objective/train/value_reward_corr": 0.4194674562674259,
+      "objective/train/value_std": 0.007694244384765625,
+      "objective/train/weight_avg": 1.0014029741287231,
+      "objective/train/weighted_lm_loss": 1.5860118865966797,
+      "objective/train/weights_max": 1.3310092687606812,
+      "objective/train/weights_min": 0.3714952766895294,
+      "theoretical_loss": 3.818195956606262,
+      "tokens_seen": 681574400
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008014764885251164,
+      "loss": 0.0807,
+      "theoretical_loss": 3.818195956606262,
+      "tokens_seen": 681574400
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000801396244583534,
+      "loss": 0.0802,
+      "theoretical_loss": 3.8180472881658707,
+      "tokens_seen": 681836544
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008013160006419515,
+      "loss": 0.0807,
+      "theoretical_loss": 3.817898692869961,
+      "tokens_seen": 682098688
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008012357567003692,
+      "loss": 0.0812,
+      "theoretical_loss": 3.817750170654455,
+      "tokens_seen": 682360832
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008011555127587867,
+      "loss": 0.0797,
+      "theoretical_loss": 3.8176017214553575,
+      "tokens_seen": 682622976
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008010752688172042,
+      "loss": 0.0803,
+      "theoretical_loss": 3.8174533452087513,
+      "tokens_seen": 682885120
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008009950248756219,
+      "loss": 0.0774,
+      "theoretical_loss": 3.817305041850802,
+      "tokens_seen": 683147264
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008009147809340394,
+      "loss": 0.0806,
+      "theoretical_loss": 3.817156811317754,
+      "tokens_seen": 683409408
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008008345369924572,
+      "loss": 0.0802,
+      "theoretical_loss": 3.8170086535459333,
+      "tokens_seen": 683671552
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008007542930508747,
+      "loss": 0.0799,
+      "theoretical_loss": 3.8168605684717454,
+      "tokens_seen": 683933696
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008006740491092923,
+      "loss": 0.08,
+      "theoretical_loss": 3.816712556031675,
+      "tokens_seen": 684195840
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008005938051677099,
+      "loss": 0.079,
+      "theoretical_loss": 3.816564616162287,
+      "tokens_seen": 684457984
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008005135612261274,
+      "loss": 0.078,
+      "theoretical_loss": 3.8164167488002265,
+      "tokens_seen": 684720128
+    },
+    {
+      "epoch": 0.21,
+      "objective/train/advantage_avg": 0.001422036555595696,
+      "objective/train/docs_used": 254674,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.559423565864563,
+      "objective/train/original_loss": 1.5594233274459839,
+      "objective/train/theoretical_loss": 3.816342842289668,
+      "objective/train/tokens_used": 705311200,
+      "objective/train/value_avg": -0.006954193115234375,
+      "objective/train/value_loss": 0.00011249267845414579,
+      "objective/train/value_max": -0.0001823902130126953,
+      "objective/train/value_min": -0.30419921875,
+      "objective/train/value_reward_corr": 0.6799434767604788,
+      "objective/train/value_std": 0.01035308837890625,
+      "objective/train/weight_avg": 1.0014777183532715,
+      "objective/train/weighted_lm_loss": 1.5613473653793335,
+      "objective/train/weights_max": 1.134773850440979,
+      "objective/train/weights_min": 0.7942905426025391,
+      "theoretical_loss": 3.816342842289668,
+      "tokens_seen": 684851200
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000800433317284545,
+      "loss": 0.0798,
+      "theoretical_loss": 3.816268953882218,
+      "tokens_seen": 684982272
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008003530733429626,
+      "loss": 0.0777,
+      "theoretical_loss": 3.8161212313450648,
+      "tokens_seen": 685244416
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008002728294013802,
+      "loss": 0.0793,
+      "theoretical_loss": 3.81597358112565,
+      "tokens_seen": 685506560
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008001925854597977,
+      "loss": 0.0798,
+      "theoretical_loss": 3.815826003160935,
+      "tokens_seen": 685768704
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0008001123415182155,
+      "loss": 0.0811,
+      "theoretical_loss": 3.815678497387962,
+      "tokens_seen": 686030848
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000800032097576633,
+      "loss": 0.0788,
+      "theoretical_loss": 3.8155310637438506,
+      "tokens_seen": 686292992
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007999518536350505,
+      "loss": 0.079,
+      "theoretical_loss": 3.8153837021657995,
+      "tokens_seen": 686555136
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007998716096934682,
+      "loss": 0.0776,
+      "theoretical_loss": 3.8152364125910863,
+      "tokens_seen": 686817280
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007997913657518857,
+      "loss": 0.0781,
+      "theoretical_loss": 3.8150891949570664,
+      "tokens_seen": 687079424
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007997111218103033,
+      "loss": 0.0782,
+      "theoretical_loss": 3.814942049201175,
+      "tokens_seen": 687341568
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007996308778687209,
+      "loss": 0.0817,
+      "theoretical_loss": 3.8147949752609236,
+      "tokens_seen": 687603712
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007995506339271385,
+      "loss": 0.0789,
+      "theoretical_loss": 3.814647973073903,
+      "tokens_seen": 687865856
+    },
+    {
+      "epoch": 0.21,
+      "objective/train/advantage_avg": 0.0019079339690506458,
+      "objective/train/docs_used": 255827,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5390352010726929,
+      "objective/train/original_loss": 1.5390353202819824,
+      "objective/train/theoretical_loss": 3.8145010425777826,
+      "objective/train/tokens_used": 708588000,
+      "objective/train/value_avg": -0.0116119384765625,
+      "objective/train/value_loss": 0.0002800696238409728,
+      "objective/train/value_max": -0.00017261505126953125,
+      "objective/train/value_min": -0.5703125,
+      "objective/train/value_reward_corr": 0.7645409376422845,
+      "objective/train/value_std": 0.020599365234375,
+      "objective/train/weight_avg": 1.0020438432693481,
+      "objective/train/weighted_lm_loss": 1.5430036783218384,
+      "objective/train/weights_max": 1.2020081281661987,
+      "objective/train/weights_min": 0.627696692943573,
+      "theoretical_loss": 3.8145010425777826,
+      "tokens_seen": 688128000
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007994703899855562,
+      "loss": 0.0788,
+      "theoretical_loss": 3.8145010425777826,
+      "tokens_seen": 688128000
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007993901460439737,
+      "loss": 0.0796,
+      "theoretical_loss": 3.814354183710308,
+      "tokens_seen": 688390144
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007993099021023913,
+      "loss": 0.0796,
+      "theoretical_loss": 3.8142073964093046,
+      "tokens_seen": 688652288
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007992296581608089,
+      "loss": 0.0782,
+      "theoretical_loss": 3.8140606806126733,
+      "tokens_seen": 688914432
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007991494142192265,
+      "loss": 0.0796,
+      "theoretical_loss": 3.813914036258393,
+      "tokens_seen": 689176576
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000799069170277644,
+      "loss": 0.0789,
+      "theoretical_loss": 3.813767463284522,
+      "tokens_seen": 689438720
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007989889263360617,
+      "loss": 0.0797,
+      "theoretical_loss": 3.8136209616291934,
+      "tokens_seen": 689700864
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007989086823944792,
+      "loss": 0.0796,
+      "theoretical_loss": 3.813474531230618,
+      "tokens_seen": 689963008
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007988284384528968,
+      "loss": 0.0785,
+      "theoretical_loss": 3.8133281720270835,
+      "tokens_seen": 690225152
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007987481945113144,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8131818839569562,
+      "tokens_seen": 690487296
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000798667950569732,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8130356669586765,
+      "tokens_seen": 690749440
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007985877066281496,
+      "loss": 0.0768,
+      "theoretical_loss": 3.8128895209707627,
+      "tokens_seen": 691011584
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007985074626865672,
+      "loss": 0.0819,
+      "theoretical_loss": 3.81274344593181,
+      "tokens_seen": 691273728
+    },
+    {
+      "epoch": 0.21,
+      "objective/train/advantage_avg": 0.001648097182624042,
+      "objective/train/docs_used": 257040,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.549683690071106,
+      "objective/train/original_loss": 1.5496838092803955,
+      "objective/train/theoretical_loss": 3.812670434999026,
+      "objective/train/tokens_used": 711864800,
+      "objective/train/value_avg": -0.0057830810546875,
+      "objective/train/value_loss": 0.00016456494631711394,
+      "objective/train/value_max": -0.00016605854034423828,
+      "objective/train/value_min": -0.3056640625,
+      "objective/train/value_reward_corr": 0.6402803699022559,
+      "objective/train/value_std": 0.00890350341796875,
+      "objective/train/weight_avg": 1.0017253160476685,
+      "objective/train/weighted_lm_loss": 1.5529325008392334,
+      "objective/train/weights_max": 1.138139009475708,
+      "objective/train/weights_min": 0.368623286485672,
+      "theoretical_loss": 3.812670434999026,
+      "tokens_seen": 691404800
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007984272187449848,
+      "loss": 0.0774,
+      "theoretical_loss": 3.8125974417804893,
+      "tokens_seen": 691535872
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007983469748034023,
+      "loss": 0.0767,
+      "theoretical_loss": 3.8124515084555477,
+      "tokens_seen": 691798016
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.00079826673086182,
+      "loss": 0.0786,
+      "theoretical_loss": 3.8123056458958087,
+      "tokens_seen": 692060160
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007981864869202375,
+      "loss": 0.078,
+      "theoretical_loss": 3.812159854040172,
+      "tokens_seen": 692322304
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007981062429786551,
+      "loss": 0.0796,
+      "theoretical_loss": 3.8120141328276125,
+      "tokens_seen": 692584448
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007980259990370727,
+      "loss": 0.0813,
+      "theoretical_loss": 3.8118684821971813,
+      "tokens_seen": 692846592
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007979457550954902,
+      "loss": 0.0807,
+      "theoretical_loss": 3.8117229020880057,
+      "tokens_seen": 693108736
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000797865511153908,
+      "loss": 0.0804,
+      "theoretical_loss": 3.811577392439287,
+      "tokens_seen": 693370880
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007977852672123255,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8114319531903025,
+      "tokens_seen": 693633024
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007977050232707431,
+      "loss": 0.0807,
+      "theoretical_loss": 3.811286584280406,
+      "tokens_seen": 693895168
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007976247793291607,
+      "loss": 0.0799,
+      "theoretical_loss": 3.8111412856490245,
+      "tokens_seen": 694157312
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007975445353875782,
+      "loss": 0.08,
+      "theoretical_loss": 3.810996057235661,
+      "tokens_seen": 694419456
+    },
+    {
+      "epoch": 0.21,
+      "objective/train/advantage_avg": 0.0017769966507330537,
+      "objective/train/docs_used": 258341,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6005815267562866,
+      "objective/train/original_loss": 1.6005815267562866,
+      "objective/train/theoretical_loss": 3.810850898979894,
+      "objective/train/tokens_used": 715141600,
+      "objective/train/value_avg": -0.00868988037109375,
+      "objective/train/value_loss": 0.00037193644675426185,
+      "objective/train/value_max": -0.00015234947204589844,
+      "objective/train/value_min": -0.72021484375,
+      "objective/train/value_reward_corr": 0.5291458465205756,
+      "objective/train/value_std": 0.0122528076171875,
+      "objective/train/weight_avg": 1.001926302909851,
+      "objective/train/weighted_lm_loss": 1.60252046585083,
+      "objective/train/weights_max": 1.3593815565109253,
+      "objective/train/weights_min": 0.23074059188365936,
+      "theoretical_loss": 3.810850898979894,
+      "tokens_seen": 694681600
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007974642914459958,
+      "loss": 0.08,
+      "theoretical_loss": 3.810850898979894,
+      "tokens_seen": 694681600
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007973840475044134,
+      "loss": 0.0808,
+      "theoretical_loss": 3.810705810821375,
+      "tokens_seen": 694943744
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000797303803562831,
+      "loss": 0.0795,
+      "theoretical_loss": 3.8105607926998326,
+      "tokens_seen": 695205888
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007972235596212485,
+      "loss": 0.0785,
+      "theoretical_loss": 3.810415844555067,
+      "tokens_seen": 695468032
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007971433156796663,
+      "loss": 0.0778,
+      "theoretical_loss": 3.8102709663269554,
+      "tokens_seen": 695730176
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007970630717380838,
+      "loss": 0.0791,
+      "theoretical_loss": 3.810126157955448,
+      "tokens_seen": 695992320
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007969828277965014,
+      "loss": 0.0779,
+      "theoretical_loss": 3.809981419380569,
+      "tokens_seen": 696254464
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000796902583854919,
+      "loss": 0.0797,
+      "theoretical_loss": 3.809836750542418,
+      "tokens_seen": 696516608
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007968223399133365,
+      "loss": 0.0803,
+      "theoretical_loss": 3.8096921513811663,
+      "tokens_seen": 696778752
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007967420959717542,
+      "loss": 0.0804,
+      "theoretical_loss": 3.809547621837061,
+      "tokens_seen": 697040896
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007966618520301717,
+      "loss": 0.0765,
+      "theoretical_loss": 3.809403161850423,
+      "tokens_seen": 697303040
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007965816080885893,
+      "loss": 0.0823,
+      "theoretical_loss": 3.8092587713616446,
+      "tokens_seen": 697565184
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000796501364147007,
+      "loss": 0.081,
+      "theoretical_loss": 3.809114450311193,
+      "tokens_seen": 697827328
+    },
+    {
+      "epoch": 0.21,
+      "objective/train/advantage_avg": -0.0008030504104681313,
+      "objective/train/docs_used": 259086,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6132134199142456,
+      "objective/train/original_loss": 1.613213300704956,
+      "objective/train/theoretical_loss": 3.8090423158067566,
+      "objective/train/tokens_used": 718418400,
+      "objective/train/value_avg": -0.00884246826171875,
+      "objective/train/value_loss": 0.00026143278228119016,
+      "objective/train/value_max": -0.0002033710479736328,
+      "objective/train/value_min": -0.25,
+      "objective/train/value_reward_corr": 0.8812842327841226,
+      "objective/train/value_std": 0.0213165283203125,
+      "objective/train/weight_avg": 0.9993177652359009,
+      "objective/train/weighted_lm_loss": 1.6114174127578735,
+      "objective/train/weights_max": 1.1213833093643188,
+      "objective/train/weights_min": 0.37042829394340515,
+      "theoretical_loss": 3.8090423158067566,
+      "tokens_seen": 697958400
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007964211202054245,
+      "loss": 0.08,
+      "theoretical_loss": 3.808970198639609,
+      "tokens_seen": 698089472
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007963408762638421,
+      "loss": 0.0781,
+      "theoretical_loss": 3.808826016287507,
+      "tokens_seen": 698351616
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007962606323222597,
+      "loss": 0.0804,
+      "theoretical_loss": 3.8086819031955725,
+      "tokens_seen": 698613760
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007961803883806773,
+      "loss": 0.0787,
+      "theoretical_loss": 3.8085378593045665,
+      "tokens_seen": 698875904
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007961001444390948,
+      "loss": 0.0817,
+      "theoretical_loss": 3.8083938845553202,
+      "tokens_seen": 699138048
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007960199004975125,
+      "loss": 0.0783,
+      "theoretical_loss": 3.80824997888874,
+      "tokens_seen": 699400192
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.00079593965655593,
+      "loss": 0.0807,
+      "theoretical_loss": 3.8081061422458036,
+      "tokens_seen": 699662336
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007958594126143476,
+      "loss": 0.0785,
+      "theoretical_loss": 3.8079623745675613,
+      "tokens_seen": 699924480
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007957791686727652,
+      "loss": 0.0755,
+      "theoretical_loss": 3.8078186757951364,
+      "tokens_seen": 700186624
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007956989247311828,
+      "loss": 0.0807,
+      "theoretical_loss": 3.8076750458697237,
+      "tokens_seen": 700448768
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007956186807896005,
+      "loss": 0.0789,
+      "theoretical_loss": 3.807531484732591,
+      "tokens_seen": 700710912
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000795538436848018,
+      "loss": 0.0793,
+      "theoretical_loss": 3.8073879923250775,
+      "tokens_seen": 700973056
+    },
+    {
+      "epoch": 0.21,
+      "objective/train/advantage_avg": 0.0006472544046118855,
+      "objective/train/docs_used": 260260,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.584854006767273,
+      "objective/train/original_loss": 1.5848541259765625,
+      "objective/train/theoretical_loss": 3.8072445685885947,
+      "objective/train/tokens_used": 721695200,
+      "objective/train/value_avg": -0.007328033447265625,
+      "objective/train/value_loss": 0.0002117718249792233,
+      "objective/train/value_max": -7.784366607666016e-05,
+      "objective/train/value_min": -0.60693359375,
+      "objective/train/value_reward_corr": 0.616764286984292,
+      "objective/train/value_std": 0.011688232421875,
+      "objective/train/weight_avg": 1.000746250152588,
+      "objective/train/weighted_lm_loss": 1.5859571695327759,
+      "objective/train/weights_max": 1.590544581413269,
+      "objective/train/weights_min": 0.38592278957366943,
+      "theoretical_loss": 3.8072445685885947,
+      "tokens_seen": 701235200
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007954581929064356,
+      "loss": 0.0775,
+      "theoretical_loss": 3.8072445685885947,
+      "tokens_seen": 701235200
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007953779489648532,
+      "loss": 0.0785,
+      "theoretical_loss": 3.807101213464625,
+      "tokens_seen": 701497344
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007952977050232708,
+      "loss": 0.0803,
+      "theoretical_loss": 3.8069579268947242,
+      "tokens_seen": 701759488
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007952174610816883,
+      "loss": 0.0812,
+      "theoretical_loss": 3.806814708820519,
+      "tokens_seen": 702021632
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007951372171401059,
+      "loss": 0.0759,
+      "theoretical_loss": 3.806671559183706,
+      "tokens_seen": 702283776
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007950569731985235,
+      "loss": 0.078,
+      "theoretical_loss": 3.806528477926056,
+      "tokens_seen": 702545920
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000794976729256941,
+      "loss": 0.0811,
+      "theoretical_loss": 3.806385464989409,
+      "tokens_seen": 702808064
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007948964853153588,
+      "loss": 0.0808,
+      "theoretical_loss": 3.806242520315676,
+      "tokens_seen": 703070208
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007948162413737763,
+      "loss": 0.0788,
+      "theoretical_loss": 3.806099643846841,
+      "tokens_seen": 703332352
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007947359974321939,
+      "loss": 0.0792,
+      "theoretical_loss": 3.8059568355249564,
+      "tokens_seen": 703594496
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007946557534906115,
+      "loss": 0.0789,
+      "theoretical_loss": 3.8058140952921478,
+      "tokens_seen": 703856640
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000794575509549029,
+      "loss": 0.0796,
+      "theoretical_loss": 3.805671423090609,
+      "tokens_seen": 704118784
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007944952656074467,
+      "loss": 0.0781,
+      "theoretical_loss": 3.805528818862607,
+      "tokens_seen": 704380928
+    },
+    {
+      "epoch": 0.21,
+      "objective/train/advantage_avg": 0.00023795383458491415,
+      "objective/train/docs_used": 261464,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4358924627304077,
+      "objective/train/original_loss": 1.4358923435211182,
+      "objective/train/theoretical_loss": 3.8054575422206596,
+      "objective/train/tokens_used": 724972000,
+      "objective/train/value_avg": -0.00519561767578125,
+      "objective/train/value_loss": 0.00011082483979407698,
+      "objective/train/value_max": -0.00012433528900146484,
+      "objective/train/value_min": -0.2100830078125,
+      "objective/train/value_reward_corr": 0.568821173994195,
+      "objective/train/value_std": 0.00617218017578125,
+      "objective/train/weight_avg": 1.000292420387268,
+      "objective/train/weighted_lm_loss": 1.4362776279449463,
+      "objective/train/weights_max": 1.1238564252853394,
+      "objective/train/weights_min": 0.7808533906936646,
+      "theoretical_loss": 3.8054575422206596,
+      "tokens_seen": 704512000
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007944150216658642,
+      "loss": 0.078,
+      "theoretical_loss": 3.8053862825504776,
+      "tokens_seen": 704643072
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007943347777242818,
+      "loss": 0.0806,
+      "theoretical_loss": 3.8052438140966265,
+      "tokens_seen": 704905216
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007942545337826995,
+      "loss": 0.0804,
+      "theoretical_loss": 3.8051014134435315,
+      "tokens_seen": 705167360
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007941742898411171,
+      "loss": 0.0751,
+      "theoretical_loss": 3.804959080533739,
+      "tokens_seen": 705429504
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007940940458995346,
+      "loss": 0.0791,
+      "theoretical_loss": 3.8048168153098656,
+      "tokens_seen": 705691648
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007940138019579522,
+      "loss": 0.0787,
+      "theoretical_loss": 3.8046746177145985,
+      "tokens_seen": 705953792
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007939335580163698,
+      "loss": 0.0799,
+      "theoretical_loss": 3.804532487690694,
+      "tokens_seen": 706215936
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007938533140747873,
+      "loss": 0.0782,
+      "theoretical_loss": 3.8043904251809786,
+      "tokens_seen": 706478080
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.000793773070133205,
+      "loss": 0.0796,
+      "theoretical_loss": 3.8042484301283475,
+      "tokens_seen": 706740224
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007936928261916225,
+      "loss": 0.0805,
+      "theoretical_loss": 3.8041065024757668,
+      "tokens_seen": 707002368
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007936125822500401,
+      "loss": 0.0791,
+      "theoretical_loss": 3.8039646421662705,
+      "tokens_seen": 707264512
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007935323383084578,
+      "loss": 0.0783,
+      "theoretical_loss": 3.8038228491429624,
+      "tokens_seen": 707526656
+    },
+    {
+      "epoch": 0.21,
+      "objective/train/advantage_avg": 0.00028257613303139806,
+      "objective/train/docs_used": 262675,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7718579769134521,
+      "objective/train/original_loss": 1.7718580961227417,
+      "objective/train/theoretical_loss": 3.8036811233490164,
+      "objective/train/tokens_used": 728248800,
+      "objective/train/value_avg": -0.01105499267578125,
+      "objective/train/value_loss": 0.00044336056453175843,
+      "objective/train/value_max": -0.0001442432403564453,
+      "objective/train/value_min": -0.88037109375,
+      "objective/train/value_reward_corr": 0.5845583896713396,
+      "objective/train/value_std": 0.0162353515625,
+      "objective/train/weight_avg": 1.0004814863204956,
+      "objective/train/weighted_lm_loss": 1.772655725479126,
+      "objective/train/weights_max": 1.4637386798858643,
+      "objective/train/weights_min": 0.3728155791759491,
+      "theoretical_loss": 3.8036811233490164,
+      "tokens_seen": 707788800
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007934520943668753,
+      "loss": 0.0805,
+      "theoretical_loss": 3.8036811233490164,
+      "tokens_seen": 707788800
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007933718504252929,
+      "loss": 0.0776,
+      "theoretical_loss": 3.803539464727673,
+      "tokens_seen": 708050944
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007932916064837105,
+      "loss": 0.0815,
+      "theoretical_loss": 3.803397873222244,
+      "tokens_seen": 708313088
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007932113625421281,
+      "loss": 0.0802,
+      "theoretical_loss": 3.8032563487761095,
+      "tokens_seen": 708575232
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007931311186005457,
+      "loss": 0.0775,
+      "theoretical_loss": 3.8031148913327177,
+      "tokens_seen": 708837376
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007930508746589633,
+      "loss": 0.0807,
+      "theoretical_loss": 3.8029735008355843,
+      "tokens_seen": 709099520
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 0.0007929706307173808,
+      "loss": 0.0802,
+      "theoretical_loss": 3.8028321772282965,
+      "tokens_seen": 709361664
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007928903867757984,
+      "loss": 0.079,
+      "theoretical_loss": 3.8026909204545065,
+      "tokens_seen": 709623808
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.000792810142834216,
+      "loss": 0.0813,
+      "theoretical_loss": 3.8025497304579376,
+      "tokens_seen": 709885952
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007927298988926336,
+      "loss": 0.0786,
+      "theoretical_loss": 3.80240860718238,
+      "tokens_seen": 710148096
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007926496549510513,
+      "loss": 0.0829,
+      "theoretical_loss": 3.802267550571691,
+      "tokens_seen": 710410240
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007925694110094688,
+      "loss": 0.082,
+      "theoretical_loss": 3.802126560569798,
+      "tokens_seen": 710672384
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007924891670678864,
+      "loss": 0.0815,
+      "theoretical_loss": 3.801985637120694,
+      "tokens_seen": 710934528
+    },
+    {
+      "epoch": 0.22,
+      "objective/train/advantage_avg": 0.0011543107684701681,
+      "objective/train/docs_used": 263840,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6443710327148438,
+      "objective/train/original_loss": 1.6443710327148438,
+      "objective/train/theoretical_loss": 3.801915200335954,
+      "objective/train/tokens_used": 731525600,
+      "objective/train/value_avg": -0.00934600830078125,
+      "objective/train/value_loss": 0.00040373342926613986,
+      "objective/train/value_max": -0.0002199411392211914,
+      "objective/train/value_min": -0.466552734375,
+      "objective/train/value_reward_corr": 0.6620460795564023,
+      "objective/train/value_std": 0.01470947265625,
+      "objective/train/weight_avg": 1.0013347864151,
+      "objective/train/weighted_lm_loss": 1.6481596231460571,
+      "objective/train/weights_max": 1.1715375185012817,
+      "objective/train/weights_min": 0.37442609667778015,
+      "theoretical_loss": 3.801915200335954,
+      "tokens_seen": 711065600
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.000792408923126304,
+      "loss": 0.0788,
+      "theoretical_loss": 3.801844780168441,
+      "tokens_seen": 711196672
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007923286791847216,
+      "loss": 0.0816,
+      "theoretical_loss": 3.8017039896571685,
+      "tokens_seen": 711458816
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007922484352431391,
+      "loss": 0.0787,
+      "theoretical_loss": 3.8015632655310734,
+      "tokens_seen": 711720960
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007921681913015567,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8014226077344198,
+      "tokens_seen": 711983104
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007920879473599743,
+      "loss": 0.0779,
+      "theoretical_loss": 3.8012820162115393,
+      "tokens_seen": 712245248
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007920077034183918,
+      "loss": 0.0772,
+      "theoretical_loss": 3.801141490906831,
+      "tokens_seen": 712507392
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007919274594768096,
+      "loss": 0.078,
+      "theoretical_loss": 3.80100103176476,
+      "tokens_seen": 712769536
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007918472155352271,
+      "loss": 0.0816,
+      "theoretical_loss": 3.8008606387298594,
+      "tokens_seen": 713031680
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007917669715936448,
+      "loss": 0.0767,
+      "theoretical_loss": 3.80072031174673,
+      "tokens_seen": 713293824
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007916867276520623,
+      "loss": 0.0804,
+      "theoretical_loss": 3.800580050760036,
+      "tokens_seen": 713555968
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007916064837104798,
+      "loss": 0.0764,
+      "theoretical_loss": 3.800439855714512,
+      "tokens_seen": 713818112
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007915262397688975,
+      "loss": 0.08,
+      "theoretical_loss": 3.8002997265549574,
+      "tokens_seen": 714080256
+    },
+    {
+      "epoch": 0.22,
+      "objective/train/advantage_avg": 0.001398808672092855,
+      "objective/train/docs_used": 265036,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4737719297409058,
+      "objective/train/original_loss": 1.4737719297409058,
+      "objective/train/theoretical_loss": 3.8001596632262387,
+      "objective/train/tokens_used": 734802400,
+      "objective/train/value_avg": -0.009124755859375,
+      "objective/train/value_loss": 0.00014532258501276374,
+      "objective/train/value_max": -0.00020992755889892578,
+      "objective/train/value_min": -0.1900634765625,
+      "objective/train/value_reward_corr": 0.6496980006355382,
+      "objective/train/value_std": 0.0111541748046875,
+      "objective/train/weight_avg": 1.0014667510986328,
+      "objective/train/weighted_lm_loss": 1.4759221076965332,
+      "objective/train/weights_max": 1.1377222537994385,
+      "objective/train/weights_min": 0.3690439462661743,
+      "theoretical_loss": 3.8001596632262387,
+      "tokens_seen": 714342400
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.000791445995827315,
+      "loss": 0.0801,
+      "theoretical_loss": 3.8001596632262387,
+      "tokens_seen": 714342400
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007913657518857326,
+      "loss": 0.0779,
+      "theoretical_loss": 3.8000196656732874,
+      "tokens_seen": 714604544
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007912855079441503,
+      "loss": 0.078,
+      "theoretical_loss": 3.7998797338411032,
+      "tokens_seen": 714866688
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007912052640025679,
+      "loss": 0.0797,
+      "theoretical_loss": 3.7997398676747496,
+      "tokens_seen": 715128832
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007911250200609854,
+      "loss": 0.0797,
+      "theoretical_loss": 3.7996000671193593,
+      "tokens_seen": 715390976
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.000791044776119403,
+      "loss": 0.0777,
+      "theoretical_loss": 3.7994603321201277,
+      "tokens_seen": 715653120
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007909645321778206,
+      "loss": 0.0787,
+      "theoretical_loss": 3.7993206626223177,
+      "tokens_seen": 715915264
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007908842882362381,
+      "loss": 0.0789,
+      "theoretical_loss": 3.799181058571258,
+      "tokens_seen": 716177408
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007908040442946558,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7990415199123424,
+      "tokens_seen": 716439552
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007907238003530733,
+      "loss": 0.0809,
+      "theoretical_loss": 3.79890204659103,
+      "tokens_seen": 716701696
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.000790643556411491,
+      "loss": 0.0783,
+      "theoretical_loss": 3.7987626385528466,
+      "tokens_seen": 716963840
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007905633124699085,
+      "loss": 0.0818,
+      "theoretical_loss": 3.798623295743382,
+      "tokens_seen": 717225984
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.000790483068528326,
+      "loss": 0.0818,
+      "theoretical_loss": 3.798484018108291,
+      "tokens_seen": 717488128
+    },
+    {
+      "epoch": 0.22,
+      "objective/train/advantage_avg": 0.0011304336367174983,
+      "objective/train/docs_used": 266305,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5863089561462402,
+      "objective/train/original_loss": 1.5863087177276611,
+      "objective/train/theoretical_loss": 3.7984144037141716,
+      "objective/train/tokens_used": 738079200,
+      "objective/train/value_avg": -0.0092620849609375,
+      "objective/train/value_loss": 0.0005252858973108232,
+      "objective/train/value_max": -0.0002434253692626953,
+      "objective/train/value_min": -0.962890625,
+      "objective/train/value_reward_corr": 0.61228862282039,
+      "objective/train/value_std": 0.0162353515625,
+      "objective/train/weight_avg": 1.0013542175292969,
+      "objective/train/weighted_lm_loss": 1.5876915454864502,
+      "objective/train/weights_max": 1.5222142934799194,
+      "objective/train/weights_min": 0.23608994483947754,
+      "theoretical_loss": 3.7984144037141716,
+      "tokens_seen": 717619200
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007904028245867438,
+      "loss": 0.0783,
+      "theoretical_loss": 3.7983448055932953,
+      "tokens_seen": 717750272
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007903225806451613,
+      "loss": 0.0822,
+      "theoretical_loss": 3.79820565814418,
+      "tokens_seen": 718012416
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007902423367035789,
+      "loss": 0.0786,
+      "theoretical_loss": 3.798066575706795,
+      "tokens_seen": 718274560
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007901620927619965,
+      "loss": 0.0779,
+      "theoretical_loss": 3.797927558227056,
+      "tokens_seen": 718536704
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007900818488204141,
+      "loss": 0.0771,
+      "theoretical_loss": 3.7977886056509433,
+      "tokens_seen": 718798848
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007900016048788316,
+      "loss": 0.0781,
+      "theoretical_loss": 3.797649717924502,
+      "tokens_seen": 719060992
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007899213609372492,
+      "loss": 0.0791,
+      "theoretical_loss": 3.797510894993839,
+      "tokens_seen": 719323136
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007898411169956668,
+      "loss": 0.0783,
+      "theoretical_loss": 3.79737213680513,
+      "tokens_seen": 719585280
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007897608730540844,
+      "loss": 0.0791,
+      "theoretical_loss": 3.797233443304612,
+      "tokens_seen": 719847424
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007896806291125021,
+      "loss": 0.0805,
+      "theoretical_loss": 3.7970948144385868,
+      "tokens_seen": 720109568
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007896003851709196,
+      "loss": 0.0764,
+      "theoretical_loss": 3.796956250153421,
+      "tokens_seen": 720371712
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007895201412293372,
+      "loss": 0.0797,
+      "theoretical_loss": 3.796817750395544,
+      "tokens_seen": 720633856
+    },
+    {
+      "epoch": 0.22,
+      "objective/train/advantage_avg": 0.0022220478858798742,
+      "objective/train/docs_used": 267624,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4708417654037476,
+      "objective/train/original_loss": 1.4708420038223267,
+      "objective/train/theoretical_loss": 3.7966793151114504,
+      "objective/train/tokens_used": 741356000,
+      "objective/train/value_avg": -0.0073699951171875,
+      "objective/train/value_loss": 0.0003509256348479539,
+      "objective/train/value_max": -0.0001838207244873047,
+      "objective/train/value_min": -0.91015625,
+      "objective/train/value_reward_corr": 0.5371351152106856,
+      "objective/train/value_std": 0.0122222900390625,
+      "objective/train/weight_avg": 1.0023695230484009,
+      "objective/train/weighted_lm_loss": 1.4751582145690918,
+      "objective/train/weights_max": 1.2902494668960571,
+      "objective/train/weights_min": 0.27010083198547363,
+      "theoretical_loss": 3.7966793151114504,
+      "tokens_seen": 720896000
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007894398972877548,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7966793151114504,
+      "tokens_seen": 720896000
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007893596533461724,
+      "loss": 0.0768,
+      "theoretical_loss": 3.796540944247698,
+      "tokens_seen": 721158144
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.00078927940940459,
+      "loss": 0.0783,
+      "theoretical_loss": 3.796402637750908,
+      "tokens_seen": 721420288
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007891991654630075,
+      "loss": 0.0809,
+      "theoretical_loss": 3.796264395567766,
+      "tokens_seen": 721682432
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007891189215214251,
+      "loss": 0.0788,
+      "theoretical_loss": 3.7961262176450195,
+      "tokens_seen": 721944576
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007890386775798428,
+      "loss": 0.08,
+      "theoretical_loss": 3.795988103929482,
+      "tokens_seen": 722206720
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007889584336382604,
+      "loss": 0.0762,
+      "theoretical_loss": 3.7958500543680276,
+      "tokens_seen": 722468864
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007888781896966779,
+      "loss": 0.079,
+      "theoretical_loss": 3.795712068907596,
+      "tokens_seen": 722731008
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007887979457550956,
+      "loss": 0.0825,
+      "theoretical_loss": 3.795574147495188,
+      "tokens_seen": 722993152
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007887177018135131,
+      "loss": 0.0807,
+      "theoretical_loss": 3.795436290077868,
+      "tokens_seen": 723255296
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007886374578719306,
+      "loss": 0.081,
+      "theoretical_loss": 3.795298496602765,
+      "tokens_seen": 723517440
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007885572139303483,
+      "loss": 0.0778,
+      "theoretical_loss": 3.795160767017068,
+      "tokens_seen": 723779584
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007884769699887658,
+      "loss": 0.0787,
+      "theoretical_loss": 3.795023101268031,
+      "tokens_seen": 724041728
+    },
+    {
+      "epoch": 0.22,
+      "objective/train/advantage_avg": 0.0011378737399354577,
+      "objective/train/docs_used": 268786,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.588625431060791,
+      "objective/train/original_loss": 1.588625431060791,
+      "objective/train/theoretical_loss": 3.7949542923157935,
+      "objective/train/tokens_used": 744632800,
+      "objective/train/value_avg": -0.007007598876953125,
+      "objective/train/value_loss": 0.00012135379802202806,
+      "objective/train/value_max": -0.00010973215103149414,
+      "objective/train/value_min": -0.343505859375,
+      "objective/train/value_reward_corr": 0.6047769086501988,
+      "objective/train/value_std": 0.00875091552734375,
+      "objective/train/weight_avg": 1.0011974573135376,
+      "objective/train/weighted_lm_loss": 1.5900264978408813,
+      "objective/train/weights_max": 1.1416176557540894,
+      "objective/train/weights_min": 0.7288485765457153,
+      "theoretical_loss": 3.7949542923157935,
+      "tokens_seen": 724172800
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007883967260471834,
+      "loss": 0.0827,
+      "theoretical_loss": 3.7948854993029695,
+      "tokens_seen": 724303872
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.000788316482105601,
+      "loss": 0.0767,
+      "theoretical_loss": 3.7947479610692616,
+      "tokens_seen": 724566016
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007882362381640187,
+      "loss": 0.0784,
+      "theoretical_loss": 3.794610486514348,
+      "tokens_seen": 724828160
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007881559942224363,
+      "loss": 0.0788,
+      "theoretical_loss": 3.7944730755857323,
+      "tokens_seen": 725090304
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007880757502808538,
+      "loss": 0.0773,
+      "theoretical_loss": 3.794335728230979,
+      "tokens_seen": 725352448
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007879955063392714,
+      "loss": 0.078,
+      "theoretical_loss": 3.7941984443977157,
+      "tokens_seen": 725614592
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.000787915262397689,
+      "loss": 0.0762,
+      "theoretical_loss": 3.7940612240336327,
+      "tokens_seen": 725876736
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007878350184561066,
+      "loss": 0.0786,
+      "theoretical_loss": 3.793924067086481,
+      "tokens_seen": 726138880
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007877547745145241,
+      "loss": 0.0781,
+      "theoretical_loss": 3.793786973504073,
+      "tokens_seen": 726401024
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007876745305729418,
+      "loss": 0.0772,
+      "theoretical_loss": 3.7936499432342847,
+      "tokens_seen": 726663168
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007875942866313593,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7935129762250526,
+      "tokens_seen": 726925312
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007875140426897769,
+      "loss": 0.0813,
+      "theoretical_loss": 3.7933760724243752,
+      "tokens_seen": 727187456
+    },
+    {
+      "epoch": 0.22,
+      "objective/train/advantage_avg": 0.0019741442520171404,
+      "objective/train/docs_used": 269904,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5811312198638916,
+      "objective/train/original_loss": 1.5811312198638916,
+      "objective/train/theoretical_loss": 3.7932392317803116,
+      "objective/train/tokens_used": 747909600,
+      "objective/train/value_avg": -0.006404876708984375,
+      "objective/train/value_loss": 0.00015449750935658813,
+      "objective/train/value_max": -0.0001398324966430664,
+      "objective/train/value_min": -0.30126953125,
+      "objective/train/value_reward_corr": 0.5738474794753151,
+      "objective/train/value_std": 0.00901031494140625,
+      "objective/train/weight_avg": 1.002044916152954,
+      "objective/train/weighted_lm_loss": 1.5848934650421143,
+      "objective/train/weights_max": 1.2912355661392212,
+      "objective/train/weights_min": 0.36876535415649414,
+      "theoretical_loss": 3.7932392317803116,
+      "tokens_seen": 727449600
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007874337987481946,
+      "loss": 0.0798,
+      "theoretical_loss": 3.7932392317803116,
+      "tokens_seen": 727449600
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007873535548066121,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7931024542409837,
+      "tokens_seen": 727711744
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007872733108650297,
+      "loss": 0.0784,
+      "theoretical_loss": 3.7929657397545733,
+      "tokens_seen": 727973888
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007871930669234473,
+      "loss": 0.0767,
+      "theoretical_loss": 3.792829088269324,
+      "tokens_seen": 728236032
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007871128229818649,
+      "loss": 0.077,
+      "theoretical_loss": 3.792692499733541,
+      "tokens_seen": 728498176
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007870325790402824,
+      "loss": 0.0785,
+      "theoretical_loss": 3.7925559740955896,
+      "tokens_seen": 728760320
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007869523350987,
+      "loss": 0.0775,
+      "theoretical_loss": 3.7924195113038968,
+      "tokens_seen": 729022464
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007868720911571176,
+      "loss": 0.0788,
+      "theoretical_loss": 3.7922831113069493,
+      "tokens_seen": 729284608
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007867918472155353,
+      "loss": 0.0762,
+      "theoretical_loss": 3.792146774053296,
+      "tokens_seen": 729546752
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007867116032739529,
+      "loss": 0.077,
+      "theoretical_loss": 3.792010499491545,
+      "tokens_seen": 729808896
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007866313593323704,
+      "loss": 0.0754,
+      "theoretical_loss": 3.7918742875703657,
+      "tokens_seen": 730071040
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007865511153907881,
+      "loss": 0.0763,
+      "theoretical_loss": 3.7917381382384883,
+      "tokens_seen": 730333184
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007864708714492056,
+      "loss": 0.0769,
+      "theoretical_loss": 3.791602051444703,
+      "tokens_seen": 730595328
+    },
+    {
+      "epoch": 0.22,
+      "objective/train/advantage_avg": 0.0019119989592581987,
+      "objective/train/docs_used": 271086,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5188298225402832,
+      "objective/train/original_loss": 1.5188297033309937,
+      "objective/train/theoretical_loss": 3.7915340314836077,
+      "objective/train/tokens_used": 751186400,
+      "objective/train/value_avg": -0.00708770751953125,
+      "objective/train/value_loss": 0.00019582056847866625,
+      "objective/train/value_max": -0.00018966197967529297,
+      "objective/train/value_min": -0.70556640625,
+      "objective/train/value_reward_corr": 0.6300847957470939,
+      "objective/train/value_std": 0.0132293701171875,
+      "objective/train/weight_avg": 1.0020005702972412,
+      "objective/train/weighted_lm_loss": 1.5218416452407837,
+      "objective/train/weights_max": 2.00211238861084,
+      "objective/train/weights_min": 0.2432425171136856,
+      "theoretical_loss": 3.7915340314836077,
+      "tokens_seen": 730726400
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007863906275076232,
+      "loss": 0.0753,
+      "theoretical_loss": 3.791466027137859,
+      "tokens_seen": 730857472
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007863103835660408,
+      "loss": 0.0783,
+      "theoretical_loss": 3.7913300652668678,
+      "tokens_seen": 731119616
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007862301396244583,
+      "loss": 0.0779,
+      "theoretical_loss": 3.7911941657807002,
+      "tokens_seen": 731381760
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007861498956828759,
+      "loss": 0.0754,
+      "theoretical_loss": 3.7910583286283854,
+      "tokens_seen": 731643904
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007860696517412936,
+      "loss": 0.0786,
+      "theoretical_loss": 3.7909225537590157,
+      "tokens_seen": 731906048
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007859894077997112,
+      "loss": 0.0766,
+      "theoretical_loss": 3.790786841121739,
+      "tokens_seen": 732168192
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007859091638581287,
+      "loss": 0.0767,
+      "theoretical_loss": 3.7906511906657676,
+      "tokens_seen": 732430336
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007858289199165464,
+      "loss": 0.0766,
+      "theoretical_loss": 3.7905156023403697,
+      "tokens_seen": 732692480
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007857486759749639,
+      "loss": 0.0781,
+      "theoretical_loss": 3.7903800760948743,
+      "tokens_seen": 732954624
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007856684320333814,
+      "loss": 0.0778,
+      "theoretical_loss": 3.790244611878671,
+      "tokens_seen": 733216768
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007855881880917991,
+      "loss": 0.078,
+      "theoretical_loss": 3.790109209641206,
+      "tokens_seen": 733478912
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007855079441502166,
+      "loss": 0.0769,
+      "theoretical_loss": 3.7899738693319875,
+      "tokens_seen": 733741056
+    },
+    {
+      "epoch": 0.22,
+      "objective/train/advantage_avg": 0.0013161771930754185,
+      "objective/train/docs_used": 272255,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.631644606590271,
+      "objective/train/original_loss": 1.6316447257995605,
+      "objective/train/theoretical_loss": 3.7898385909005814,
+      "objective/train/tokens_used": 754463200,
+      "objective/train/value_avg": -0.006435394287109375,
+      "objective/train/value_loss": 0.0001979928492801264,
+      "objective/train/value_max": -0.00013446807861328125,
+      "objective/train/value_min": -0.2325439453125,
+      "objective/train/value_reward_corr": 0.5276408179266261,
+      "objective/train/value_std": 0.00835418701171875,
+      "objective/train/weight_avg": 1.0014052391052246,
+      "objective/train/weighted_lm_loss": 1.6336127519607544,
+      "objective/train/weights_max": 1.1831884384155273,
+      "objective/train/weights_min": 0.3687118589878082,
+      "theoretical_loss": 3.7898385909005814,
+      "tokens_seen": 734003200
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007854277002086343,
+      "loss": 0.0802,
+      "theoretical_loss": 3.7898385909005814,
+      "tokens_seen": 734003200
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007853474562670519,
+      "loss": 0.0776,
+      "theoretical_loss": 3.7897033742966135,
+      "tokens_seen": 734265344
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007852672123254695,
+      "loss": 0.0776,
+      "theoretical_loss": 3.789568219469767,
+      "tokens_seen": 734527488
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007851869683838871,
+      "loss": 0.08,
+      "theoretical_loss": 3.789433126369786,
+      "tokens_seen": 734789632
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007851067244423046,
+      "loss": 0.0767,
+      "theoretical_loss": 3.7892980949464716,
+      "tokens_seen": 735051776
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007850264805007222,
+      "loss": 0.0781,
+      "theoretical_loss": 3.7891631251496856,
+      "tokens_seen": 735313920
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007849462365591398,
+      "loss": 0.0775,
+      "theoretical_loss": 3.7890282169293465,
+      "tokens_seen": 735576064
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007848659926175574,
+      "loss": 0.0763,
+      "theoretical_loss": 3.7888933702354324,
+      "tokens_seen": 735838208
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007847857486759749,
+      "loss": 0.0796,
+      "theoretical_loss": 3.7887585850179786,
+      "tokens_seen": 736100352
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007847055047343926,
+      "loss": 0.0786,
+      "theoretical_loss": 3.788623861227081,
+      "tokens_seen": 736362496
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007846252607928101,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7884891988128926,
+      "tokens_seen": 736624640
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007845450168512277,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7883545977256228,
+      "tokens_seen": 736886784
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007844647729096454,
+      "loss": 0.0781,
+      "theoretical_loss": 3.7882200579155416,
+      "tokens_seen": 737148928
+    },
+    {
+      "epoch": 0.22,
+      "objective/train/advantage_avg": 0.0006363503634929657,
+      "objective/train/docs_used": 273341,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6099265813827515,
+      "objective/train/original_loss": 1.609926462173462,
+      "objective/train/theoretical_loss": 3.788152810973922,
+      "objective/train/tokens_used": 757740000,
+      "objective/train/value_avg": -0.0074005126953125,
+      "objective/train/value_loss": 0.00028310378547757864,
+      "objective/train/value_max": -0.0001596212387084961,
+      "objective/train/value_min": -0.5478515625,
+      "objective/train/value_reward_corr": 0.5414502644341082,
+      "objective/train/value_std": 0.0122528076171875,
+      "objective/train/weight_avg": 1.000755786895752,
+      "objective/train/weighted_lm_loss": 1.6105283498764038,
+      "objective/train/weights_max": 1.28411865234375,
+      "objective/train/weights_min": 0.3686028718948364,
+      "theoretical_loss": 3.788152810973922,
+      "tokens_seen": 737280000
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007843845289680629,
+      "loss": 0.0792,
+      "theoretical_loss": 3.788085579332977,
+      "tokens_seen": 737411072
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007843042850264806,
+      "loss": 0.0773,
+      "theoretical_loss": 3.787951161928312,
+      "tokens_seen": 737673216
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007842240410848981,
+      "loss": 0.0786,
+      "theoretical_loss": 3.7878168056519916,
+      "tokens_seen": 737935360
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007841437971433157,
+      "loss": 0.0767,
+      "theoretical_loss": 3.787682510454515,
+      "tokens_seen": 738197504
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007840635532017333,
+      "loss": 0.0779,
+      "theoretical_loss": 3.7875482762864405,
+      "tokens_seen": 738459648
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007839833092601508,
+      "loss": 0.0761,
+      "theoretical_loss": 3.787414103098384,
+      "tokens_seen": 738721792
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007839030653185684,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7872799908410193,
+      "tokens_seen": 738983936
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007838228213769861,
+      "loss": 0.0765,
+      "theoretical_loss": 3.787145939465076,
+      "tokens_seen": 739246080
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007837425774354037,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7870119489213425,
+      "tokens_seen": 739508224
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007836623334938212,
+      "loss": 0.0795,
+      "theoretical_loss": 3.786878019160664,
+      "tokens_seen": 739770368
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007835820895522389,
+      "loss": 0.0763,
+      "theoretical_loss": 3.7867441501339427,
+      "tokens_seen": 740032512
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007835018456106564,
+      "loss": 0.0783,
+      "theoretical_loss": 3.7866103417921373,
+      "tokens_seen": 740294656
+    },
+    {
+      "epoch": 0.22,
+      "objective/train/advantage_avg": 0.0014227567007765174,
+      "objective/train/docs_used": 274540,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.675157070159912,
+      "objective/train/original_loss": 1.6751569509506226,
+      "objective/train/theoretical_loss": 3.786476594086265,
+      "objective/train/tokens_used": 761016800,
+      "objective/train/value_avg": -0.00887298583984375,
+      "objective/train/value_loss": 0.0004343730106484145,
+      "objective/train/value_max": -0.00012731552124023438,
+      "objective/train/value_min": -0.580078125,
+      "objective/train/value_reward_corr": 0.5689424667790532,
+      "objective/train/value_std": 0.0146942138671875,
+      "objective/train/weight_avg": 1.0016205310821533,
+      "objective/train/weighted_lm_loss": 1.6775554418563843,
+      "objective/train/weights_max": 1.65679132938385,
+      "objective/train/weights_min": 0.3746204078197479,
+      "theoretical_loss": 3.786476594086265,
+      "tokens_seen": 740556800
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007834216016690739,
+      "loss": 0.0818,
+      "theoretical_loss": 3.786476594086265,
+      "tokens_seen": 740556800
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007833413577274916,
+      "loss": 0.0761,
+      "theoretical_loss": 3.7863429069673984,
+      "tokens_seen": 740818944
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007832611137859091,
+      "loss": 0.0774,
+      "theoretical_loss": 3.7862092803866663,
+      "tokens_seen": 741081088
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007831808698443267,
+      "loss": 0.0745,
+      "theoretical_loss": 3.786075714295257,
+      "tokens_seen": 741343232
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007831006259027444,
+      "loss": 0.0797,
+      "theoretical_loss": 3.7859422086444123,
+      "tokens_seen": 741605376
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.000783020381961162,
+      "loss": 0.078,
+      "theoretical_loss": 3.7858087633854325,
+      "tokens_seen": 741867520
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007829401380195796,
+      "loss": 0.0772,
+      "theoretical_loss": 3.785675378469673,
+      "tokens_seen": 742129664
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 0.0007828598940779972,
+      "loss": 0.0772,
+      "theoretical_loss": 3.7855420538485474,
+      "tokens_seen": 742391808
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007827796501364147,
+      "loss": 0.0787,
+      "theoretical_loss": 3.7854087894735233,
+      "tokens_seen": 742653952
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007826994061948323,
+      "loss": 0.0751,
+      "theoretical_loss": 3.7852755852961257,
+      "tokens_seen": 742916096
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007826191622532499,
+      "loss": 0.0791,
+      "theoretical_loss": 3.785142441267936,
+      "tokens_seen": 743178240
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007825389183116674,
+      "loss": 0.0778,
+      "theoretical_loss": 3.7850093573405905,
+      "tokens_seen": 743440384
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007824586743700851,
+      "loss": 0.076,
+      "theoretical_loss": 3.7848763334657827,
+      "tokens_seen": 743702528
+    },
+    {
+      "epoch": 0.23,
+      "objective/train/advantage_avg": 0.0006977500161156058,
+      "objective/train/docs_used": 275779,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6511344909667969,
+      "objective/train/original_loss": 1.6511344909667969,
+      "objective/train/theoretical_loss": 3.784809844033,
+      "objective/train/tokens_used": 764293600,
+      "objective/train/value_avg": -0.00855255126953125,
+      "objective/train/value_loss": 0.0002167196071241051,
+      "objective/train/value_max": -0.00013768672943115234,
+      "objective/train/value_min": -0.63427734375,
+      "objective/train/value_reward_corr": 0.6324548991472896,
+      "objective/train/value_std": 0.01303863525390625,
+      "objective/train/weight_avg": 1.0008044242858887,
+      "objective/train/weighted_lm_loss": 1.6532789468765259,
+      "objective/train/weights_max": 1.774770736694336,
+      "objective/train/weights_min": 0.6139143705368042,
+      "theoretical_loss": 3.784809844033,
+      "tokens_seen": 743833600
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007823784304285027,
+      "loss": 0.0807,
+      "theoretical_loss": 3.7847433695952617,
+      "tokens_seen": 743964672
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007822981864869203,
+      "loss": 0.0785,
+      "theoretical_loss": 3.7846104656808306,
+      "tokens_seen": 744226816
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007822179425453379,
+      "loss": 0.0759,
+      "theoretical_loss": 3.7844776216743505,
+      "tokens_seen": 744488960
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007821376986037554,
+      "loss": 0.0772,
+      "theoretical_loss": 3.784344837527737,
+      "tokens_seen": 744751104
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000782057454662173,
+      "loss": 0.0761,
+      "theoretical_loss": 3.784212113192961,
+      "tokens_seen": 745013248
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007819772107205906,
+      "loss": 0.0775,
+      "theoretical_loss": 3.7840794486220495,
+      "tokens_seen": 745275392
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007818969667790082,
+      "loss": 0.078,
+      "theoretical_loss": 3.783946843767084,
+      "tokens_seen": 745537536
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007818167228374257,
+      "loss": 0.0786,
+      "theoretical_loss": 3.783814298580203,
+      "tokens_seen": 745799680
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007817364788958434,
+      "loss": 0.0774,
+      "theoretical_loss": 3.7836818130135974,
+      "tokens_seen": 746061824
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000781656234954261,
+      "loss": 0.0759,
+      "theoretical_loss": 3.783549387019515,
+      "tokens_seen": 746323968
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007815759910126786,
+      "loss": 0.0781,
+      "theoretical_loss": 3.7834170205502584,
+      "tokens_seen": 746586112
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007814957470710962,
+      "loss": 0.0767,
+      "theoretical_loss": 3.783284713558186,
+      "tokens_seen": 746848256
+    },
+    {
+      "epoch": 0.23,
+      "objective/train/advantage_avg": 0.0004192414053250104,
+      "objective/train/docs_used": 277055,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5548205375671387,
+      "objective/train/original_loss": 1.5548205375671387,
+      "objective/train/theoretical_loss": 3.783152465995708,
+      "objective/train/tokens_used": 767570400,
+      "objective/train/value_avg": -0.0106964111328125,
+      "objective/train/value_loss": 0.0003498530713841319,
+      "objective/train/value_max": -0.0001881122589111328,
+      "objective/train/value_min": -0.37255859375,
+      "objective/train/value_reward_corr": 0.7141878653817549,
+      "objective/train/value_std": 0.01702880859375,
+      "objective/train/weight_avg": 1.0005861520767212,
+      "objective/train/weighted_lm_loss": 1.5560520887374878,
+      "objective/train/weights_max": 1.2006479501724243,
+      "objective/train/weights_min": 0.36900386214256287,
+      "theoretical_loss": 3.783152465995708,
+      "tokens_seen": 747110400
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007814155031295137,
+      "loss": 0.0792,
+      "theoretical_loss": 3.783152465995708,
+      "tokens_seen": 747110400
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007813352591879314,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7830202778152935,
+      "tokens_seen": 747372544
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007812550152463489,
+      "loss": 0.0751,
+      "theoretical_loss": 3.7828881489694632,
+      "tokens_seen": 747634688
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007811747713047665,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7827560794107926,
+      "tokens_seen": 747896832
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007810945273631841,
+      "loss": 0.0767,
+      "theoretical_loss": 3.7826240690919137,
+      "tokens_seen": 748158976
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007810142834216016,
+      "loss": 0.0792,
+      "theoretical_loss": 3.7824921179655115,
+      "tokens_seen": 748421120
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007809340394800192,
+      "loss": 0.0775,
+      "theoretical_loss": 3.782360225984325,
+      "tokens_seen": 748683264
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007808537955384369,
+      "loss": 0.0771,
+      "theoretical_loss": 3.782228393101149,
+      "tokens_seen": 748945408
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007807735515968545,
+      "loss": 0.0775,
+      "theoretical_loss": 3.78209661926883,
+      "tokens_seen": 749207552
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000780693307655272,
+      "loss": 0.0774,
+      "theoretical_loss": 3.781964904440271,
+      "tokens_seen": 749469696
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007806130637136897,
+      "loss": 0.0796,
+      "theoretical_loss": 3.7818332485684283,
+      "tokens_seen": 749731840
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007805328197721072,
+      "loss": 0.0772,
+      "theoretical_loss": 3.781701651606311,
+      "tokens_seen": 749993984
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007804525758305248,
+      "loss": 0.0795,
+      "theoretical_loss": 3.7815701135069846,
+      "tokens_seen": 750256128
+    },
+    {
+      "epoch": 0.23,
+      "objective/train/advantage_avg": 0.002483784221112728,
+      "objective/train/docs_used": 278273,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4896938800811768,
+      "objective/train/original_loss": 1.4896938800811768,
+      "objective/train/theoretical_loss": 3.7815043665162147,
+      "objective/train/tokens_used": 770847200,
+      "objective/train/value_avg": -0.00934600830078125,
+      "objective/train/value_loss": 0.00018718333740253001,
+      "objective/train/value_max": -0.00016736984252929688,
+      "objective/train/value_min": -0.252685546875,
+      "objective/train/value_reward_corr": 0.7167813323322847,
+      "objective/train/value_std": 0.014862060546875,
+      "objective/train/weight_avg": 1.002573013305664,
+      "objective/train/weighted_lm_loss": 1.4927213191986084,
+      "objective/train/weights_max": 1.2056413888931274,
+      "objective/train/weights_min": 0.3711751401424408,
+      "theoretical_loss": 3.7815043665162147,
+      "tokens_seen": 750387200
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007803723318889424,
+      "loss": 0.0767,
+      "theoretical_loss": 3.7814386342235653,
+      "tokens_seen": 750518272
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007802920879473599,
+      "loss": 0.0727,
+      "theoretical_loss": 3.7813072137092254,
+      "tokens_seen": 750780416
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007802118440057776,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7811758519171894,
+      "tokens_seen": 751042560
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007801316000641952,
+      "loss": 0.079,
+      "theoretical_loss": 3.781044548800736,
+      "tokens_seen": 751304704
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007800513561226128,
+      "loss": 0.0795,
+      "theoretical_loss": 3.7809133043131973,
+      "tokens_seen": 751566848
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007799711121810304,
+      "loss": 0.0786,
+      "theoretical_loss": 3.7807821184079584,
+      "tokens_seen": 751828992
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000779890868239448,
+      "loss": 0.0776,
+      "theoretical_loss": 3.780650991038459,
+      "tokens_seen": 752091136
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007798106242978655,
+      "loss": 0.0775,
+      "theoretical_loss": 3.7805199221581893,
+      "tokens_seen": 752353280
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007797303803562831,
+      "loss": 0.0797,
+      "theoretical_loss": 3.7803889117206957,
+      "tokens_seen": 752615424
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007796501364147007,
+      "loss": 0.079,
+      "theoretical_loss": 3.7802579596795756,
+      "tokens_seen": 752877568
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007795698924731182,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7801270659884807,
+      "tokens_seen": 753139712
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007794896485315359,
+      "loss": 0.078,
+      "theoretical_loss": 3.7799962306011143,
+      "tokens_seen": 753401856
+    },
+    {
+      "epoch": 0.23,
+      "objective/train/advantage_avg": 0.0004187932063359767,
+      "objective/train/docs_used": 279492,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5806915760040283,
+      "objective/train/original_loss": 1.5806913375854492,
+      "objective/train/theoretical_loss": 3.779865453471234,
+      "objective/train/tokens_used": 774124000,
+      "objective/train/value_avg": -0.01265716552734375,
+      "objective/train/value_loss": 0.00038256868720054626,
+      "objective/train/value_max": -0.00014770030975341797,
+      "objective/train/value_min": -0.55712890625,
+      "objective/train/value_reward_corr": 0.7937497238851523,
+      "objective/train/value_std": 0.0217437744140625,
+      "objective/train/weight_avg": 1.0005953311920166,
+      "objective/train/weighted_lm_loss": 1.5813276767730713,
+      "objective/train/weights_max": 1.2988438606262207,
+      "objective/train/weights_min": 0.3966088891029358,
+      "theoretical_loss": 3.779865453471234,
+      "tokens_seen": 753664000
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007794094045899534,
+      "loss": 0.0791,
+      "theoretical_loss": 3.779865453471234,
+      "tokens_seen": 753664000
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007793291606483711,
+      "loss": 0.0777,
+      "theoretical_loss": 3.7797347345526484,
+      "tokens_seen": 753926144
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007792489167067887,
+      "loss": 0.0777,
+      "theoretical_loss": 3.7796040737992205,
+      "tokens_seen": 754188288
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007791686727652062,
+      "loss": 0.0784,
+      "theoretical_loss": 3.7794734711648648,
+      "tokens_seen": 754450432
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007790884288236239,
+      "loss": 0.0792,
+      "theoretical_loss": 3.779342926603549,
+      "tokens_seen": 754712576
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007790081848820414,
+      "loss": 0.0769,
+      "theoretical_loss": 3.7792124400692924,
+      "tokens_seen": 754974720
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000778927940940459,
+      "loss": 0.0763,
+      "theoretical_loss": 3.7790820115161674,
+      "tokens_seen": 755236864
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007788476969988766,
+      "loss": 0.0792,
+      "theoretical_loss": 3.778951640898298,
+      "tokens_seen": 755499008
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007787674530572942,
+      "loss": 0.0781,
+      "theoretical_loss": 3.7788213281698617,
+      "tokens_seen": 755761152
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007786872091157117,
+      "loss": 0.0759,
+      "theoretical_loss": 3.778691073285086,
+      "tokens_seen": 756023296
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007786069651741294,
+      "loss": 0.0796,
+      "theoretical_loss": 3.7785608761982523,
+      "tokens_seen": 756285440
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000778526721232547,
+      "loss": 0.0765,
+      "theoretical_loss": 3.7784307368636934,
+      "tokens_seen": 756547584
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007784464772909645,
+      "loss": 0.0756,
+      "theoretical_loss": 3.7783006552357934,
+      "tokens_seen": 756809728
+    },
+    {
+      "epoch": 0.23,
+      "objective/train/advantage_avg": 0.0009907586500048637,
+      "objective/train/docs_used": 280659,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6565080881118774,
+      "objective/train/original_loss": 1.6565080881118774,
+      "objective/train/theoretical_loss": 3.7782356360476,
+      "objective/train/tokens_used": 777400800,
+      "objective/train/value_avg": -0.006893157958984375,
+      "objective/train/value_loss": 0.00026105440338142216,
+      "objective/train/value_max": -9.316205978393555e-05,
+      "objective/train/value_min": -0.25146484375,
+      "objective/train/value_reward_corr": 0.6604508877653548,
+      "objective/train/value_std": 0.011566162109375,
+      "objective/train/weight_avg": 1.0011118650436401,
+      "objective/train/weighted_lm_loss": 1.6582515239715576,
+      "objective/train/weights_max": 1.2315527200698853,
+      "objective/train/weights_min": 0.3999635875225067,
+      "theoretical_loss": 3.7782356360476,
+      "tokens_seen": 756940800
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007783662333493822,
+      "loss": 0.0793,
+      "theoretical_loss": 3.7781706312689893,
+      "tokens_seen": 757071872
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007782859894077997,
+      "loss": 0.078,
+      "theoretical_loss": 3.7780406649177696,
+      "tokens_seen": 757334016
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007782057454662173,
+      "loss": 0.0774,
+      "theoretical_loss": 3.777910756136673,
+      "tokens_seen": 757596160
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007781255015246349,
+      "loss": 0.0779,
+      "theoretical_loss": 3.777780904880292,
+      "tokens_seen": 757858304
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007780452575830524,
+      "loss": 0.0792,
+      "theoretical_loss": 3.777651111103269,
+      "tokens_seen": 758120448
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007779650136414702,
+      "loss": 0.0768,
+      "theoretical_loss": 3.777521374760298,
+      "tokens_seen": 758382592
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007778847696998877,
+      "loss": 0.0795,
+      "theoretical_loss": 3.7773916958061253,
+      "tokens_seen": 758644736
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007778045257583053,
+      "loss": 0.0812,
+      "theoretical_loss": 3.777262074195548,
+      "tokens_seen": 758906880
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007777242818167229,
+      "loss": 0.0791,
+      "theoretical_loss": 3.777132509883413,
+      "tokens_seen": 759169024
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007776440378751405,
+      "loss": 0.0789,
+      "theoretical_loss": 3.7770030028246215,
+      "tokens_seen": 759431168
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000777563793933558,
+      "loss": 0.078,
+      "theoretical_loss": 3.7768735529741226,
+      "tokens_seen": 759693312
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007774835499919756,
+      "loss": 0.0757,
+      "theoretical_loss": 3.776744160286918,
+      "tokens_seen": 759955456
+    },
+    {
+      "epoch": 0.23,
+      "objective/train/advantage_avg": -0.0008638743311166763,
+      "objective/train/docs_used": 281753,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.592660665512085,
+      "objective/train/original_loss": 1.592660665512085,
+      "objective/train/theoretical_loss": 3.77661482471806,
+      "objective/train/tokens_used": 780677600,
+      "objective/train/value_avg": -0.0077362060546875,
+      "objective/train/value_loss": 0.0008141609723679721,
+      "objective/train/value_max": -8.416175842285156e-05,
+      "objective/train/value_min": -0.5966796875,
+      "objective/train/value_reward_corr": 0.6082159732147249,
+      "objective/train/value_std": 0.015960693359375,
+      "objective/train/weight_avg": 0.9994794726371765,
+      "objective/train/weighted_lm_loss": 1.5920522212982178,
+      "objective/train/weights_max": 1.4571453332901,
+      "objective/train/weights_min": 0.26885083317756653,
+      "theoretical_loss": 3.77661482471806,
+      "tokens_seen": 760217600
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007774033060503932,
+      "loss": 0.0807,
+      "theoretical_loss": 3.77661482471806,
+      "tokens_seen": 760217600
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007773230621088107,
+      "loss": 0.0799,
+      "theoretical_loss": 3.776485546222651,
+      "tokens_seen": 760479744
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007772428181672284,
+      "loss": 0.0785,
+      "theoretical_loss": 3.776356324755847,
+      "tokens_seen": 760741888
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000777162574225646,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7762271602728497,
+      "tokens_seen": 761004032
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007770823302840636,
+      "loss": 0.0786,
+      "theoretical_loss": 3.7760980527289156,
+      "tokens_seen": 761266176
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007770020863424812,
+      "loss": 0.0756,
+      "theoretical_loss": 3.77596900207935,
+      "tokens_seen": 761528320
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007769218424008987,
+      "loss": 0.0763,
+      "theoretical_loss": 3.775840008279509,
+      "tokens_seen": 761790464
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007768415984593163,
+      "loss": 0.0788,
+      "theoretical_loss": 3.7757110712847997,
+      "tokens_seen": 762052608
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007767613545177339,
+      "loss": 0.0792,
+      "theoretical_loss": 3.775582191050678,
+      "tokens_seen": 762314752
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007766811105761515,
+      "loss": 0.0783,
+      "theoretical_loss": 3.775453367532651,
+      "tokens_seen": 762576896
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007766008666345691,
+      "loss": 0.0783,
+      "theoretical_loss": 3.775324600686276,
+      "tokens_seen": 762839040
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007765206226929867,
+      "loss": 0.0776,
+      "theoretical_loss": 3.7751958904671614,
+      "tokens_seen": 763101184
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007764403787514042,
+      "loss": 0.0786,
+      "theoretical_loss": 3.7750672368309623,
+      "tokens_seen": 763363328
+    },
+    {
+      "epoch": 0.23,
+      "objective/train/advantage_avg": 0.0011741745984181762,
+      "objective/train/docs_used": 282838,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.454331398010254,
+      "objective/train/original_loss": 1.454331398010254,
+      "objective/train/theoretical_loss": 3.775002931217613,
+      "objective/train/tokens_used": 783954400,
+      "objective/train/value_avg": -0.00695037841796875,
+      "objective/train/value_loss": 0.0001637246459722519,
+      "objective/train/value_max": -6.657838821411133e-05,
+      "objective/train/value_min": -0.275146484375,
+      "objective/train/value_reward_corr": 0.6056801602468862,
+      "objective/train/value_std": 0.01113128662109375,
+      "objective/train/weight_avg": 1.001247763633728,
+      "objective/train/weighted_lm_loss": 1.45583176612854,
+      "objective/train/weights_max": 1.1704035997390747,
+      "objective/train/weights_min": 0.3682715594768524,
+      "theoretical_loss": 3.775002931217613,
+      "tokens_seen": 763494400
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000776360134809822,
+      "loss": 0.0812,
+      "theoretical_loss": 3.7749386397333873,
+      "tokens_seen": 763625472
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007762798908682395,
+      "loss": 0.0781,
+      "theoretical_loss": 3.774810099130193,
+      "tokens_seen": 763887616
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000776199646926657,
+      "loss": 0.0786,
+      "theoretical_loss": 3.7746816149771862,
+      "tokens_seen": 764149760
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007761194029850747,
+      "loss": 0.0774,
+      "theoretical_loss": 3.7745531872302234,
+      "tokens_seen": 764411904
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007760391590434922,
+      "loss": 0.0782,
+      "theoretical_loss": 3.774424815845211,
+      "tokens_seen": 764674048
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007759589151019098,
+      "loss": 0.0781,
+      "theoretical_loss": 3.774296500778105,
+      "tokens_seen": 764936192
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007758786711603274,
+      "loss": 0.0784,
+      "theoretical_loss": 3.77416824198491,
+      "tokens_seen": 765198336
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000775798427218745,
+      "loss": 0.0764,
+      "theoretical_loss": 3.7740400394216813,
+      "tokens_seen": 765460480
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007757181832771625,
+      "loss": 0.0811,
+      "theoretical_loss": 3.7739118930445223,
+      "tokens_seen": 765722624
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007756379393355802,
+      "loss": 0.0799,
+      "theoretical_loss": 3.7737838028095867,
+      "tokens_seen": 765984768
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007755576953939978,
+      "loss": 0.0772,
+      "theoretical_loss": 3.773655768673077,
+      "tokens_seen": 766246912
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007754774514524153,
+      "loss": 0.0806,
+      "theoretical_loss": 3.7735277905912445,
+      "tokens_seen": 766509056
+    },
+    {
+      "epoch": 0.23,
+      "objective/train/advantage_avg": -0.0006871665827929974,
+      "objective/train/docs_used": 283988,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5568292140960693,
+      "objective/train/original_loss": 1.5568292140960693,
+      "objective/train/theoretical_loss": 3.773399868520391,
+      "objective/train/tokens_used": 787231200,
+      "objective/train/value_avg": -0.0089263916015625,
+      "objective/train/value_loss": 0.000296746235108003,
+      "objective/train/value_max": -0.00017535686492919922,
+      "objective/train/value_min": -0.2440185546875,
+      "objective/train/value_reward_corr": 0.6491560262447342,
+      "objective/train/value_std": 0.01305389404296875,
+      "objective/train/weight_avg": 0.9994507431983948,
+      "objective/train/weighted_lm_loss": 1.5556931495666504,
+      "objective/train/weights_max": 1.159029483795166,
+      "objective/train/weights_min": 0.37362140417099,
+      "theoretical_loss": 3.773399868520391,
+      "tokens_seen": 766771200
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000775397207510833,
+      "loss": 0.0784,
+      "theoretical_loss": 3.773399868520391,
+      "tokens_seen": 766771200
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007753169635692505,
+      "loss": 0.0784,
+      "theoretical_loss": 3.7732720024168644,
+      "tokens_seen": 767033344
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007752367196276682,
+      "loss": 0.0773,
+      "theoretical_loss": 3.773144192237065,
+      "tokens_seen": 767295488
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007751564756860857,
+      "loss": 0.0783,
+      "theoretical_loss": 3.7730164379374402,
+      "tokens_seen": 767557632
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007750762317445032,
+      "loss": 0.0777,
+      "theoretical_loss": 3.772888739474485,
+      "tokens_seen": 767819776
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000774995987802921,
+      "loss": 0.0769,
+      "theoretical_loss": 3.772761096804745,
+      "tokens_seen": 768081920
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007749157438613385,
+      "loss": 0.0777,
+      "theoretical_loss": 3.7726335098848143,
+      "tokens_seen": 768344064
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007748354999197561,
+      "loss": 0.0775,
+      "theoretical_loss": 3.772505978671334,
+      "tokens_seen": 768606208
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007747552559781737,
+      "loss": 0.0812,
+      "theoretical_loss": 3.772378503120996,
+      "tokens_seen": 768868352
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007746750120365913,
+      "loss": 0.0787,
+      "theoretical_loss": 3.7722510831905387,
+      "tokens_seen": 769130496
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007745947680950088,
+      "loss": 0.0793,
+      "theoretical_loss": 3.7721237188367494,
+      "tokens_seen": 769392640
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007745145241534264,
+      "loss": 0.0775,
+      "theoretical_loss": 3.771996410016464,
+      "tokens_seen": 769654784
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000774434280211844,
+      "loss": 0.0759,
+      "theoretical_loss": 3.7718691566865665,
+      "tokens_seen": 769916928
+    },
+    {
+      "epoch": 0.23,
+      "objective/train/advantage_avg": 0.001441980479285121,
+      "objective/train/docs_used": 285055,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.580965518951416,
+      "objective/train/original_loss": 1.580965518951416,
+      "objective/train/theoretical_loss": 3.7718055508170525,
+      "objective/train/tokens_used": 790508000,
+      "objective/train/value_avg": -0.00901031494140625,
+      "objective/train/value_loss": 0.00021419981203507632,
+      "objective/train/value_max": -0.00022876262664794922,
+      "objective/train/value_min": -0.2083740234375,
+      "objective/train/value_reward_corr": 0.6527358181867582,
+      "objective/train/value_std": 0.0116729736328125,
+      "objective/train/weight_avg": 1.0015414953231812,
+      "objective/train/weighted_lm_loss": 1.5829206705093384,
+      "objective/train/weights_max": 1.149869441986084,
+      "objective/train/weights_min": 0.4065597653388977,
+      "theoretical_loss": 3.7718055508170525,
+      "tokens_seen": 770048000
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007743540362702615,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7717419588039887,
+      "tokens_seen": 770179072
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007742737923286792,
+      "loss": 0.078,
+      "theoretical_loss": 3.7716148163257115,
+      "tokens_seen": 770441216
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007741935483870968,
+      "loss": 0.0749,
+      "theoretical_loss": 3.7714877292087623,
+      "tokens_seen": 770703360
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007741133044455145,
+      "loss": 0.0774,
+      "theoretical_loss": 3.7713606974102167,
+      "tokens_seen": 770965504
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000774033060503932,
+      "loss": 0.0808,
+      "theoretical_loss": 3.7712337208872,
+      "tokens_seen": 771227648
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007739528165623495,
+      "loss": 0.0765,
+      "theoretical_loss": 3.7711067995968826,
+      "tokens_seen": 771489792
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007738725726207672,
+      "loss": 0.0798,
+      "theoretical_loss": 3.770979933496485,
+      "tokens_seen": 771751936
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007737923286791847,
+      "loss": 0.0771,
+      "theoretical_loss": 3.770853122543274,
+      "tokens_seen": 772014080
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007737120847376023,
+      "loss": 0.0752,
+      "theoretical_loss": 3.770726366694564,
+      "tokens_seen": 772276224
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007736318407960199,
+      "loss": 0.0788,
+      "theoretical_loss": 3.7705996659077172,
+      "tokens_seen": 772538368
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007735515968544375,
+      "loss": 0.0766,
+      "theoretical_loss": 3.770473020140143,
+      "tokens_seen": 772800512
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000773471352912855,
+      "loss": 0.0817,
+      "theoretical_loss": 3.770346429349299,
+      "tokens_seen": 773062656
+    },
+    {
+      "epoch": 0.23,
+      "objective/train/advantage_avg": 0.0003448604547884315,
+      "objective/train/docs_used": 286255,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4529708623886108,
+      "objective/train/original_loss": 1.4529707431793213,
+      "objective/train/theoretical_loss": 3.7702198934926896,
+      "objective/train/tokens_used": 793784800,
+      "objective/train/value_avg": -0.009613037109375,
+      "objective/train/value_loss": 0.000516877043992281,
+      "objective/train/value_max": -0.0001233816146850586,
+      "objective/train/value_min": -0.37451171875,
+      "objective/train/value_reward_corr": 0.5522317075530438,
+      "objective/train/value_std": 0.016510009765625,
+      "objective/train/weight_avg": 1.0005626678466797,
+      "objective/train/weighted_lm_loss": 1.4538027048110962,
+      "objective/train/weights_max": 1.3192977905273438,
+      "objective/train/weights_min": 0.36989736557006836,
+      "theoretical_loss": 3.7702198934926896,
+      "tokens_seen": 773324800
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007733911089712728,
+      "loss": 0.0771,
+      "theoretical_loss": 3.7702198934926896,
+      "tokens_seen": 773324800
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007733108650296903,
+      "loss": 0.0767,
+      "theoretical_loss": 3.7700934125278653,
+      "tokens_seen": 773586944
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007732306210881078,
+      "loss": 0.0805,
+      "theoretical_loss": 3.7699669864124266,
+      "tokens_seen": 773849088
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007731503771465255,
+      "loss": 0.0776,
+      "theoretical_loss": 3.769840615104018,
+      "tokens_seen": 774111232
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.000773070133204943,
+      "loss": 0.0808,
+      "theoretical_loss": 3.7697142985603325,
+      "tokens_seen": 774373376
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007729898892633606,
+      "loss": 0.0797,
+      "theoretical_loss": 3.76958803673911,
+      "tokens_seen": 774635520
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007729096453217782,
+      "loss": 0.08,
+      "theoretical_loss": 3.7694618295981375,
+      "tokens_seen": 774897664
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007728294013801958,
+      "loss": 0.0779,
+      "theoretical_loss": 3.769335677095248,
+      "tokens_seen": 775159808
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 0.0007727491574386135,
+      "loss": 0.0805,
+      "theoretical_loss": 3.769209579188323,
+      "tokens_seen": 775421952
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000772668913497031,
+      "loss": 0.078,
+      "theoretical_loss": 3.7690835358352883,
+      "tokens_seen": 775684096
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007725886695554486,
+      "loss": 0.0805,
+      "theoretical_loss": 3.7689575469941183,
+      "tokens_seen": 775946240
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007725084256138662,
+      "loss": 0.0791,
+      "theoretical_loss": 3.768831612622833,
+      "tokens_seen": 776208384
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007724281816722838,
+      "loss": 0.0774,
+      "theoretical_loss": 3.7687057326794986,
+      "tokens_seen": 776470528
+    },
+    {
+      "epoch": 0.24,
+      "objective/train/advantage_avg": 0.0015384306898340583,
+      "objective/train/docs_used": 287303,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5866420269012451,
+      "objective/train/original_loss": 1.5866421461105347,
+      "objective/train/theoretical_loss": 3.768642813105222,
+      "objective/train/tokens_used": 797061600,
+      "objective/train/value_avg": -0.006908416748046875,
+      "objective/train/value_loss": 0.00025792099768295884,
+      "objective/train/value_max": -9.995698928833008e-05,
+      "objective/train/value_min": -0.2237548828125,
+      "objective/train/value_reward_corr": 0.42940502937583597,
+      "objective/train/value_std": 0.0088653564453125,
+      "objective/train/weight_avg": 1.0016469955444336,
+      "objective/train/weighted_lm_loss": 1.5896674394607544,
+      "objective/train/weights_max": 1.1481863260269165,
+      "objective/train/weights_min": 0.3687625229358673,
+      "theoretical_loss": 3.768642813105222,
+      "tokens_seen": 776601600
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007723479377307013,
+      "loss": 0.0805,
+      "theoretical_loss": 3.768579907122229,
+      "tokens_seen": 776732672
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000772267693789119,
+      "loss": 0.078,
+      "theoretical_loss": 3.768454135909183,
+      "tokens_seen": 776994816
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007721874498475365,
+      "loss": 0.078,
+      "theoretical_loss": 3.768328418998567,
+      "tokens_seen": 777256960
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000772107205905954,
+      "loss": 0.0778,
+      "theoretical_loss": 3.7682027563486327,
+      "tokens_seen": 777519104
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007720269619643718,
+      "loss": 0.0806,
+      "theoretical_loss": 3.768077147917678,
+      "tokens_seen": 777781248
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007719467180227893,
+      "loss": 0.0772,
+      "theoretical_loss": 3.7679515936640477,
+      "tokens_seen": 778043392
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007718664740812069,
+      "loss": 0.0758,
+      "theoretical_loss": 3.7678260935461316,
+      "tokens_seen": 778305536
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007717862301396245,
+      "loss": 0.0741,
+      "theoretical_loss": 3.767700647522366,
+      "tokens_seen": 778567680
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007717059861980421,
+      "loss": 0.0735,
+      "theoretical_loss": 3.7675752555512334,
+      "tokens_seen": 778829824
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007716257422564597,
+      "loss": 0.0783,
+      "theoretical_loss": 3.7674499175912617,
+      "tokens_seen": 779091968
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007715454983148772,
+      "loss": 0.0776,
+      "theoretical_loss": 3.767324633601024,
+      "tokens_seen": 779354112
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007714652543732948,
+      "loss": 0.0788,
+      "theoretical_loss": 3.7671994035391405,
+      "tokens_seen": 779616256
+    },
+    {
+      "epoch": 0.24,
+      "objective/train/advantage_avg": 0.0009066300117410719,
+      "objective/train/docs_used": 288522,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5356212854385376,
+      "objective/train/original_loss": 1.535621166229248,
+      "objective/train/theoretical_loss": 3.767074227364275,
+      "objective/train/tokens_used": 800338400,
+      "objective/train/value_avg": -0.005023956298828125,
+      "objective/train/value_loss": 0.0001440069027012214,
+      "objective/train/value_max": -0.00011414289474487305,
+      "objective/train/value_min": -0.1871337890625,
+      "objective/train/value_reward_corr": 0.5712434670773441,
+      "objective/train/value_std": 0.006877899169921875,
+      "objective/train/weight_avg": 1.0009735822677612,
+      "objective/train/weighted_lm_loss": 1.5371370315551758,
+      "objective/train/weights_max": 1.0837024450302124,
+      "objective/train/weights_min": 0.3689109683036804,
+      "theoretical_loss": 3.767074227364275,
+      "tokens_seen": 779878400
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007713850104317124,
+      "loss": 0.0805,
+      "theoretical_loss": 3.767074227364275,
+      "tokens_seen": 779878400
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.00077130476649013,
+      "loss": 0.0804,
+      "theoretical_loss": 3.7669491050351396,
+      "tokens_seen": 780140544
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007712245225485476,
+      "loss": 0.0793,
+      "theoretical_loss": 3.7668240365104895,
+      "tokens_seen": 780402688
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007711442786069653,
+      "loss": 0.08,
+      "theoretical_loss": 3.7666990217491265,
+      "tokens_seen": 780664832
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007710640346653828,
+      "loss": 0.0778,
+      "theoretical_loss": 3.7665740607098974,
+      "tokens_seen": 780926976
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007709837907238003,
+      "loss": 0.0771,
+      "theoretical_loss": 3.7664491533516946,
+      "tokens_seen": 781189120
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000770903546782218,
+      "loss": 0.0773,
+      "theoretical_loss": 3.766324299633455,
+      "tokens_seen": 781451264
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007708233028406355,
+      "loss": 0.0787,
+      "theoretical_loss": 3.766199499514162,
+      "tokens_seen": 781713408
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007707430588990531,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7660747529528424,
+      "tokens_seen": 781975552
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007706628149574707,
+      "loss": 0.079,
+      "theoretical_loss": 3.76595005990857,
+      "tokens_seen": 782237696
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007705825710158883,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7658254203404615,
+      "tokens_seen": 782499840
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007705023270743058,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7657008342076796,
+      "tokens_seen": 782761984
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007704220831327236,
+      "loss": 0.0761,
+      "theoretical_loss": 3.765576301469433,
+      "tokens_seen": 783024128
+    },
+    {
+      "epoch": 0.24,
+      "objective/train/advantage_avg": 0.0012172440765425563,
+      "objective/train/docs_used": 289672,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6650699377059937,
+      "objective/train/original_loss": 1.6650700569152832,
+      "objective/train/theoretical_loss": 3.7655140551105246,
+      "objective/train/tokens_used": 803615200,
+      "objective/train/value_avg": -0.0089111328125,
+      "objective/train/value_loss": 0.000253047764999792,
+      "objective/train/value_max": -0.00019872188568115234,
+      "objective/train/value_min": -0.36279296875,
+      "objective/train/value_reward_corr": 0.6359651215028251,
+      "objective/train/value_std": 0.0158538818359375,
+      "objective/train/weight_avg": 1.0013352632522583,
+      "objective/train/weighted_lm_loss": 1.6672289371490479,
+      "objective/train/weights_max": 1.2659153938293457,
+      "objective/train/weights_min": 0.3686739206314087,
+      "theoretical_loss": 3.7655140551105246,
+      "tokens_seen": 783155200
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007703418391911411,
+      "loss": 0.0775,
+      "theoretical_loss": 3.7654518220849726,
+      "tokens_seen": 783286272
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007702615952495587,
+      "loss": 0.0769,
+      "theoretical_loss": 3.7653273960135962,
+      "tokens_seen": 783548416
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007701813513079763,
+      "loss": 0.081,
+      "theoretical_loss": 3.765203023214645,
+      "tokens_seen": 783810560
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007701011073663938,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7650787036475055,
+      "tokens_seen": 784072704
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007700208634248115,
+      "loss": 0.0768,
+      "theoretical_loss": 3.764954437271608,
+      "tokens_seen": 784334848
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000769940619483229,
+      "loss": 0.0718,
+      "theoretical_loss": 3.7648302240464284,
+      "tokens_seen": 784596992
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007698603755416466,
+      "loss": 0.0764,
+      "theoretical_loss": 3.764706063931486,
+      "tokens_seen": 784859136
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007697801316000643,
+      "loss": 0.0786,
+      "theoretical_loss": 3.764581956886345,
+      "tokens_seen": 785121280
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007696998876584818,
+      "loss": 0.0778,
+      "theoretical_loss": 3.7644579028706135,
+      "tokens_seen": 785383424
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007696196437168994,
+      "loss": 0.0759,
+      "theoretical_loss": 3.764333901843944,
+      "tokens_seen": 785645568
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000769539399775317,
+      "loss": 0.0775,
+      "theoretical_loss": 3.764209953766033,
+      "tokens_seen": 785907712
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007694591558337346,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7640860585966207,
+      "tokens_seen": 786169856
+    },
+    {
+      "epoch": 0.24,
+      "objective/train/advantage_avg": 0.0007303394959308207,
+      "objective/train/docs_used": 290810,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4556148052215576,
+      "objective/train/original_loss": 1.4556148052215576,
+      "objective/train/theoretical_loss": 3.763962216295493,
+      "objective/train/tokens_used": 806892000,
+      "objective/train/value_avg": -0.0081787109375,
+      "objective/train/value_loss": 0.00028056377777829766,
+      "objective/train/value_max": -0.0001881122589111328,
+      "objective/train/value_min": -0.37158203125,
+      "objective/train/value_reward_corr": 0.6735432986137638,
+      "objective/train/value_std": 0.01500701904296875,
+      "objective/train/weight_avg": 1.0008543729782104,
+      "objective/train/weighted_lm_loss": 1.4567698240280151,
+      "objective/train/weights_max": 1.129281997680664,
+      "objective/train/weights_min": 0.36916786432266235,
+      "theoretical_loss": 3.763962216295493,
+      "tokens_seen": 786432000
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007693789118921521,
+      "loss": 0.0757,
+      "theoretical_loss": 3.763962216295493,
+      "tokens_seen": 786432000
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007692986679505698,
+      "loss": 0.0762,
+      "theoretical_loss": 3.7638384268224776,
+      "tokens_seen": 786694144
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007692184240089873,
+      "loss": 0.0775,
+      "theoretical_loss": 3.7637146901374474,
+      "tokens_seen": 786956288
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007691381800674048,
+      "loss": 0.0788,
+      "theoretical_loss": 3.7635910062003193,
+      "tokens_seen": 787218432
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007690579361258225,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7634673749710523,
+      "tokens_seen": 787480576
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.00076897769218424,
+      "loss": 0.0755,
+      "theoretical_loss": 3.763343796409651,
+      "tokens_seen": 787742720
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007688974482426578,
+      "loss": 0.0761,
+      "theoretical_loss": 3.7632202704761637,
+      "tokens_seen": 788004864
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007688172043010753,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7630967971306797,
+      "tokens_seen": 788267008
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007687369603594929,
+      "loss": 0.0768,
+      "theoretical_loss": 3.762973376333335,
+      "tokens_seen": 788529152
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007686567164179105,
+      "loss": 0.0777,
+      "theoretical_loss": 3.7628500080443077,
+      "tokens_seen": 788791296
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000768576472476328,
+      "loss": 0.0771,
+      "theoretical_loss": 3.7627266922238185,
+      "tokens_seen": 789053440
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007684962285347456,
+      "loss": 0.0769,
+      "theoretical_loss": 3.762603428832133,
+      "tokens_seen": 789315584
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007684159845931632,
+      "loss": 0.0751,
+      "theoretical_loss": 3.7624802178295584,
+      "tokens_seen": 789577728
+    },
+    {
+      "epoch": 0.24,
+      "objective/train/advantage_avg": 0.0018347672885283828,
+      "objective/train/docs_used": 292000,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7165638208389282,
+      "objective/train/original_loss": 1.7165637016296387,
+      "objective/train/theoretical_loss": 3.762418631961796,
+      "objective/train/tokens_used": 810168800,
+      "objective/train/value_avg": -0.00826263427734375,
+      "objective/train/value_loss": 9.183614747598767e-05,
+      "objective/train/value_max": -0.00014090538024902344,
+      "objective/train/value_min": -0.263671875,
+      "objective/train/value_reward_corr": 0.8270739313759928,
+      "objective/train/value_std": 0.01273345947265625,
+      "objective/train/weight_avg": 1.0018802881240845,
+      "objective/train/weighted_lm_loss": 1.7201040983200073,
+      "objective/train/weights_max": 1.1363343000411987,
+      "objective/train/weights_min": 0.8306474089622498,
+      "theoretical_loss": 3.762418631961796,
+      "tokens_seen": 789708800
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007683357406515808,
+      "loss": 0.078,
+      "theoretical_loss": 3.762357059176447,
+      "tokens_seen": 789839872
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007682554967099984,
+      "loss": 0.0776,
+      "theoretical_loss": 3.762233952833193,
+      "tokens_seen": 790102016
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007681752527684161,
+      "loss": 0.0799,
+      "theoretical_loss": 3.7621108987602336,
+      "tokens_seen": 790364160
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007680950088268336,
+      "loss": 0.0766,
+      "theoretical_loss": 3.76198789691805,
+      "tokens_seen": 790626304
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007680147648852511,
+      "loss": 0.0798,
+      "theoretical_loss": 3.7618649472671652,
+      "tokens_seen": 790888448
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007679345209436688,
+      "loss": 0.0783,
+      "theoretical_loss": 3.761742049768146,
+      "tokens_seen": 791150592
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007678542770020863,
+      "loss": 0.0789,
+      "theoretical_loss": 3.761619204381602,
+      "tokens_seen": 791412736
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000767774033060504,
+      "loss": 0.0753,
+      "theoretical_loss": 3.7614964110681846,
+      "tokens_seen": 791674880
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007676937891189215,
+      "loss": 0.0771,
+      "theoretical_loss": 3.761373669788589,
+      "tokens_seen": 791937024
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007676135451773391,
+      "loss": 0.0766,
+      "theoretical_loss": 3.7612509805035526,
+      "tokens_seen": 792199168
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007675333012357568,
+      "loss": 0.0751,
+      "theoretical_loss": 3.761128343173856,
+      "tokens_seen": 792461312
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007674530572941743,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7610057577603215,
+      "tokens_seen": 792723456
+    },
+    {
+      "epoch": 0.24,
+      "objective/train/advantage_avg": 0.00022212705516722053,
+      "objective/train/docs_used": 293128,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.473114252090454,
+      "objective/train/original_loss": 1.473114252090454,
+      "objective/train/theoretical_loss": 3.7608832242238144,
+      "objective/train/tokens_used": 813445600,
+      "objective/train/value_avg": -0.0098419189453125,
+      "objective/train/value_loss": 0.00027089385548606515,
+      "objective/train/value_max": -0.00020182132720947266,
+      "objective/train/value_min": -0.5673828125,
+      "objective/train/value_reward_corr": 0.6538193866153705,
+      "objective/train/value_std": 0.01425933837890625,
+      "objective/train/weight_avg": 1.0003442764282227,
+      "objective/train/weighted_lm_loss": 1.473488211631775,
+      "objective/train/weights_max": 1.16512930393219,
+      "objective/train/weights_min": 0.36934465169906616,
+      "theoretical_loss": 3.7608832242238144,
+      "tokens_seen": 792985600
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007673728133525919,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7608832242238144,
+      "tokens_seen": 792985600
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007672925694110095,
+      "loss": 0.0779,
+      "theoretical_loss": 3.7607607425252416,
+      "tokens_seen": 793247744
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007672123254694271,
+      "loss": 0.079,
+      "theoretical_loss": 3.7606383126255536,
+      "tokens_seen": 793509888
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007671320815278446,
+      "loss": 0.0794,
+      "theoretical_loss": 3.760515934485743,
+      "tokens_seen": 793772032
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007670518375862623,
+      "loss": 0.0777,
+      "theoretical_loss": 3.760393608066843,
+      "tokens_seen": 794034176
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007669715936446798,
+      "loss": 0.076,
+      "theoretical_loss": 3.760271333329932,
+      "tokens_seen": 794296320
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007668913497030974,
+      "loss": 0.078,
+      "theoretical_loss": 3.7601491102361275,
+      "tokens_seen": 794558464
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000766811105761515,
+      "loss": 0.0777,
+      "theoretical_loss": 3.7600269387465914,
+      "tokens_seen": 794820608
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007667308618199326,
+      "loss": 0.0781,
+      "theoretical_loss": 3.759904818822525,
+      "tokens_seen": 795082752
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007666506178783502,
+      "loss": 0.0789,
+      "theoretical_loss": 3.759782750425175,
+      "tokens_seen": 795344896
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007665703739367678,
+      "loss": 0.076,
+      "theoretical_loss": 3.759660733515826,
+      "tokens_seen": 795607040
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007664901299951854,
+      "loss": 0.0767,
+      "theoretical_loss": 3.7595387680558088,
+      "tokens_seen": 795869184
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000766409886053603,
+      "loss": 0.0792,
+      "theoretical_loss": 3.759416854006492,
+      "tokens_seen": 796131328
+    },
+    {
+      "epoch": 0.24,
+      "objective/train/advantage_avg": 0.001882580341771245,
+      "objective/train/docs_used": 294439,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6294305324554443,
+      "objective/train/original_loss": 1.6294307708740234,
+      "objective/train/theoretical_loss": 3.7593559162487864,
+      "objective/train/tokens_used": 816722400,
+      "objective/train/value_avg": -0.00921630859375,
+      "objective/train/value_loss": 0.00032186240423470736,
+      "objective/train/value_max": -0.00012636184692382812,
+      "objective/train/value_min": -0.79052734375,
+      "objective/train/value_reward_corr": 0.6324796724873081,
+      "objective/train/value_std": 0.0162353515625,
+      "objective/train/weight_avg": 1.0020285844802856,
+      "objective/train/weighted_lm_loss": 1.6332080364227295,
+      "objective/train/weights_max": 1.9955464601516724,
+      "objective/train/weights_min": 0.39590752124786377,
+      "theoretical_loss": 3.7593559162487864,
+      "tokens_seen": 796262400
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007663296421120206,
+      "loss": 0.0795,
+      "theoretical_loss": 3.7592949913292886,
+      "tokens_seen": 796393472
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007662493981704381,
+      "loss": 0.0787,
+      "theoretical_loss": 3.759173179985652,
+      "tokens_seen": 796655616
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007661691542288557,
+      "loss": 0.0749,
+      "theoretical_loss": 3.7590514199370775,
+      "tokens_seen": 796917760
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007660889102872733,
+      "loss": 0.0737,
+      "theoretical_loss": 3.758929711145101,
+      "tokens_seen": 797179904
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007660086663456909,
+      "loss": 0.0756,
+      "theoretical_loss": 3.758808053571302,
+      "tokens_seen": 797442048
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007659284224041086,
+      "loss": 0.0805,
+      "theoretical_loss": 3.7586864471772996,
+      "tokens_seen": 797704192
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007658481784625261,
+      "loss": 0.0772,
+      "theoretical_loss": 3.758564891924755,
+      "tokens_seen": 797966336
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007657679345209437,
+      "loss": 0.0794,
+      "theoretical_loss": 3.758443387775371,
+      "tokens_seen": 798228480
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007656876905793613,
+      "loss": 0.0764,
+      "theoretical_loss": 3.7583219346908905,
+      "tokens_seen": 798490624
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007656074466377788,
+      "loss": 0.0772,
+      "theoretical_loss": 3.758200532633099,
+      "tokens_seen": 798752768
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007655272026961964,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7580791815638213,
+      "tokens_seen": 799014912
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000765446958754614,
+      "loss": 0.0785,
+      "theoretical_loss": 3.7579578814449253,
+      "tokens_seen": 799277056
+    },
+    {
+      "epoch": 0.24,
+      "objective/train/advantage_avg": 0.0015031647635623813,
+      "objective/train/docs_used": 295724,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.440466046333313,
+      "objective/train/original_loss": 1.4404661655426025,
+      "objective/train/theoretical_loss": 3.7578366322383188,
+      "objective/train/tokens_used": 819999200,
+      "objective/train/value_avg": -0.005229949951171875,
+      "objective/train/value_loss": 0.00010367185313953087,
+      "objective/train/value_max": -0.0001398324966430664,
+      "objective/train/value_min": -0.90234375,
+      "objective/train/value_reward_corr": 0.7276015339903671,
+      "objective/train/value_std": 0.01129150390625,
+      "objective/train/weight_avg": 1.0015531778335571,
+      "objective/train/weighted_lm_loss": 1.4429603815078735,
+      "objective/train/weights_max": 1.3412903547286987,
+      "objective/train/weights_min": 0.5442660450935364,
+      "theoretical_loss": 3.7578366322383188,
+      "tokens_seen": 799539200
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007653667148130316,
+      "loss": 0.0781,
+      "theoretical_loss": 3.7578366322383188,
+      "tokens_seen": 799539200
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007652864708714491,
+      "loss": 0.0763,
+      "theoretical_loss": 3.7577154339059504,
+      "tokens_seen": 799801344
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007652062269298669,
+      "loss": 0.077,
+      "theoretical_loss": 3.7575942864098106,
+      "tokens_seen": 800063488
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007651259829882844,
+      "loss": 0.0758,
+      "theoretical_loss": 3.75747318971193,
+      "tokens_seen": 800325632
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.000765045739046702,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7573521437743795,
+      "tokens_seen": 800587776
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007649654951051196,
+      "loss": 0.0775,
+      "theoretical_loss": 3.7572311485592715,
+      "tokens_seen": 800849920
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007648852511635371,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7571102040287596,
+      "tokens_seen": 801112064
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007648050072219548,
+      "loss": 0.0804,
+      "theoretical_loss": 3.7569893101450367,
+      "tokens_seen": 801374208
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007647247632803723,
+      "loss": 0.0805,
+      "theoretical_loss": 3.756868466870337,
+      "tokens_seen": 801636352
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007646445193387899,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7567476741669346,
+      "tokens_seen": 801898496
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007645642753972076,
+      "loss": 0.078,
+      "theoretical_loss": 3.756626931997145,
+      "tokens_seen": 802160640
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007644840314556251,
+      "loss": 0.0746,
+      "theoretical_loss": 3.7565062403233234,
+      "tokens_seen": 802422784
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007644037875140427,
+      "loss": 0.0743,
+      "theoretical_loss": 3.7563855991078654,
+      "tokens_seen": 802684928
+    },
+    {
+      "epoch": 0.24,
+      "objective/train/advantage_avg": -0.0006523468182422221,
+      "objective/train/docs_used": 296852,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5035277605056763,
+      "objective/train/original_loss": 1.5035279989242554,
+      "objective/train/theoretical_loss": 3.7563252974102825,
+      "objective/train/tokens_used": 823276000,
+      "objective/train/value_avg": -0.00974273681640625,
+      "objective/train/value_loss": 0.00043232380994595587,
+      "objective/train/value_max": -0.0001823902130126953,
+      "objective/train/value_min": -0.93505859375,
+      "objective/train/value_reward_corr": 0.643231403068373,
+      "objective/train/value_std": 0.01873779296875,
+      "objective/train/weight_avg": 0.9995439648628235,
+      "objective/train/weighted_lm_loss": 1.5028263330459595,
+      "objective/train/weights_max": 2.1485886573791504,
+      "objective/train/weights_min": 0.3683769106864929,
+      "theoretical_loss": 3.7563252974102825,
+      "tokens_seen": 802816000
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007643235435724603,
+      "loss": 0.0728,
+      "theoretical_loss": 3.7562650083132074,
+      "tokens_seen": 802947072
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007642432996308779,
+      "loss": 0.0785,
+      "theoretical_loss": 3.756144467901825,
+      "tokens_seen": 803209216
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007641630556892954,
+      "loss": 0.0779,
+      "theoretical_loss": 3.756023977836235,
+      "tokens_seen": 803471360
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007640828117477131,
+      "loss": 0.0768,
+      "theoretical_loss": 3.755903538078994,
+      "tokens_seen": 803733504
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007640025678061306,
+      "loss": 0.0804,
+      "theoretical_loss": 3.7557831485926982,
+      "tokens_seen": 803995648
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007639223238645483,
+      "loss": 0.0787,
+      "theoretical_loss": 3.7556628093399835,
+      "tokens_seen": 804257792
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007638420799229659,
+      "loss": 0.0803,
+      "theoretical_loss": 3.7555425202835275,
+      "tokens_seen": 804519936
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007637618359813834,
+      "loss": 0.0738,
+      "theoretical_loss": 3.7554222813860463,
+      "tokens_seen": 804782080
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007636815920398011,
+      "loss": 0.0771,
+      "theoretical_loss": 3.7553020926102954,
+      "tokens_seen": 805044224
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007636013480982186,
+      "loss": 0.0785,
+      "theoretical_loss": 3.755181953919071,
+      "tokens_seen": 805306368
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007635211041566362,
+      "loss": 0.0768,
+      "theoretical_loss": 3.755061865275209,
+      "tokens_seen": 805568512
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007634408602150538,
+      "loss": 0.078,
+      "theoretical_loss": 3.754941826641584,
+      "tokens_seen": 805830656
+    },
+    {
+      "epoch": 0.24,
+      "objective/train/advantage_avg": 0.001965285511687398,
+      "objective/train/docs_used": 297892,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5463677644729614,
+      "objective/train/original_loss": 1.546367883682251,
+      "objective/train/theoretical_loss": 3.754821837981112,
+      "objective/train/tokens_used": 826552800,
+      "objective/train/value_avg": -0.005489349365234375,
+      "objective/train/value_loss": 7.032585563138127e-05,
+      "objective/train/value_max": -0.00011771917343139648,
+      "objective/train/value_min": -0.2242431640625,
+      "objective/train/value_reward_corr": 0.6211484433775776,
+      "objective/train/value_std": 0.007572174072265625,
+      "objective/train/weight_avg": 1.0019999742507935,
+      "objective/train/weighted_lm_loss": 1.5493464469909668,
+      "objective/train/weights_max": 1.1479060649871826,
+      "objective/train/weights_min": 0.7884321212768555,
+      "theoretical_loss": 3.754821837981112,
+      "tokens_seen": 806092800
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007633606162734714,
+      "loss": 0.0756,
+      "theoretical_loss": 3.754821837981112,
+      "tokens_seen": 806092800
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007632803723318889,
+      "loss": 0.0761,
+      "theoretical_loss": 3.7547018992567462,
+      "tokens_seen": 806354944
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007632001283903065,
+      "loss": 0.0792,
+      "theoretical_loss": 3.7545820104314815,
+      "tokens_seen": 806617088
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007631198844487241,
+      "loss": 0.0745,
+      "theoretical_loss": 3.7544621714683517,
+      "tokens_seen": 806879232
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007630396405071417,
+      "loss": 0.0765,
+      "theoretical_loss": 3.754342382330428,
+      "tokens_seen": 807141376
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007629593965655594,
+      "loss": 0.0749,
+      "theoretical_loss": 3.7542226429808236,
+      "tokens_seen": 807403520
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007628791526239769,
+      "loss": 0.0744,
+      "theoretical_loss": 3.7541029533826893,
+      "tokens_seen": 807665664
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007627989086823945,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7539833134992158,
+      "tokens_seen": 807927808
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007627186647408121,
+      "loss": 0.0784,
+      "theoretical_loss": 3.753863723293634,
+      "tokens_seen": 808189952
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 0.0007626384207992296,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7537441827292106,
+      "tokens_seen": 808452096
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007625581768576473,
+      "loss": 0.0773,
+      "theoretical_loss": 3.753624691769255,
+      "tokens_seen": 808714240
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007624779329160648,
+      "loss": 0.0793,
+      "theoretical_loss": 3.7535052503771142,
+      "tokens_seen": 808976384
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007623976889744824,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7533858585161735,
+      "tokens_seen": 809238528
+    },
+    {
+      "epoch": 0.25,
+      "objective/train/advantage_avg": 0.0006120595498941839,
+      "objective/train/docs_used": 298985,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5744049549102783,
+      "objective/train/original_loss": 1.5744049549102783,
+      "objective/train/theoretical_loss": 3.753326181148472,
+      "objective/train/tokens_used": 829829600,
+      "objective/train/value_avg": -0.0063629150390625,
+      "objective/train/value_loss": 0.00018711324082687497,
+      "objective/train/value_max": -8.153915405273438e-05,
+      "objective/train/value_min": -0.1954345703125,
+      "objective/train/value_reward_corr": 0.5172567162546362,
+      "objective/train/value_std": 0.0081634521484375,
+      "objective/train/weight_avg": 1.0006967782974243,
+      "objective/train/weighted_lm_loss": 1.575119972229004,
+      "objective/train/weights_max": 1.1109601259231567,
+      "objective/train/weights_min": 0.37364134192466736,
+      "theoretical_loss": 3.753326181148472,
+      "tokens_seen": 809369600
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007623174450329001,
+      "loss": 0.0797,
+      "theoretical_loss": 3.753266516149858,
+      "tokens_seen": 809500672
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007622372010913177,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7531472232416316,
+      "tokens_seen": 809762816
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007621569571497352,
+      "loss": 0.0739,
+      "theoretical_loss": 3.7530279797549957,
+      "tokens_seen": 810024960
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007620767132081528,
+      "loss": 0.075,
+      "theoretical_loss": 3.752908785653492,
+      "tokens_seen": 810287104
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007619964692665704,
+      "loss": 0.0758,
+      "theoretical_loss": 3.7527896409007004,
+      "tokens_seen": 810549248
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007619162253249879,
+      "loss": 0.0781,
+      "theoretical_loss": 3.7526705454602394,
+      "tokens_seen": 810811392
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007618359813834056,
+      "loss": 0.0772,
+      "theoretical_loss": 3.752551499295766,
+      "tokens_seen": 811073536
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007617557374418231,
+      "loss": 0.0767,
+      "theoretical_loss": 3.7524325023709757,
+      "tokens_seen": 811335680
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007616754935002407,
+      "loss": 0.0738,
+      "theoretical_loss": 3.7523135546496023,
+      "tokens_seen": 811597824
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007615952495586584,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7521946560954182,
+      "tokens_seen": 811859968
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007615150056170759,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7520758066722344,
+      "tokens_seen": 812122112
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007614347616754936,
+      "loss": 0.0751,
+      "theoretical_loss": 3.7519570063438996,
+      "tokens_seen": 812384256
+    },
+    {
+      "epoch": 0.25,
+      "objective/train/advantage_avg": 0.002464708173647523,
+      "objective/train/docs_used": 300102,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4953174591064453,
+      "objective/train/original_loss": 1.4953173398971558,
+      "objective/train/theoretical_loss": 3.7518382550743024,
+      "objective/train/tokens_used": 833106400,
+      "objective/train/value_avg": -0.0084686279296875,
+      "objective/train/value_loss": 0.00014236473361961544,
+      "objective/train/value_max": -0.0001323223114013672,
+      "objective/train/value_min": -0.236572265625,
+      "objective/train/value_reward_corr": 0.6225013810544121,
+      "objective/train/value_std": 0.01258087158203125,
+      "objective/train/weight_avg": 1.00253164768219,
+      "objective/train/weighted_lm_loss": 1.4988830089569092,
+      "objective/train/weights_max": 1.1583771705627441,
+      "objective/train/weights_min": 0.36844298243522644,
+      "theoretical_loss": 3.7518382550743024,
+      "tokens_seen": 812646400
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007613545177339111,
+      "loss": 0.0764,
+      "theoretical_loss": 3.7518382550743024,
+      "tokens_seen": 812646400
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007612742737923287,
+      "loss": 0.0778,
+      "theoretical_loss": 3.7517195528273666,
+      "tokens_seen": 812908544
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007611940298507463,
+      "loss": 0.0758,
+      "theoretical_loss": 3.751600899567057,
+      "tokens_seen": 813170688
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007611137859091639,
+      "loss": 0.0725,
+      "theoretical_loss": 3.7514822952573743,
+      "tokens_seen": 813432832
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007610335419675814,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7513637398623603,
+      "tokens_seen": 813694976
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007609532980259991,
+      "loss": 0.0759,
+      "theoretical_loss": 3.751245233346091,
+      "tokens_seen": 813957120
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007608730540844167,
+      "loss": 0.077,
+      "theoretical_loss": 3.7511267756726823,
+      "tokens_seen": 814219264
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007607928101428342,
+      "loss": 0.0786,
+      "theoretical_loss": 3.7510083668062886,
+      "tokens_seen": 814481408
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007607125662012519,
+      "loss": 0.075,
+      "theoretical_loss": 3.750890006711101,
+      "tokens_seen": 814743552
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007606323222596694,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7507716953513492,
+      "tokens_seen": 815005696
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000760552078318087,
+      "loss": 0.0787,
+      "theoretical_loss": 3.7506534326912995,
+      "tokens_seen": 815267840
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007604718343765046,
+      "loss": 0.0772,
+      "theoretical_loss": 3.7505352186952567,
+      "tokens_seen": 815529984
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007603915904349222,
+      "loss": 0.074,
+      "theoretical_loss": 3.7504170533275634,
+      "tokens_seen": 815792128
+    },
+    {
+      "epoch": 0.25,
+      "objective/train/advantage_avg": 0.0003943704068660736,
+      "objective/train/docs_used": 301260,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4938346147537231,
+      "objective/train/original_loss": 1.4938344955444336,
+      "objective/train/theoretical_loss": 3.7503579888682155,
+      "objective/train/tokens_used": 836383200,
+      "objective/train/value_avg": -0.00901031494140625,
+      "objective/train/value_loss": 0.0003670316655188799,
+      "objective/train/value_max": -0.0001080632209777832,
+      "objective/train/value_min": -0.38720703125,
+      "objective/train/value_reward_corr": 0.6623424603359667,
+      "objective/train/value_std": 0.0164642333984375,
+      "objective/train/weight_avg": 1.0005619525909424,
+      "objective/train/weighted_lm_loss": 1.4935811758041382,
+      "objective/train/weights_max": 1.3037108182907104,
+      "objective/train/weights_min": 0.36955323815345764,
+      "theoretical_loss": 3.7503579888682155,
+      "tokens_seen": 815923200
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007603113464933397,
+      "loss": 0.0733,
+      "theoretical_loss": 3.7502989365526,
+      "tokens_seen": 816054272
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007602311025517573,
+      "loss": 0.0764,
+      "theoretical_loss": 3.7501808683347826,
+      "tokens_seen": 816316416
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000760150858610175,
+      "loss": 0.0753,
+      "theoretical_loss": 3.7500628486385668,
+      "tokens_seen": 816578560
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007600706146685926,
+      "loss": 0.0747,
+      "theoretical_loss": 3.7499448774284447,
+      "tokens_seen": 816840704
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007599903707270102,
+      "loss": 0.0771,
+      "theoretical_loss": 3.749826954668946,
+      "tokens_seen": 817102848
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007599101267854277,
+      "loss": 0.0774,
+      "theoretical_loss": 3.7497090803246387,
+      "tokens_seen": 817364992
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007598298828438454,
+      "loss": 0.075,
+      "theoretical_loss": 3.7495912543601246,
+      "tokens_seen": 817627136
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007597496389022629,
+      "loss": 0.0778,
+      "theoretical_loss": 3.7494734767400475,
+      "tokens_seen": 817889280
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007596693949606804,
+      "loss": 0.0766,
+      "theoretical_loss": 3.7493557474290853,
+      "tokens_seen": 818151424
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007595891510190981,
+      "loss": 0.0745,
+      "theoretical_loss": 3.7492380663919533,
+      "tokens_seen": 818413568
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007595089070775156,
+      "loss": 0.0762,
+      "theoretical_loss": 3.7491204335934043,
+      "tokens_seen": 818675712
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007594286631359332,
+      "loss": 0.0761,
+      "theoretical_loss": 3.7490028489982286,
+      "tokens_seen": 818937856
+    },
+    {
+      "epoch": 0.25,
+      "objective/train/advantage_avg": 0.0018716322956606746,
+      "objective/train/docs_used": 302476,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5703623294830322,
+      "objective/train/original_loss": 1.5703623294830322,
+      "objective/train/theoretical_loss": 3.7488853125712525,
+      "objective/train/tokens_used": 839660000,
+      "objective/train/value_avg": -0.011810302734375,
+      "objective/train/value_loss": 0.0005407900898717344,
+      "objective/train/value_max": -0.00021660327911376953,
+      "objective/train/value_min": -0.86669921875,
+      "objective/train/value_reward_corr": 0.6016890575229221,
+      "objective/train/value_std": 0.01898193359375,
+      "objective/train/weight_avg": 1.0021065473556519,
+      "objective/train/weighted_lm_loss": 1.5734965801239014,
+      "objective/train/weights_max": 1.6662120819091797,
+      "objective/train/weights_min": 0.3839375674724579,
+      "theoretical_loss": 3.7488853125712525,
+      "tokens_seen": 819200000
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007593484191943509,
+      "loss": 0.073,
+      "theoretical_loss": 3.7488853125712525,
+      "tokens_seen": 819200000
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007592681752527685,
+      "loss": 0.0737,
+      "theoretical_loss": 3.7487678242773406,
+      "tokens_seen": 819462144
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000759187931311186,
+      "loss": 0.0774,
+      "theoretical_loss": 3.748650384081392,
+      "tokens_seen": 819724288
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007591076873696036,
+      "loss": 0.0769,
+      "theoretical_loss": 3.7485329919483448,
+      "tokens_seen": 819986432
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007590274434280212,
+      "loss": 0.0746,
+      "theoretical_loss": 3.7484156478431734,
+      "tokens_seen": 820248576
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007589471994864387,
+      "loss": 0.0751,
+      "theoretical_loss": 3.748298351730888,
+      "tokens_seen": 820510720
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007588669555448564,
+      "loss": 0.0756,
+      "theoretical_loss": 3.748181103576537,
+      "tokens_seen": 820772864
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007587867116032739,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7480639033452032,
+      "tokens_seen": 821035008
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007587064676616916,
+      "loss": 0.078,
+      "theoretical_loss": 3.747946751002009,
+      "tokens_seen": 821297152
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007586262237201092,
+      "loss": 0.0783,
+      "theoretical_loss": 3.747829646512109,
+      "tokens_seen": 821559296
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007585459797785267,
+      "loss": 0.0765,
+      "theoretical_loss": 3.747712589840699,
+      "tokens_seen": 821821440
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007584657358369444,
+      "loss": 0.075,
+      "theoretical_loss": 3.7475955809530084,
+      "tokens_seen": 822083584
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007583854918953619,
+      "loss": 0.0732,
+      "theoretical_loss": 3.747478619814303,
+      "tokens_seen": 822345728
+    },
+    {
+      "epoch": 0.25,
+      "objective/train/advantage_avg": 0.000402974575990811,
+      "objective/train/docs_used": 303734,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5402162075042725,
+      "objective/train/original_loss": 1.540216088294983,
+      "objective/train/theoretical_loss": 3.7474201571399757,
+      "objective/train/tokens_used": 842936800,
+      "objective/train/value_avg": -0.011871337890625,
+      "objective/train/value_loss": 0.00018918554997071624,
+      "objective/train/value_max": -0.00019872188568115234,
+      "objective/train/value_min": -0.302734375,
+      "objective/train/value_reward_corr": 0.9061185714934559,
+      "objective/train/value_std": 0.0229949951171875,
+      "objective/train/weight_avg": 1.000493049621582,
+      "objective/train/weighted_lm_loss": 1.5424861907958984,
+      "objective/train/weights_max": 1.3256474733352661,
+      "objective/train/weights_min": 0.37776029109954834,
+      "theoretical_loss": 3.7474201571399757,
+      "tokens_seen": 822476800
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007583052479537795,
+      "loss": 0.0791,
+      "theoretical_loss": 3.7473617063898863,
+      "tokens_seen": 822607872
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007582250040121971,
+      "loss": 0.0768,
+      "theoretical_loss": 3.747244840645097,
+      "tokens_seen": 822870016
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007581447600706147,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7471280225453096,
+      "tokens_seen": 823132160
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007580645161290322,
+      "loss": 0.0766,
+      "theoretical_loss": 3.747011252055936,
+      "tokens_seen": 823394304
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007579842721874498,
+      "loss": 0.075,
+      "theoretical_loss": 3.746894529142424,
+      "tokens_seen": 823656448
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007579040282458675,
+      "loss": 0.0764,
+      "theoretical_loss": 3.746777853770256,
+      "tokens_seen": 823918592
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000757823784304285,
+      "loss": 0.0752,
+      "theoretical_loss": 3.746661225904953,
+      "tokens_seen": 824180736
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007577435403627027,
+      "loss": 0.0776,
+      "theoretical_loss": 3.746544645512069,
+      "tokens_seen": 824442880
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007576632964211202,
+      "loss": 0.0754,
+      "theoretical_loss": 3.7464281125571963,
+      "tokens_seen": 824705024
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007575830524795379,
+      "loss": 0.0761,
+      "theoretical_loss": 3.7463116270059618,
+      "tokens_seen": 824967168
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007575028085379554,
+      "loss": 0.0751,
+      "theoretical_loss": 3.7461951888240286,
+      "tokens_seen": 825229312
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000757422564596373,
+      "loss": 0.077,
+      "theoretical_loss": 3.7460787979770958,
+      "tokens_seen": 825491456
+    },
+    {
+      "epoch": 0.25,
+      "objective/train/advantage_avg": 0.001048018573783338,
+      "objective/train/docs_used": 304950,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.441100835800171,
+      "objective/train/original_loss": 1.441100835800171,
+      "objective/train/theoretical_loss": 3.745962454430897,
+      "objective/train/tokens_used": 846213600,
+      "objective/train/value_avg": -0.00592803955078125,
+      "objective/train/value_loss": 0.00011939887917833403,
+      "objective/train/value_max": -0.00017404556274414062,
+      "objective/train/value_min": -0.1895751953125,
+      "objective/train/value_reward_corr": 0.6019524468909858,
+      "objective/train/value_std": 0.0082244873046875,
+      "objective/train/weight_avg": 1.0011035203933716,
+      "objective/train/weighted_lm_loss": 1.4430818557739258,
+      "objective/train/weights_max": 1.2087359428405762,
+      "objective/train/weights_min": 0.3807637691497803,
+      "theoretical_loss": 3.745962454430897,
+      "tokens_seen": 825753600
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007573423206547906,
+      "loss": 0.0766,
+      "theoretical_loss": 3.745962454430897,
+      "tokens_seen": 825753600
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007572620767132081,
+      "loss": 0.0787,
+      "theoretical_loss": 3.745846158151204,
+      "tokens_seen": 826015744
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007571818327716257,
+      "loss": 0.0766,
+      "theoretical_loss": 3.7457299091038214,
+      "tokens_seen": 826277888
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007571015888300434,
+      "loss": 0.0748,
+      "theoretical_loss": 3.745613707254591,
+      "tokens_seen": 826540032
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000757021344888461,
+      "loss": 0.0764,
+      "theoretical_loss": 3.7454975525693897,
+      "tokens_seen": 826802176
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007569411009468785,
+      "loss": 0.076,
+      "theoretical_loss": 3.7453814450141305,
+      "tokens_seen": 827064320
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007568608570052962,
+      "loss": 0.0808,
+      "theoretical_loss": 3.7452653845547603,
+      "tokens_seen": 827326464
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007567806130637137,
+      "loss": 0.08,
+      "theoretical_loss": 3.745149371157263,
+      "tokens_seen": 827588608
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007567003691221312,
+      "loss": 0.078,
+      "theoretical_loss": 3.7450334047876574,
+      "tokens_seen": 827850752
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007566201251805489,
+      "loss": 0.0779,
+      "theoretical_loss": 3.744917485411997,
+      "tokens_seen": 828112896
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007565398812389664,
+      "loss": 0.0803,
+      "theoretical_loss": 3.744801612996371,
+      "tokens_seen": 828375040
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000756459637297384,
+      "loss": 0.0767,
+      "theoretical_loss": 3.744685787506903,
+      "tokens_seen": 828637184
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007563793933558017,
+      "loss": 0.0776,
+      "theoretical_loss": 3.7445700089097533,
+      "tokens_seen": 828899328
+    },
+    {
+      "epoch": 0.25,
+      "objective/train/advantage_avg": 0.0003027545753866434,
+      "objective/train/docs_used": 306096,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4149469137191772,
+      "objective/train/original_loss": 1.4149470329284668,
+      "objective/train/theoretical_loss": 3.7445121371852323,
+      "objective/train/tokens_used": 849490400,
+      "objective/train/value_avg": -0.00829315185546875,
+      "objective/train/value_loss": 0.00021588837262243032,
+      "objective/train/value_max": -0.0001926422119140625,
+      "objective/train/value_min": -0.2091064453125,
+      "objective/train/value_reward_corr": 0.6806189676548946,
+      "objective/train/value_std": 0.013153076171875,
+      "objective/train/weight_avg": 1.0003987550735474,
+      "objective/train/weighted_lm_loss": 1.4160531759262085,
+      "objective/train/weights_max": 1.1402249336242676,
+      "objective/train/weights_min": 0.3719916045665741,
+      "theoretical_loss": 3.7445121371852323,
+      "tokens_seen": 829030400
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007562991494142193,
+      "loss": 0.0764,
+      "theoretical_loss": 3.7444542771711165,
+      "tokens_seen": 829161472
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007562189054726369,
+      "loss": 0.0747,
+      "theoretical_loss": 3.744338592257222,
+      "tokens_seen": 829423616
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007561386615310544,
+      "loss": 0.0762,
+      "theoretical_loss": 3.744222954134334,
+      "tokens_seen": 829685760
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000756058417589472,
+      "loss": 0.0747,
+      "theoretical_loss": 3.7441073627687524,
+      "tokens_seen": 829947904
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007559781736478896,
+      "loss": 0.0769,
+      "theoretical_loss": 3.743991818126812,
+      "tokens_seen": 830210048
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007558979297063072,
+      "loss": 0.0764,
+      "theoretical_loss": 3.7438763201748815,
+      "tokens_seen": 830472192
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007558176857647247,
+      "loss": 0.0766,
+      "theoretical_loss": 3.743760868879365,
+      "tokens_seen": 830734336
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007557374418231424,
+      "loss": 0.075,
+      "theoretical_loss": 3.743645464206702,
+      "tokens_seen": 830996480
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.00075565719788156,
+      "loss": 0.0759,
+      "theoretical_loss": 3.743530106123365,
+      "tokens_seen": 831258624
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007555769539399775,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7434147945958642,
+      "tokens_seen": 831520768
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007554967099983952,
+      "loss": 0.0767,
+      "theoretical_loss": 3.7432995295907405,
+      "tokens_seen": 831782912
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007554164660568127,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7431843110745726,
+      "tokens_seen": 832045056
+    },
+    {
+      "epoch": 0.25,
+      "objective/train/advantage_avg": 0.0011427036952227354,
+      "objective/train/docs_used": 307306,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.53184974193573,
+      "objective/train/original_loss": 1.5318498611450195,
+      "objective/train/theoretical_loss": 3.743069139013972,
+      "objective/train/tokens_used": 852767200,
+      "objective/train/value_avg": -0.0099029541015625,
+      "objective/train/value_loss": 0.0002651831309776753,
+      "objective/train/value_max": -0.0002551078796386719,
+      "objective/train/value_min": -0.54150390625,
+      "objective/train/value_reward_corr": 0.6319940300493742,
+      "objective/train/value_std": 0.01386260986328125,
+      "objective/train/weight_avg": 1.0012651681900024,
+      "objective/train/weighted_lm_loss": 1.5327504873275757,
+      "objective/train/weights_max": 1.5147995948791504,
+      "objective/train/weights_min": 0.38184595108032227,
+      "theoretical_loss": 3.743069139013972,
+      "tokens_seen": 832307200
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007553362221152303,
+      "loss": 0.0752,
+      "theoretical_loss": 3.743069139013972,
+      "tokens_seen": 832307200
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007552559781736479,
+      "loss": 0.0763,
+      "theoretical_loss": 3.742954013375586,
+      "tokens_seen": 832569344
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007551757342320655,
+      "loss": 0.0769,
+      "theoretical_loss": 3.742838934126094,
+      "tokens_seen": 832831488
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000755095490290483,
+      "loss": 0.0764,
+      "theoretical_loss": 3.742723901232213,
+      "tokens_seen": 833093632
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007550152463489006,
+      "loss": 0.0767,
+      "theoretical_loss": 3.742608914660692,
+      "tokens_seen": 833355776
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007549350024073182,
+      "loss": 0.0781,
+      "theoretical_loss": 3.742493974378314,
+      "tokens_seen": 833617920
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007548547584657359,
+      "loss": 0.0753,
+      "theoretical_loss": 3.742379080351899,
+      "tokens_seen": 833880064
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007547745145241535,
+      "loss": 0.079,
+      "theoretical_loss": 3.7422642325482975,
+      "tokens_seen": 834142208
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000754694270582571,
+      "loss": 0.081,
+      "theoretical_loss": 3.742149430934398,
+      "tokens_seen": 834404352
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007546140266409887,
+      "loss": 0.0807,
+      "theoretical_loss": 3.7420346754771208,
+      "tokens_seen": 834666496
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007545337826994062,
+      "loss": 0.0763,
+      "theoretical_loss": 3.7419199661434197,
+      "tokens_seen": 834928640
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007544535387578238,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7418053029002842,
+      "tokens_seen": 835190784
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007543732948162414,
+      "loss": 0.074,
+      "theoretical_loss": 3.7416906857147367,
+      "tokens_seen": 835452928
+    },
+    {
+      "epoch": 0.25,
+      "objective/train/advantage_avg": 4.4281885493546724e-05,
+      "objective/train/docs_used": 308483,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.479179859161377,
+      "objective/train/original_loss": 1.479179859161377,
+      "objective/train/theoretical_loss": 3.741633394383263,
+      "objective/train/tokens_used": 856044000,
+      "objective/train/value_avg": -0.01134490966796875,
+      "objective/train/value_loss": 0.0005562845035456121,
+      "objective/train/value_max": -0.00019562244415283203,
+      "objective/train/value_min": -0.66552734375,
+      "objective/train/value_reward_corr": 0.6574823323493614,
+      "objective/train/value_std": 0.0205535888671875,
+      "objective/train/weight_avg": 1.0002927780151367,
+      "objective/train/weighted_lm_loss": 1.477531909942627,
+      "objective/train/weights_max": 1.606209397315979,
+      "objective/train/weights_min": 0.37909993529319763,
+      "theoretical_loss": 3.741633394383263,
+      "tokens_seen": 835584000
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007542930508746589,
+      "loss": 0.0756,
+      "theoretical_loss": 3.741576114553835,
+      "tokens_seen": 835715072
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007542128069330765,
+      "loss": 0.0778,
+      "theoretical_loss": 3.7414615893846683,
+      "tokens_seen": 835977216
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007541325629914942,
+      "loss": 0.0768,
+      "theoretical_loss": 3.741347110174362,
+      "tokens_seen": 836239360
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007540523190499118,
+      "loss": 0.0776,
+      "theoretical_loss": 3.741232676890074,
+      "tokens_seen": 836501504
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007539720751083293,
+      "loss": 0.0785,
+      "theoretical_loss": 3.7411182894989965,
+      "tokens_seen": 836763648
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000753891831166747,
+      "loss": 0.078,
+      "theoretical_loss": 3.7410039479683546,
+      "tokens_seen": 837025792
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007538115872251645,
+      "loss": 0.0758,
+      "theoretical_loss": 3.740889652265408,
+      "tokens_seen": 837287936
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007537313432835821,
+      "loss": 0.0775,
+      "theoretical_loss": 3.7407754023574507,
+      "tokens_seen": 837550080
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007536510993419997,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7406611982118076,
+      "tokens_seen": 837812224
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007535708554004172,
+      "loss": 0.0768,
+      "theoretical_loss": 3.74054703979584,
+      "tokens_seen": 838074368
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000753490611458835,
+      "loss": 0.0783,
+      "theoretical_loss": 3.7404329270769403,
+      "tokens_seen": 838336512
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007534103675172525,
+      "loss": 0.076,
+      "theoretical_loss": 3.740318860022537,
+      "tokens_seen": 838598656
+    },
+    {
+      "epoch": 0.25,
+      "objective/train/advantage_avg": 0.0013807304203510284,
+      "objective/train/docs_used": 309674,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6753649711608887,
+      "objective/train/original_loss": 1.6753649711608887,
+      "objective/train/theoretical_loss": 3.7402048386000892,
+      "objective/train/tokens_used": 859320800,
+      "objective/train/value_avg": -0.01296234130859375,
+      "objective/train/value_loss": 0.0006226726691238582,
+      "objective/train/value_max": -0.0002065896987915039,
+      "objective/train/value_min": -0.6162109375,
+      "objective/train/value_reward_corr": 0.639937529812378,
+      "objective/train/value_std": 0.019622802734375,
+      "objective/train/weight_avg": 1.001650094985962,
+      "objective/train/weighted_lm_loss": 1.6781058311462402,
+      "objective/train/weights_max": 1.614859700202942,
+      "objective/train/weights_min": 0.37846994400024414,
+      "theoretical_loss": 3.7402048386000892,
+      "tokens_seen": 838860800
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007533301235756701,
+      "loss": 0.0754,
+      "theoretical_loss": 3.7402048386000892,
+      "tokens_seen": 838860800
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007532498796340877,
+      "loss": 0.0743,
+      "theoretical_loss": 3.740090862777091,
+      "tokens_seen": 839122944
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007531696356925052,
+      "loss": 0.0779,
+      "theoretical_loss": 3.7399769325210697,
+      "tokens_seen": 839385088
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007530893917509228,
+      "loss": 0.0771,
+      "theoretical_loss": 3.7398630477995853,
+      "tokens_seen": 839647232
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007530091478093404,
+      "loss": 0.0751,
+      "theoretical_loss": 3.7397492085802315,
+      "tokens_seen": 839909376
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.000752928903867758,
+      "loss": 0.0765,
+      "theoretical_loss": 3.739635414830635,
+      "tokens_seen": 840171520
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007528486599261755,
+      "loss": 0.0763,
+      "theoretical_loss": 3.7395216665184554,
+      "tokens_seen": 840433664
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007527684159845932,
+      "loss": 0.0793,
+      "theoretical_loss": 3.739407963611386,
+      "tokens_seen": 840695808
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007526881720430108,
+      "loss": 0.074,
+      "theoretical_loss": 3.739294306077152,
+      "tokens_seen": 840957952
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 0.0007526079281014283,
+      "loss": 0.0785,
+      "theoretical_loss": 3.7391806938835126,
+      "tokens_seen": 841220096
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000752527684159846,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7390671269982603,
+      "tokens_seen": 841482240
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007524474402182635,
+      "loss": 0.0764,
+      "theoretical_loss": 3.7389536053892187,
+      "tokens_seen": 841744384
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007523671962766812,
+      "loss": 0.0744,
+      "theoretical_loss": 3.738840129024246,
+      "tokens_seen": 842006528
+    },
+    {
+      "epoch": 0.26,
+      "objective/train/advantage_avg": 0.002575838938355446,
+      "objective/train/docs_used": 310992,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4314990043640137,
+      "objective/train/original_loss": 1.4314990043640137,
+      "objective/train/theoretical_loss": 3.73878340779825,
+      "objective/train/tokens_used": 862597600,
+      "objective/train/value_avg": -0.00510406494140625,
+      "objective/train/value_loss": 6.991349073359743e-05,
+      "objective/train/value_max": -0.00016605854034423828,
+      "objective/train/value_min": -0.186279296875,
+      "objective/train/value_reward_corr": 0.4218440665138546,
+      "objective/train/value_std": 0.005443572998046875,
+      "objective/train/weight_avg": 1.0026094913482666,
+      "objective/train/weighted_lm_loss": 1.4359185695648193,
+      "objective/train/weights_max": 1.1117740869522095,
+      "objective/train/weights_min": 0.5176728963851929,
+      "theoretical_loss": 3.73878340779825,
+      "tokens_seen": 842137600
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007522869523350987,
+      "loss": 0.0771,
+      "theoretical_loss": 3.738726697871233,
+      "tokens_seen": 842268672
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007522067083935163,
+      "loss": 0.0796,
+      "theoretical_loss": 3.738613311898103,
+      "tokens_seen": 842530816
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007521264644519339,
+      "loss": 0.0764,
+      "theoretical_loss": 3.7384999710728106,
+      "tokens_seen": 842792960
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007520462205103514,
+      "loss": 0.0769,
+      "theoretical_loss": 3.738386675363346,
+      "tokens_seen": 843055104
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000751965976568769,
+      "loss": 0.0764,
+      "theoretical_loss": 3.738273424737729,
+      "tokens_seen": 843317248
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007518857326271867,
+      "loss": 0.0792,
+      "theoretical_loss": 3.7381602191640146,
+      "tokens_seen": 843579392
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007518054886856043,
+      "loss": 0.077,
+      "theoretical_loss": 3.738047058610289,
+      "tokens_seen": 843841536
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007517252447440218,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7379339430446707,
+      "tokens_seen": 844103680
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007516450008024395,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7378208724353117,
+      "tokens_seen": 844365824
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000751564756860857,
+      "loss": 0.0776,
+      "theoretical_loss": 3.7377078467503955,
+      "tokens_seen": 844627968
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007514845129192746,
+      "loss": 0.0782,
+      "theoretical_loss": 3.737594865958138,
+      "tokens_seen": 844890112
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007514042689776922,
+      "loss": 0.076,
+      "theoretical_loss": 3.7374819300267883,
+      "tokens_seen": 845152256
+    },
+    {
+      "epoch": 0.26,
+      "objective/train/advantage_avg": 0.000701575365383178,
+      "objective/train/docs_used": 312167,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.512547492980957,
+      "objective/train/original_loss": 1.512547492980957,
+      "objective/train/theoretical_loss": 3.7373690389246272,
+      "objective/train/tokens_used": 865874400,
+      "objective/train/value_avg": -0.00689697265625,
+      "objective/train/value_loss": 0.00012069241347489879,
+      "objective/train/value_max": -0.00014317035675048828,
+      "objective/train/value_min": -0.318115234375,
+      "objective/train/value_reward_corr": 0.6602302605762631,
+      "objective/train/value_std": 0.01007843017578125,
+      "objective/train/weight_avg": 1.0007611513137817,
+      "objective/train/weighted_lm_loss": 1.5141348838806152,
+      "objective/train/weights_max": 1.115106463432312,
+      "objective/train/weights_min": 0.8200203776359558,
+      "theoretical_loss": 3.7373690389246272,
+      "tokens_seen": 845414400
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007513240250361097,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7373690389246272,
+      "tokens_seen": 845414400
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007512437810945275,
+      "loss": 0.0783,
+      "theoretical_loss": 3.737256192619967,
+      "tokens_seen": 845676544
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000751163537152945,
+      "loss": 0.0753,
+      "theoretical_loss": 3.737143391081154,
+      "tokens_seen": 845938688
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007510832932113626,
+      "loss": 0.0762,
+      "theoretical_loss": 3.7370306342765653,
+      "tokens_seen": 846200832
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007510030492697802,
+      "loss": 0.0737,
+      "theoretical_loss": 3.73691792217461,
+      "tokens_seen": 846462976
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007509228053281978,
+      "loss": 0.0779,
+      "theoretical_loss": 3.7368052547437305,
+      "tokens_seen": 846725120
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007508425613866153,
+      "loss": 0.0772,
+      "theoretical_loss": 3.7366926319524003,
+      "tokens_seen": 846987264
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007507623174450329,
+      "loss": 0.0747,
+      "theoretical_loss": 3.736580053769125,
+      "tokens_seen": 847249408
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007506820735034505,
+      "loss": 0.0728,
+      "theoretical_loss": 3.736467520162442,
+      "tokens_seen": 847511552
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000750601829561868,
+      "loss": 0.0772,
+      "theoretical_loss": 3.736355031100922,
+      "tokens_seen": 847773696
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007505215856202858,
+      "loss": 0.0765,
+      "theoretical_loss": 3.7362425865531654,
+      "tokens_seen": 848035840
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007504413416787033,
+      "loss": 0.0773,
+      "theoretical_loss": 3.736130186487806,
+      "tokens_seen": 848297984
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007503610977371209,
+      "loss": 0.0749,
+      "theoretical_loss": 3.736017830873508,
+      "tokens_seen": 848560128
+    },
+    {
+      "epoch": 0.26,
+      "objective/train/advantage_avg": 0.0018955293344333768,
+      "objective/train/docs_used": 313372,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.561277151107788,
+      "objective/train/original_loss": 1.561277151107788,
+      "objective/train/theoretical_loss": 3.7359616697257243,
+      "objective/train/tokens_used": 869151200,
+      "objective/train/value_avg": -0.00714111328125,
+      "objective/train/value_loss": 0.00045251031406223774,
+      "objective/train/value_max": -0.00013136863708496094,
+      "objective/train/value_min": -0.984375,
+      "objective/train/value_reward_corr": 0.6359393678379087,
+      "objective/train/value_std": 0.0159912109375,
+      "objective/train/weight_avg": 1.0020655393600464,
+      "objective/train/weighted_lm_loss": 1.5643846988677979,
+      "objective/train/weights_max": 2.523839235305786,
+      "objective/train/weights_min": 0.07859423011541367,
+      "theoretical_loss": 3.7359616697257243,
+      "tokens_seen": 848691200
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007502808537955385,
+      "loss": 0.0772,
+      "theoretical_loss": 3.7359055196789694,
+      "tokens_seen": 848822272
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000750200609853956,
+      "loss": 0.075,
+      "theoretical_loss": 3.7357932528729183,
+      "tokens_seen": 849084416
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007501203659123736,
+      "loss": 0.0787,
+      "theoretical_loss": 3.7356810304241144,
+      "tokens_seen": 849346560
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007500401219707912,
+      "loss": 0.0783,
+      "theoretical_loss": 3.73556885230135,
+      "tokens_seen": 849608704
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007499598780292088,
+      "loss": 0.0756,
+      "theoretical_loss": 3.735456718473449,
+      "tokens_seen": 849870848
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007498796340876264,
+      "loss": 0.0718,
+      "theoretical_loss": 3.7353446289092647,
+      "tokens_seen": 850132992
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000749799390146044,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7352325835776856,
+      "tokens_seen": 850395136
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007497191462044616,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7351205824476277,
+      "tokens_seen": 850657280
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007496389022628792,
+      "loss": 0.0791,
+      "theoretical_loss": 3.7350086254880415,
+      "tokens_seen": 850919424
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007495586583212968,
+      "loss": 0.0743,
+      "theoretical_loss": 3.734896712667907,
+      "tokens_seen": 851181568
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007494784143797143,
+      "loss": 0.0754,
+      "theoretical_loss": 3.734784843956236,
+      "tokens_seen": 851443712
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000749398170438132,
+      "loss": 0.0747,
+      "theoretical_loss": 3.7346730193220727,
+      "tokens_seen": 851705856
+    },
+    {
+      "epoch": 0.26,
+      "objective/train/advantage_avg": 0.0007218181854113936,
+      "objective/train/docs_used": 314587,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5503199100494385,
+      "objective/train/original_loss": 1.5503199100494385,
+      "objective/train/theoretical_loss": 3.7345612387344906,
+      "objective/train/tokens_used": 872428000,
+      "objective/train/value_avg": -0.007656097412109375,
+      "objective/train/value_loss": 0.00033693682053126395,
+      "objective/train/value_max": -0.00014650821685791016,
+      "objective/train/value_min": -0.345947265625,
+      "objective/train/value_reward_corr": 0.6609246880890598,
+      "objective/train/value_std": 0.01441192626953125,
+      "objective/train/weight_avg": 1.0008713006973267,
+      "objective/train/weighted_lm_loss": 1.550735354423523,
+      "objective/train/weights_max": 1.2893656492233276,
+      "objective/train/weights_min": 0.3895317018032074,
+      "theoretical_loss": 3.7345612387344906,
+      "tokens_seen": 851968000
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007493179264965495,
+      "loss": 0.0758,
+      "theoretical_loss": 3.7345612387344906,
+      "tokens_seen": 851968000
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007492376825549671,
+      "loss": 0.0782,
+      "theoretical_loss": 3.734449502162596,
+      "tokens_seen": 852230144
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007491574386133847,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7343378095755257,
+      "tokens_seen": 852492288
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007490771946718022,
+      "loss": 0.0745,
+      "theoretical_loss": 3.7342261609424483,
+      "tokens_seen": 852754432
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007489969507302198,
+      "loss": 0.0788,
+      "theoretical_loss": 3.7341145562325613,
+      "tokens_seen": 853016576
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007489167067886375,
+      "loss": 0.0761,
+      "theoretical_loss": 3.734002995415096,
+      "tokens_seen": 853278720
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007488364628470551,
+      "loss": 0.0772,
+      "theoretical_loss": 3.7338914784593134,
+      "tokens_seen": 853540864
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007487562189054726,
+      "loss": 0.0755,
+      "theoretical_loss": 3.733780005334505,
+      "tokens_seen": 853803008
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007486759749638903,
+      "loss": 0.0789,
+      "theoretical_loss": 3.733668576009995,
+      "tokens_seen": 854065152
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007485957310223078,
+      "loss": 0.0768,
+      "theoretical_loss": 3.733557190455136,
+      "tokens_seen": 854327296
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007485154870807254,
+      "loss": 0.0742,
+      "theoretical_loss": 3.733445848639313,
+      "tokens_seen": 854589440
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000748435243139143,
+      "loss": 0.0754,
+      "theoretical_loss": 3.733334550531942,
+      "tokens_seen": 854851584
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007483549991975605,
+      "loss": 0.0776,
+      "theoretical_loss": 3.7332232961024694,
+      "tokens_seen": 855113728
+    },
+    {
+      "epoch": 0.26,
+      "objective/train/advantage_avg": 0.001274409587495029,
+      "objective/train/docs_used": 315806,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4173961877822876,
+      "objective/train/original_loss": 1.417395830154419,
+      "objective/train/theoretical_loss": 3.733167685257405,
+      "objective/train/tokens_used": 875704800,
+      "objective/train/value_avg": -0.006504058837890625,
+      "objective/train/value_loss": 0.00017858008504845202,
+      "objective/train/value_max": -0.0001647472381591797,
+      "objective/train/value_min": -0.2578125,
+      "objective/train/value_reward_corr": 0.5306423754138953,
+      "objective/train/value_std": 0.00859832763671875,
+      "objective/train/weight_avg": 1.001355767250061,
+      "objective/train/weighted_lm_loss": 1.420598030090332,
+      "objective/train/weights_max": 1.1989800930023193,
+      "objective/train/weights_min": 0.38221901655197144,
+      "theoretical_loss": 3.733167685257405,
+      "tokens_seen": 855244800
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007482747552559783,
+      "loss": 0.0759,
+      "theoretical_loss": 3.7331120853203714,
+      "tokens_seen": 855375872
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007481945113143958,
+      "loss": 0.0752,
+      "theoretical_loss": 3.733000918155156,
+      "tokens_seen": 855638016
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007481142673728134,
+      "loss": 0.0726,
+      "theoretical_loss": 3.7328897945763617,
+      "tokens_seen": 855900160
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000748034023431231,
+      "loss": 0.0781,
+      "theoretical_loss": 3.7327787145535574,
+      "tokens_seen": 856162304
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007479537794896486,
+      "loss": 0.0738,
+      "theoretical_loss": 3.732667678056342,
+      "tokens_seen": 856424448
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007478735355480661,
+      "loss": 0.073,
+      "theoretical_loss": 3.732556685054346,
+      "tokens_seen": 856686592
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007477932916064837,
+      "loss": 0.0745,
+      "theoretical_loss": 3.7324457355172296,
+      "tokens_seen": 856948736
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007477130476649013,
+      "loss": 0.0748,
+      "theoretical_loss": 3.7323348294146843,
+      "tokens_seen": 857210880
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007476328037233188,
+      "loss": 0.0758,
+      "theoretical_loss": 3.73222396671643,
+      "tokens_seen": 857473024
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007475525597817365,
+      "loss": 0.0766,
+      "theoretical_loss": 3.73211314739222,
+      "tokens_seen": 857735168
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000747472315840154,
+      "loss": 0.0748,
+      "theoretical_loss": 3.732002371411835,
+      "tokens_seen": 857997312
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007473920718985718,
+      "loss": 0.0738,
+      "theoretical_loss": 3.7318916387450876,
+      "tokens_seen": 858259456
+    },
+    {
+      "epoch": 0.26,
+      "objective/train/advantage_avg": 0.00041493060416541994,
+      "objective/train/docs_used": 317142,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5313217639923096,
+      "objective/train/original_loss": 1.5313217639923096,
+      "objective/train/theoretical_loss": 3.7317809493618204,
+      "objective/train/tokens_used": 878981600,
+      "objective/train/value_avg": -0.007602691650390625,
+      "objective/train/value_loss": 0.00031841802410781384,
+      "objective/train/value_max": -0.0001596212387084961,
+      "objective/train/value_min": -0.71923828125,
+      "objective/train/value_reward_corr": 0.6602214596414195,
+      "objective/train/value_std": 0.0145111083984375,
+      "objective/train/weight_avg": 1.0005568265914917,
+      "objective/train/weighted_lm_loss": 1.5318763256072998,
+      "objective/train/weights_max": 1.2434570789337158,
+      "objective/train/weights_min": 0.3680877387523651,
+      "theoretical_loss": 3.7317809493618204,
+      "tokens_seen": 858521600
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007473118279569893,
+      "loss": 0.0776,
+      "theoretical_loss": 3.7317809493618204,
+      "tokens_seen": 858521600
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007472315840154068,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7316703032319056,
+      "tokens_seen": 858783744
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007471513400738245,
+      "loss": 0.0787,
+      "theoretical_loss": 3.7315597003252474,
+      "tokens_seen": 859045888
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000747071096132242,
+      "loss": 0.0784,
+      "theoretical_loss": 3.731449140611777,
+      "tokens_seen": 859308032
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007469908521906596,
+      "loss": 0.0747,
+      "theoretical_loss": 3.7313386240614577,
+      "tokens_seen": 859570176
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007469106082490772,
+      "loss": 0.0761,
+      "theoretical_loss": 3.7312281506442835,
+      "tokens_seen": 859832320
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007468303643074948,
+      "loss": 0.0758,
+      "theoretical_loss": 3.7311177203302766,
+      "tokens_seen": 860094464
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007467501203659124,
+      "loss": 0.0769,
+      "theoretical_loss": 3.7310073330894906,
+      "tokens_seen": 860356608
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.00074666987642433,
+      "loss": 0.074,
+      "theoretical_loss": 3.730896988892008,
+      "tokens_seen": 860618752
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007465896324827476,
+      "loss": 0.0809,
+      "theoretical_loss": 3.7307866877079414,
+      "tokens_seen": 860880896
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007465093885411651,
+      "loss": 0.0758,
+      "theoretical_loss": 3.730676429507435,
+      "tokens_seen": 861143040
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007464291445995828,
+      "loss": 0.0736,
+      "theoretical_loss": 3.730566214260659,
+      "tokens_seen": 861405184
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007463489006580003,
+      "loss": 0.0762,
+      "theoretical_loss": 3.730456041937817,
+      "tokens_seen": 861667328
+    },
+    {
+      "epoch": 0.26,
+      "objective/train/advantage_avg": 0.0012228551786392927,
+      "objective/train/docs_used": 318349,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4355342388153076,
+      "objective/train/original_loss": 1.4355342388153076,
+      "objective/train/theoretical_loss": 3.730400971863568,
+      "objective/train/tokens_used": 882258400,
+      "objective/train/value_avg": -0.00885772705078125,
+      "objective/train/value_loss": 0.0004532379098236561,
+      "objective/train/value_max": -0.00011235475540161133,
+      "objective/train/value_min": -0.8740234375,
+      "objective/train/value_reward_corr": 0.5636366361372568,
+      "objective/train/value_std": 0.015533447265625,
+      "objective/train/weight_avg": 1.001416563987732,
+      "objective/train/weighted_lm_loss": 1.4377250671386719,
+      "objective/train/weights_max": 1.8811969757080078,
+      "objective/train/weights_min": 0.22597576677799225,
+      "theoretical_loss": 3.730400971863568,
+      "tokens_seen": 861798400
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007462686567164179,
+      "loss": 0.0767,
+      "theoretical_loss": 3.730345912509141,
+      "tokens_seen": 861929472
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007461884127748355,
+      "loss": 0.0742,
+      "theoretical_loss": 3.7302358259448924,
+      "tokens_seen": 862191616
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000746108168833253,
+      "loss": 0.0778,
+      "theoretical_loss": 3.730125782215362,
+      "tokens_seen": 862453760
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007460279248916708,
+      "loss": 0.0766,
+      "theoretical_loss": 3.730015781290872,
+      "tokens_seen": 862715904
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007459476809500883,
+      "loss": 0.0784,
+      "theoretical_loss": 3.729905823141771,
+      "tokens_seen": 862978048
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007458674370085059,
+      "loss": 0.0739,
+      "theoretical_loss": 3.729795907738441,
+      "tokens_seen": 863240192
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007457871930669235,
+      "loss": 0.0748,
+      "theoretical_loss": 3.729686035051291,
+      "tokens_seen": 863502336
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007457069491253411,
+      "loss": 0.0762,
+      "theoretical_loss": 3.7295762050507593,
+      "tokens_seen": 863764480
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007456267051837586,
+      "loss": 0.0754,
+      "theoretical_loss": 3.7294664177073145,
+      "tokens_seen": 864026624
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007455464612421762,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7293566729914547,
+      "tokens_seen": 864288768
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007454662173005938,
+      "loss": 0.0756,
+      "theoretical_loss": 3.7292469708737066,
+      "tokens_seen": 864550912
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007453859733590113,
+      "loss": 0.0767,
+      "theoretical_loss": 3.729137311324627,
+      "tokens_seen": 864813056
+    },
+    {
+      "epoch": 0.26,
+      "objective/train/advantage_avg": 6.832103099441156e-05,
+      "objective/train/docs_used": 319421,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.434070348739624,
+      "objective/train/original_loss": 1.4340702295303345,
+      "objective/train/theoretical_loss": 3.7290276943148015,
+      "objective/train/tokens_used": 885535200,
+      "objective/train/value_avg": -0.00914764404296875,
+      "objective/train/value_loss": 0.0007328266510739923,
+      "objective/train/value_max": -9.387731552124023e-05,
+      "objective/train/value_min": -0.91455078125,
+      "objective/train/value_reward_corr": 0.5348672218461172,
+      "objective/train/value_std": 0.0240936279296875,
+      "objective/train/weight_avg": 1.0004147291183472,
+      "objective/train/weighted_lm_loss": 1.43430757522583,
+      "objective/train/weights_max": 2.415832042694092,
+      "objective/train/weights_min": 0.3684886395931244,
+      "theoretical_loss": 3.7290276943148015,
+      "tokens_seen": 865075200
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000745305729417429,
+      "loss": 0.073,
+      "theoretical_loss": 3.7290276943148015,
+      "tokens_seen": 865075200
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007452254854758466,
+      "loss": 0.0777,
+      "theoretical_loss": 3.7289181198148458,
+      "tokens_seen": 865337344
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007451452415342642,
+      "loss": 0.0774,
+      "theoretical_loss": 3.7288085877954025,
+      "tokens_seen": 865599488
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007450649975926818,
+      "loss": 0.0766,
+      "theoretical_loss": 3.728699098227146,
+      "tokens_seen": 865861632
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007449847536510994,
+      "loss": 0.0757,
+      "theoretical_loss": 3.728589651080779,
+      "tokens_seen": 866123776
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000744904509709517,
+      "loss": 0.0754,
+      "theoretical_loss": 3.728480246327032,
+      "tokens_seen": 866385920
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007448242657679345,
+      "loss": 0.0745,
+      "theoretical_loss": 3.7283708839366656,
+      "tokens_seen": 866648064
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007447440218263521,
+      "loss": 0.0766,
+      "theoretical_loss": 3.72826156388047,
+      "tokens_seen": 866910208
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007446637778847697,
+      "loss": 0.0763,
+      "theoretical_loss": 3.728152286129263,
+      "tokens_seen": 867172352
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007445835339431873,
+      "loss": 0.0747,
+      "theoretical_loss": 3.728043050653893,
+      "tokens_seen": 867434496
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007445032900016049,
+      "loss": 0.0798,
+      "theoretical_loss": 3.7279338574252354,
+      "tokens_seen": 867696640
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007444230460600226,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7278247064141956,
+      "tokens_seen": 867958784
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007443428021184401,
+      "loss": 0.0772,
+      "theoretical_loss": 3.7277155975917076,
+      "tokens_seen": 868220928
+    },
+    {
+      "epoch": 0.26,
+      "objective/train/advantage_avg": 0.0010427895467728376,
+      "objective/train/docs_used": 320668,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5272066593170166,
+      "objective/train/original_loss": 1.5272066593170166,
+      "objective/train/theoretical_loss": 3.727661058992095,
+      "objective/train/tokens_used": 888812000,
+      "objective/train/value_avg": -0.0088348388671875,
+      "objective/train/value_loss": 0.00025395132252015173,
+      "objective/train/value_max": -0.0001398324966430664,
+      "objective/train/value_min": -0.421630859375,
+      "objective/train/value_reward_corr": 0.7319601341651985,
+      "objective/train/value_std": 0.0164642333984375,
+      "objective/train/weight_avg": 1.0011564493179321,
+      "objective/train/weighted_lm_loss": 1.5287580490112305,
+      "objective/train/weights_max": 1.296626091003418,
+      "objective/train/weights_min": 0.3688490390777588,
+      "theoretical_loss": 3.727661058992095,
+      "tokens_seen": 868352000
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007442625581768576,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7276065309287345,
+      "tokens_seen": 868483072
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007441823142352753,
+      "loss": 0.0775,
+      "theoretical_loss": 3.727497506396267,
+      "tokens_seen": 868745216
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007441020702936928,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7273885239653266,
+      "tokens_seen": 869007360
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007440218263521104,
+      "loss": 0.0753,
+      "theoretical_loss": 3.727279583606961,
+      "tokens_seen": 869269504
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000743941582410528,
+      "loss": 0.0763,
+      "theoretical_loss": 3.727170685292248,
+      "tokens_seen": 869531648
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007438613384689456,
+      "loss": 0.077,
+      "theoretical_loss": 3.7270618289922943,
+      "tokens_seen": 869793792
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007437810945273631,
+      "loss": 0.0753,
+      "theoretical_loss": 3.7269530146782337,
+      "tokens_seen": 870055936
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007437008505857808,
+      "loss": 0.0809,
+      "theoretical_loss": 3.72684424232123,
+      "tokens_seen": 870318080
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007436206066441984,
+      "loss": 0.0772,
+      "theoretical_loss": 3.7267355118924748,
+      "tokens_seen": 870580224
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000743540362702616,
+      "loss": 0.0753,
+      "theoretical_loss": 3.726626823363188,
+      "tokens_seen": 870842368
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007434601187610336,
+      "loss": 0.0754,
+      "theoretical_loss": 3.7265181767046176,
+      "tokens_seen": 871104512
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007433798748194511,
+      "loss": 0.0773,
+      "theoretical_loss": 3.726409571888042,
+      "tokens_seen": 871366656
+    },
+    {
+      "epoch": 0.26,
+      "objective/train/advantage_avg": 0.0008427270222455263,
+      "objective/train/docs_used": 321822,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6887977123260498,
+      "objective/train/original_loss": 1.6887975931167603,
+      "objective/train/theoretical_loss": 3.7263010088847652,
+      "objective/train/tokens_used": 892088800,
+      "objective/train/value_avg": -0.00997161865234375,
+      "objective/train/value_loss": 0.00024173302517738193,
+      "objective/train/value_max": -0.00015115737915039062,
+      "objective/train/value_min": -0.336181640625,
+      "objective/train/value_reward_corr": 0.7177472974340894,
+      "objective/train/value_std": 0.01708984375,
+      "objective/train/weight_avg": 1.000956654548645,
+      "objective/train/weighted_lm_loss": 1.689243197441101,
+      "objective/train/weights_max": 1.1999152898788452,
+      "objective/train/weights_min": 0.40030547976493835,
+      "theoretical_loss": 3.7263010088847652,
+      "tokens_seen": 871628800
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007432996308778688,
+      "loss": 0.0784,
+      "theoretical_loss": 3.7263010088847652,
+      "tokens_seen": 871628800
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007432193869362863,
+      "loss": 0.0753,
+      "theoretical_loss": 3.726192487666121,
+      "tokens_seen": 871890944
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007431391429947038,
+      "loss": 0.0765,
+      "theoretical_loss": 3.7260840082034714,
+      "tokens_seen": 872153088
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007430588990531216,
+      "loss": 0.0759,
+      "theoretical_loss": 3.7259755704682065,
+      "tokens_seen": 872415232
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007429786551115391,
+      "loss": 0.0766,
+      "theoretical_loss": 3.7258671744317446,
+      "tokens_seen": 872677376
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007428984111699567,
+      "loss": 0.0768,
+      "theoretical_loss": 3.725758820065531,
+      "tokens_seen": 872939520
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007428181672283743,
+      "loss": 0.0761,
+      "theoretical_loss": 3.725650507341042,
+      "tokens_seen": 873201664
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007427379232867919,
+      "loss": 0.0774,
+      "theoretical_loss": 3.7255422362297788,
+      "tokens_seen": 873463808
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007426576793452094,
+      "loss": 0.0759,
+      "theoretical_loss": 3.7254340067032725,
+      "tokens_seen": 873725952
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.000742577435403627,
+      "loss": 0.0766,
+      "theoretical_loss": 3.7253258187330816,
+      "tokens_seen": 873988096
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 0.0007424971914620446,
+      "loss": 0.0763,
+      "theoretical_loss": 3.7252176722907926,
+      "tokens_seen": 874250240
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007424169475204621,
+      "loss": 0.0788,
+      "theoretical_loss": 3.725109567348021,
+      "tokens_seen": 874512384
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007423367035788799,
+      "loss": 0.0778,
+      "theoretical_loss": 3.725001503876408,
+      "tokens_seen": 874774528
+    },
+    {
+      "epoch": 0.27,
+      "objective/train/advantage_avg": 0.00147084123454988,
+      "objective/train/docs_used": 323008,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5208227634429932,
+      "objective/train/original_loss": 1.520822525024414,
+      "objective/train/theoretical_loss": 3.7249474876834325,
+      "objective/train/tokens_used": 895365600,
+      "objective/train/value_avg": -0.007843017578125,
+      "objective/train/value_loss": 0.00020401620713528246,
+      "objective/train/value_max": -0.00010311603546142578,
+      "objective/train/value_min": -0.30029296875,
+      "objective/train/value_reward_corr": 0.789625392294421,
+      "objective/train/value_std": 0.017059326171875,
+      "objective/train/weight_avg": 1.0015629529953003,
+      "objective/train/weighted_lm_loss": 1.5230070352554321,
+      "objective/train/weights_max": 1.1601619720458984,
+      "objective/train/weights_min": 0.37002718448638916,
+      "theoretical_loss": 3.7249474876834325,
+      "tokens_seen": 874905600
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007422564596372974,
+      "loss": 0.0756,
+      "theoretical_loss": 3.7248934818476247,
+      "tokens_seen": 875036672
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007421762156957151,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7247855012333693,
+      "tokens_seen": 875298816
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007420959717541326,
+      "loss": 0.0759,
+      "theoretical_loss": 3.7246775620053665,
+      "tokens_seen": 875560960
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007420157278125502,
+      "loss": 0.0768,
+      "theoretical_loss": 3.724569664135372,
+      "tokens_seen": 875823104
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007419354838709678,
+      "loss": 0.0763,
+      "theoretical_loss": 3.7244618075951657,
+      "tokens_seen": 876085248
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007418552399293853,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7243539923565576,
+      "tokens_seen": 876347392
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007417749959878029,
+      "loss": 0.0773,
+      "theoretical_loss": 3.724246218391384,
+      "tokens_seen": 876609536
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007416947520462205,
+      "loss": 0.0778,
+      "theoretical_loss": 3.7241384856715096,
+      "tokens_seen": 876871680
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007416145081046381,
+      "loss": 0.0754,
+      "theoretical_loss": 3.724030794168826,
+      "tokens_seen": 877133824
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007415342641630557,
+      "loss": 0.0752,
+      "theoretical_loss": 3.723923143855253,
+      "tokens_seen": 877395968
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007414540202214734,
+      "loss": 0.0746,
+      "theoretical_loss": 3.723815534702738,
+      "tokens_seen": 877658112
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007413737762798909,
+      "loss": 0.0785,
+      "theoretical_loss": 3.7237079666832553,
+      "tokens_seen": 877920256
+    },
+    {
+      "epoch": 0.27,
+      "objective/train/advantage_avg": -0.0005045576835982502,
+      "objective/train/docs_used": 324331,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4083878993988037,
+      "objective/train/original_loss": 1.4083881378173828,
+      "objective/train/theoretical_loss": 3.7236004397688065,
+      "objective/train/tokens_used": 898642400,
+      "objective/train/value_avg": -0.01111602783203125,
+      "objective/train/value_loss": 0.0004869569675065577,
+      "objective/train/value_max": -0.00018668174743652344,
+      "objective/train/value_min": -0.7138671875,
+      "objective/train/value_reward_corr": 0.598061434276012,
+      "objective/train/value_std": 0.0183563232421875,
+      "objective/train/weight_avg": 0.9997174143791199,
+      "objective/train/weighted_lm_loss": 1.406802773475647,
+      "objective/train/weights_max": 1.8978021144866943,
+      "objective/train/weights_min": 0.38027602434158325,
+      "theoretical_loss": 3.7236004397688065,
+      "tokens_seen": 878182400
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007412935323383084,
+      "loss": 0.078,
+      "theoretical_loss": 3.7236004397688065,
+      "tokens_seen": 878182400
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007412132883967261,
+      "loss": 0.0754,
+      "theoretical_loss": 3.723492953931421,
+      "tokens_seen": 878444544
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007411330444551436,
+      "loss": 0.0759,
+      "theoretical_loss": 3.7233855091431565,
+      "tokens_seen": 878706688
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007410528005135613,
+      "loss": 0.0759,
+      "theoretical_loss": 3.723278105376096,
+      "tokens_seen": 878968832
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007409725565719788,
+      "loss": 0.0748,
+      "theoretical_loss": 3.723170742602351,
+      "tokens_seen": 879230976
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007408923126303964,
+      "loss": 0.0775,
+      "theoretical_loss": 3.7230634207940607,
+      "tokens_seen": 879493120
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007408120686888141,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7229561399233906,
+      "tokens_seen": 879755264
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007407318247472316,
+      "loss": 0.0791,
+      "theoretical_loss": 3.7228488999625338,
+      "tokens_seen": 880017408
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007406515808056492,
+      "loss": 0.0775,
+      "theoretical_loss": 3.722741700883711,
+      "tokens_seen": 880279552
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007405713368640668,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7226345426591694,
+      "tokens_seen": 880541696
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007404910929224844,
+      "loss": 0.0732,
+      "theoretical_loss": 3.722527425261183,
+      "tokens_seen": 880803840
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007404108489809019,
+      "loss": 0.0783,
+      "theoretical_loss": 3.7224203486620535,
+      "tokens_seen": 881065984
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007403306050393196,
+      "loss": 0.0777,
+      "theoretical_loss": 3.7223133128341104,
+      "tokens_seen": 881328128
+    },
+    {
+      "epoch": 0.27,
+      "objective/train/advantage_avg": 0.0017001149244606495,
+      "objective/train/docs_used": 325483,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4574103355407715,
+      "objective/train/original_loss": 1.4574103355407715,
+      "objective/train/theoretical_loss": 3.722259810200693,
+      "objective/train/tokens_used": 901919200,
+      "objective/train/value_avg": -0.007709503173828125,
+      "objective/train/value_loss": 0.00018067251949105412,
+      "objective/train/value_max": -0.00011593103408813477,
+      "objective/train/value_min": -0.2108154296875,
+      "objective/train/value_reward_corr": 0.625343074103676,
+      "objective/train/value_std": 0.011138916015625,
+      "objective/train/weight_avg": 1.0017832517623901,
+      "objective/train/weighted_lm_loss": 1.4603242874145508,
+      "objective/train/weights_max": 1.2097421884536743,
+      "objective/train/weights_min": 0.36907634139060974,
+      "theoretical_loss": 3.722259810200693,
+      "tokens_seen": 881459200
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007402503610977371,
+      "loss": 0.0752,
+      "theoretical_loss": 3.722206317749708,
+      "tokens_seen": 881590272
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007401701171561546,
+      "loss": 0.0778,
+      "theoretical_loss": 3.722099363381229,
+      "tokens_seen": 881852416
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007400898732145724,
+      "loss": 0.0769,
+      "theoretical_loss": 3.7219924497010837,
+      "tokens_seen": 882114560
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007400096292729899,
+      "loss": 0.0766,
+      "theoretical_loss": 3.721885576681708,
+      "tokens_seen": 882376704
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007399293853314075,
+      "loss": 0.078,
+      "theoretical_loss": 3.7217787442955643,
+      "tokens_seen": 882638848
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007398491413898251,
+      "loss": 0.0767,
+      "theoretical_loss": 3.721671952515144,
+      "tokens_seen": 882900992
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007397688974482427,
+      "loss": 0.0762,
+      "theoretical_loss": 3.7215652013129628,
+      "tokens_seen": 883163136
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007396886535066603,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7214584906615644,
+      "tokens_seen": 883425280
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007396084095650778,
+      "loss": 0.0732,
+      "theoretical_loss": 3.7213518205335196,
+      "tokens_seen": 883687424
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007395281656234954,
+      "loss": 0.0764,
+      "theoretical_loss": 3.721245190901425,
+      "tokens_seen": 883949568
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.000739447921681913,
+      "loss": 0.077,
+      "theoretical_loss": 3.721138601737904,
+      "tokens_seen": 884211712
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007393676777403307,
+      "loss": 0.0755,
+      "theoretical_loss": 3.721032053015607,
+      "tokens_seen": 884473856
+    },
+    {
+      "epoch": 0.27,
+      "objective/train/advantage_avg": 0.0006489785737358034,
+      "objective/train/docs_used": 326745,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.7073997259140015,
+      "objective/train/original_loss": 1.707399606704712,
+      "objective/train/theoretical_loss": 3.720925544707211,
+      "objective/train/tokens_used": 905196000,
+      "objective/train/value_avg": -0.0076141357421875,
+      "objective/train/value_loss": 0.00017974516958929598,
+      "objective/train/value_max": -0.00010389089584350586,
+      "objective/train/value_min": -0.5849609375,
+      "objective/train/value_reward_corr": 0.7308202130586332,
+      "objective/train/value_std": 0.0149078369140625,
+      "objective/train/weight_avg": 1.0007365942001343,
+      "objective/train/weighted_lm_loss": 1.7088409662246704,
+      "objective/train/weights_max": 1.28644597530365,
+      "objective/train/weights_min": 0.5981242656707764,
+      "theoretical_loss": 3.720925544707211,
+      "tokens_seen": 884736000
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007392874337987482,
+      "loss": 0.0761,
+      "theoretical_loss": 3.720925544707211,
+      "tokens_seen": 884736000
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007392071898571659,
+      "loss": 0.078,
+      "theoretical_loss": 3.720819076785419,
+      "tokens_seen": 884998144
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007391269459155834,
+      "loss": 0.0741,
+      "theoretical_loss": 3.720712649222961,
+      "tokens_seen": 885260288
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007390467019740009,
+      "loss": 0.0752,
+      "theoretical_loss": 3.720606261992593,
+      "tokens_seen": 885522432
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007389664580324186,
+      "loss": 0.0777,
+      "theoretical_loss": 3.7204999150670988,
+      "tokens_seen": 885784576
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007388862140908361,
+      "loss": 0.0753,
+      "theoretical_loss": 3.7203936084192866,
+      "tokens_seen": 886046720
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007388059701492537,
+      "loss": 0.0775,
+      "theoretical_loss": 3.720287342021992,
+      "tokens_seen": 886308864
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007387257262076713,
+      "loss": 0.0748,
+      "theoretical_loss": 3.720181115848078,
+      "tokens_seen": 886571008
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.000738645482266089,
+      "loss": 0.0742,
+      "theoretical_loss": 3.7200749298704316,
+      "tokens_seen": 886833152
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007385652383245065,
+      "loss": 0.0778,
+      "theoretical_loss": 3.7199687840619675,
+      "tokens_seen": 887095296
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007384849943829242,
+      "loss": 0.0767,
+      "theoretical_loss": 3.719862678395627,
+      "tokens_seen": 887357440
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007384047504413417,
+      "loss": 0.0755,
+      "theoretical_loss": 3.719756612844377,
+      "tokens_seen": 887619584
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007383245064997593,
+      "loss": 0.0738,
+      "theoretical_loss": 3.7196505873812105,
+      "tokens_seen": 887881728
+    },
+    {
+      "epoch": 0.27,
+      "objective/train/advantage_avg": 0.0009603702928870916,
+      "objective/train/docs_used": 327973,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3757928609848022,
+      "objective/train/original_loss": 1.3757928609848022,
+      "objective/train/theoretical_loss": 3.719597589674226,
+      "objective/train/tokens_used": 908472800,
+      "objective/train/value_avg": -0.006237030029296875,
+      "objective/train/value_loss": 0.00027475733077153563,
+      "objective/train/value_max": -0.00013554096221923828,
+      "objective/train/value_min": -0.5302734375,
+      "objective/train/value_reward_corr": 0.5065483364707875,
+      "objective/train/value_std": 0.0123748779296875,
+      "objective/train/weight_avg": 1.0010859966278076,
+      "objective/train/weighted_lm_loss": 1.3766732215881348,
+      "objective/train/weights_max": 1.5861471891403198,
+      "objective/train/weights_min": 0.37259092926979065,
+      "theoretical_loss": 3.719597589674226,
+      "tokens_seen": 888012800
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007382442625581769,
+      "loss": 0.0765,
+      "theoretical_loss": 3.7195446019791465,
+      "tokens_seen": 888143872
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007381640186165944,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7194386566112314,
+      "tokens_seen": 888406016
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007380837746750121,
+      "loss": 0.0739,
+      "theoretical_loss": 3.7193327512505356,
+      "tokens_seen": 888668160
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007380035307334296,
+      "loss": 0.0773,
+      "theoretical_loss": 3.7192268858701576,
+      "tokens_seen": 888930304
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007379232867918472,
+      "loss": 0.0743,
+      "theoretical_loss": 3.7191210604432205,
+      "tokens_seen": 889192448
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007378430428502649,
+      "loss": 0.0776,
+      "theoretical_loss": 3.7190152749428735,
+      "tokens_seen": 889454592
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007377627989086824,
+      "loss": 0.0758,
+      "theoretical_loss": 3.7189095293422927,
+      "tokens_seen": 889716736
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007376825549671,
+      "loss": 0.0774,
+      "theoretical_loss": 3.71880382361468,
+      "tokens_seen": 889978880
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007376023110255176,
+      "loss": 0.0768,
+      "theoretical_loss": 3.7186981577332614,
+      "tokens_seen": 890241024
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007375220670839352,
+      "loss": 0.0737,
+      "theoretical_loss": 3.718592531671291,
+      "tokens_seen": 890503168
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007374418231423527,
+      "loss": 0.0761,
+      "theoretical_loss": 3.7184869454020477,
+      "tokens_seen": 890765312
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007373615792007704,
+      "loss": 0.075,
+      "theoretical_loss": 3.7183813988988357,
+      "tokens_seen": 891027456
+    },
+    {
+      "epoch": 0.27,
+      "objective/train/advantage_avg": 0.0005341742653399706,
+      "objective/train/docs_used": 329220,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4387054443359375,
+      "objective/train/original_loss": 1.4387052059173584,
+      "objective/train/theoretical_loss": 3.7182758921349865,
+      "objective/train/tokens_used": 911749600,
+      "objective/train/value_avg": -0.00934600830078125,
+      "objective/train/value_loss": 0.0006961746839806437,
+      "objective/train/value_max": -0.0002532005310058594,
+      "objective/train/value_min": -0.74853515625,
+      "objective/train/value_reward_corr": 0.5610245926975079,
+      "objective/train/value_std": 0.018096923828125,
+      "objective/train/weight_avg": 1.0008351802825928,
+      "objective/train/weighted_lm_loss": 1.439525842666626,
+      "objective/train/weights_max": 1.8936995267868042,
+      "objective/train/weights_min": 0.36917349696159363,
+      "theoretical_loss": 3.7182758921349865,
+      "tokens_seen": 891289600
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007372813352591879,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7182758921349865,
+      "tokens_seen": 891289600
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007372010913176055,
+      "loss": 0.0762,
+      "theoretical_loss": 3.718170425083856,
+      "tokens_seen": 891551744
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007371208473760232,
+      "loss": 0.0755,
+      "theoretical_loss": 3.718064997718826,
+      "tokens_seen": 891813888
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007370406034344407,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7179596100133034,
+      "tokens_seen": 892076032
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007369603594928584,
+      "loss": 0.0782,
+      "theoretical_loss": 3.7178542619407233,
+      "tokens_seen": 892338176
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007368801155512759,
+      "loss": 0.0786,
+      "theoretical_loss": 3.7177489534745427,
+      "tokens_seen": 892600320
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007367998716096935,
+      "loss": 0.0789,
+      "theoretical_loss": 3.717643684588247,
+      "tokens_seen": 892862464
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007367196276681111,
+      "loss": 0.077,
+      "theoretical_loss": 3.7175384552553457,
+      "tokens_seen": 893124608
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007366393837265286,
+      "loss": 0.0743,
+      "theoretical_loss": 3.7174332654493742,
+      "tokens_seen": 893386752
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007365591397849462,
+      "loss": 0.0768,
+      "theoretical_loss": 3.717328115143894,
+      "tokens_seen": 893648896
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007364788958433638,
+      "loss": 0.0768,
+      "theoretical_loss": 3.717223004312491,
+      "tokens_seen": 893911040
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007363986519017815,
+      "loss": 0.077,
+      "theoretical_loss": 3.717117932928777,
+      "tokens_seen": 894173184
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.000736318407960199,
+      "loss": 0.0754,
+      "theoretical_loss": 3.7170129009663886,
+      "tokens_seen": 894435328
+    },
+    {
+      "epoch": 0.27,
+      "objective/train/advantage_avg": 0.0014942382695153356,
+      "objective/train/docs_used": 330336,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3993629217147827,
+      "objective/train/original_loss": 1.3993628025054932,
+      "objective/train/theoretical_loss": 3.7169603997599605,
+      "objective/train/tokens_used": 915026400,
+      "objective/train/value_avg": -0.006420135498046875,
+      "objective/train/value_loss": 0.00015116189024411142,
+      "objective/train/value_max": -0.00015473365783691406,
+      "objective/train/value_min": -0.1751708984375,
+      "objective/train/value_reward_corr": 0.49777587233171894,
+      "objective/train/value_std": 0.0080718994140625,
+      "objective/train/weight_avg": 1.0015615224838257,
+      "objective/train/weighted_lm_loss": 1.4017528295516968,
+      "objective/train/weights_max": 1.12247896194458,
+      "objective/train/weights_min": 0.36870554089546204,
+      "theoretical_loss": 3.7169603997599605,
+      "tokens_seen": 894566400
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007362381640186167,
+      "loss": 0.0778,
+      "theoretical_loss": 3.716907908398989,
+      "tokens_seen": 894697472
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007361579200770342,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7168029552002655,
+      "tokens_seen": 894959616
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007360776761354517,
+      "loss": 0.0766,
+      "theoretical_loss": 3.716698041343931,
+      "tokens_seen": 895221760
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007359974321938694,
+      "loss": 0.078,
+      "theoretical_loss": 3.716593166803724,
+      "tokens_seen": 895483904
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007359171882522869,
+      "loss": 0.0735,
+      "theoretical_loss": 3.7164883315534087,
+      "tokens_seen": 895746048
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007358369443107046,
+      "loss": 0.0775,
+      "theoretical_loss": 3.7163835355667723,
+      "tokens_seen": 896008192
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007357567003691221,
+      "loss": 0.074,
+      "theoretical_loss": 3.7162787788176295,
+      "tokens_seen": 896270336
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007356764564275397,
+      "loss": 0.0763,
+      "theoretical_loss": 3.716174061279819,
+      "tokens_seen": 896532480
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007355962124859574,
+      "loss": 0.0741,
+      "theoretical_loss": 3.7160693829272047,
+      "tokens_seen": 896794624
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.000735515968544375,
+      "loss": 0.0747,
+      "theoretical_loss": 3.715964743733676,
+      "tokens_seen": 897056768
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007354357246027925,
+      "loss": 0.0779,
+      "theoretical_loss": 3.7158601436731464,
+      "tokens_seen": 897318912
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007353554806612101,
+      "loss": 0.0751,
+      "theoretical_loss": 3.715755582719556,
+      "tokens_seen": 897581056
+    },
+    {
+      "epoch": 0.27,
+      "objective/train/advantage_avg": 0.0011574587551876903,
+      "objective/train/docs_used": 331577,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4930155277252197,
+      "objective/train/original_loss": 1.4930155277252197,
+      "objective/train/theoretical_loss": 3.7156510608468674,
+      "objective/train/tokens_used": 918303200,
+      "objective/train/value_avg": -0.007770538330078125,
+      "objective/train/value_loss": 0.0005507190944626927,
+      "objective/train/value_max": -7.843971252441406e-05,
+      "objective/train/value_min": -0.568359375,
+      "objective/train/value_reward_corr": 0.6120014802789004,
+      "objective/train/value_std": 0.01477813720703125,
+      "objective/train/weight_avg": 1.0013858079910278,
+      "objective/train/weighted_lm_loss": 1.4944040775299072,
+      "objective/train/weights_max": 1.3329410552978516,
+      "objective/train/weights_min": 0.22743035852909088,
+      "theoretical_loss": 3.7156510608468674,
+      "tokens_seen": 897843200
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007352752367196277,
+      "loss": 0.0722,
+      "theoretical_loss": 3.7156510608468674,
+      "tokens_seen": 897843200
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007351949927780452,
+      "loss": 0.0771,
+      "theoretical_loss": 3.7155465780290706,
+      "tokens_seen": 898105344
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007351147488364629,
+      "loss": 0.0756,
+      "theoretical_loss": 3.7154421342401793,
+      "tokens_seen": 898367488
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007350345048948804,
+      "loss": 0.076,
+      "theoretical_loss": 3.7153377294542325,
+      "tokens_seen": 898629632
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.000734954260953298,
+      "loss": 0.0757,
+      "theoretical_loss": 3.715233363645293,
+      "tokens_seen": 898891776
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007348740170117157,
+      "loss": 0.0759,
+      "theoretical_loss": 3.7151290367874497,
+      "tokens_seen": 899153920
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007347937730701332,
+      "loss": 0.0747,
+      "theoretical_loss": 3.715024748854815,
+      "tokens_seen": 899416064
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007347135291285509,
+      "loss": 0.0754,
+      "theoretical_loss": 3.714920499821528,
+      "tokens_seen": 899678208
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007346332851869684,
+      "loss": 0.0746,
+      "theoretical_loss": 3.7148162896617505,
+      "tokens_seen": 899940352
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.000734553041245386,
+      "loss": 0.0742,
+      "theoretical_loss": 3.714712118349669,
+      "tokens_seen": 900202496
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007344727973038036,
+      "loss": 0.0789,
+      "theoretical_loss": 3.7146079858594976,
+      "tokens_seen": 900464640
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007343925533622212,
+      "loss": 0.0753,
+      "theoretical_loss": 3.714503892165471,
+      "tokens_seen": 900726784
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007343123094206387,
+      "loss": 0.0761,
+      "theoretical_loss": 3.714399837241851,
+      "tokens_seen": 900988928
+    },
+    {
+      "epoch": 0.27,
+      "objective/train/advantage_avg": 0.0022575294133275747,
+      "objective/train/docs_used": 332819,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6288766860961914,
+      "objective/train/original_loss": 1.6288766860961914,
+      "objective/train/theoretical_loss": 3.714347824310907,
+      "objective/train/tokens_used": 921580000,
+      "objective/train/value_avg": -0.00806427001953125,
+      "objective/train/value_loss": 0.00021155290596652776,
+      "objective/train/value_max": -0.00020182132720947266,
+      "objective/train/value_min": -0.2548828125,
+      "objective/train/value_reward_corr": 0.5714205851186827,
+      "objective/train/value_std": 0.01141357421875,
+      "objective/train/weight_avg": 1.0023540258407593,
+      "objective/train/weighted_lm_loss": 1.632721185684204,
+      "objective/train/weights_max": 1.173399567604065,
+      "objective/train/weights_min": 0.3694700598716736,
+      "theoretical_loss": 3.714347824310907,
+      "tokens_seen": 901120000
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007342320654790563,
+      "loss": 0.0751,
+      "theoretical_loss": 3.7142958210629233,
+      "tokens_seen": 901251072
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.000734151821537474,
+      "loss": 0.0769,
+      "theoretical_loss": 3.714191843602998,
+      "tokens_seen": 901513216
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007340715775958915,
+      "loss": 0.077,
+      "theoretical_loss": 3.7140879048364104,
+      "tokens_seen": 901775360
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007339913336543092,
+      "loss": 0.0756,
+      "theoretical_loss": 3.7139840047375183,
+      "tokens_seen": 902037504
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007339110897127267,
+      "loss": 0.0761,
+      "theoretical_loss": 3.713880143280707,
+      "tokens_seen": 902299648
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007338308457711443,
+      "loss": 0.076,
+      "theoretical_loss": 3.713776320440383,
+      "tokens_seen": 902561792
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007337506018295619,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7136725361909795,
+      "tokens_seen": 902823936
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007336703578879794,
+      "loss": 0.0765,
+      "theoretical_loss": 3.713568790506953,
+      "tokens_seen": 903086080
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.000733590113946397,
+      "loss": 0.0747,
+      "theoretical_loss": 3.7134650833627854,
+      "tokens_seen": 903348224
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007335098700048146,
+      "loss": 0.0736,
+      "theoretical_loss": 3.7133614147329808,
+      "tokens_seen": 903610368
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007334296260632322,
+      "loss": 0.0777,
+      "theoretical_loss": 3.7132577845920696,
+      "tokens_seen": 903872512
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007333493821216499,
+      "loss": 0.0761,
+      "theoretical_loss": 3.7131541929146055,
+      "tokens_seen": 904134656
+    },
+    {
+      "epoch": 0.27,
+      "objective/train/advantage_avg": -0.001184366992674768,
+      "objective/train/docs_used": 333943,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4446207284927368,
+      "objective/train/original_loss": 1.4446208477020264,
+      "objective/train/theoretical_loss": 3.713050639675166,
+      "objective/train/tokens_used": 924856800,
+      "objective/train/value_avg": -0.00861358642578125,
+      "objective/train/value_loss": 0.00047513179015368223,
+      "objective/train/value_max": -0.00012826919555664062,
+      "objective/train/value_min": -0.440185546875,
+      "objective/train/value_reward_corr": 0.6907593542665094,
+      "objective/train/value_std": 0.01509857177734375,
+      "objective/train/weight_avg": 0.9990274310112,
+      "objective/train/weighted_lm_loss": 1.4427376985549927,
+      "objective/train/weights_max": 1.3462886810302734,
+      "objective/train/weights_min": 0.37275585532188416,
+      "theoretical_loss": 3.713050639675166,
+      "tokens_seen": 904396800
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007332691381800675,
+      "loss": 0.073,
+      "theoretical_loss": 3.713050639675166,
+      "tokens_seen": 904396800
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.000733188894238485,
+      "loss": 0.0732,
+      "theoretical_loss": 3.712947124848354,
+      "tokens_seen": 904658944
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007331086502969026,
+      "loss": 0.0772,
+      "theoretical_loss": 3.7128436484087954,
+      "tokens_seen": 904921088
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007330284063553202,
+      "loss": 0.0751,
+      "theoretical_loss": 3.712740210331141,
+      "tokens_seen": 905183232
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007329481624137377,
+      "loss": 0.0769,
+      "theoretical_loss": 3.712636810590065,
+      "tokens_seen": 905445376
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007328679184721554,
+      "loss": 0.0737,
+      "theoretical_loss": 3.7125334491602664,
+      "tokens_seen": 905707520
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007327876745305729,
+      "loss": 0.0745,
+      "theoretical_loss": 3.712430126016467,
+      "tokens_seen": 905969664
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007327074305889905,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7123268411334136,
+      "tokens_seen": 906231808
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007326271866474082,
+      "loss": 0.0738,
+      "theoretical_loss": 3.7122235944858772,
+      "tokens_seen": 906493952
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007325469427058258,
+      "loss": 0.0736,
+      "theoretical_loss": 3.712120386048652,
+      "tokens_seen": 906756096
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007324666987642433,
+      "loss": 0.0761,
+      "theoretical_loss": 3.712017215796556,
+      "tokens_seen": 907018240
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 0.0007323864548226609,
+      "loss": 0.0744,
+      "theoretical_loss": 3.7119140837044315,
+      "tokens_seen": 907280384
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007323062108810785,
+      "loss": 0.0736,
+      "theoretical_loss": 3.7118109897471445,
+      "tokens_seen": 907542528
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": -0.00023626594338566065,
+      "objective/train/docs_used": 335091,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4100664854049683,
+      "objective/train/original_loss": 1.4100666046142578,
+      "objective/train/theoretical_loss": 3.7117594570612176,
+      "objective/train/tokens_used": 928133600,
+      "objective/train/value_avg": -0.006671905517578125,
+      "objective/train/value_loss": 0.0001805531937861815,
+      "objective/train/value_max": -9.459257125854492e-05,
+      "objective/train/value_min": -0.2276611328125,
+      "objective/train/value_reward_corr": 0.687647519301903,
+      "objective/train/value_std": 0.0102081298828125,
+      "objective/train/weight_avg": 0.9998487234115601,
+      "objective/train/weighted_lm_loss": 1.4099451303482056,
+      "objective/train/weights_max": 1.0809279680252075,
+      "objective/train/weights_min": 0.37694263458251953,
+      "theoretical_loss": 3.7117594570612176,
+      "tokens_seen": 907673600
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000732225966939496,
+      "loss": 0.0734,
+      "theoretical_loss": 3.7117079338995858,
+      "tokens_seen": 907804672
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007321457229979137,
+      "loss": 0.0769,
+      "theoretical_loss": 3.7116049161366673,
+      "tokens_seen": 908066816
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007320654790563312,
+      "loss": 0.0729,
+      "theoretical_loss": 3.7115019364333275,
+      "tokens_seen": 908328960
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000731985235114749,
+      "loss": 0.0762,
+      "theoretical_loss": 3.7113989947645276,
+      "tokens_seen": 908591104
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007319049911731665,
+      "loss": 0.0758,
+      "theoretical_loss": 3.711296091105252,
+      "tokens_seen": 908853248
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000731824747231584,
+      "loss": 0.0724,
+      "theoretical_loss": 3.7111932254305096,
+      "tokens_seen": 909115392
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007317445032900017,
+      "loss": 0.075,
+      "theoretical_loss": 3.7110903977153313,
+      "tokens_seen": 909377536
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007316642593484192,
+      "loss": 0.0759,
+      "theoretical_loss": 3.710987607934774,
+      "tokens_seen": 909639680
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007315840154068368,
+      "loss": 0.0759,
+      "theoretical_loss": 3.7108848560639167,
+      "tokens_seen": 909901824
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007315037714652544,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7107821420778615,
+      "tokens_seen": 910163968
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000731423527523672,
+      "loss": 0.0747,
+      "theoretical_loss": 3.7106794659517357,
+      "tokens_seen": 910426112
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007313432835820895,
+      "loss": 0.073,
+      "theoretical_loss": 3.7105768276606885,
+      "tokens_seen": 910688256
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": 0.0003594663867261261,
+      "objective/train/docs_used": 336141,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.548851728439331,
+      "objective/train/original_loss": 1.548851490020752,
+      "objective/train/theoretical_loss": 3.710474227179893,
+      "objective/train/tokens_used": 931410400,
+      "objective/train/value_avg": -0.005245208740234375,
+      "objective/train/value_loss": 0.00010473921429365873,
+      "objective/train/value_max": -6.973743438720703e-05,
+      "objective/train/value_min": -0.466552734375,
+      "objective/train/value_reward_corr": 0.6725354994272675,
+      "objective/train/value_std": 0.0080413818359375,
+      "objective/train/weight_avg": 1.0004104375839233,
+      "objective/train/weighted_lm_loss": 1.550046443939209,
+      "objective/train/weights_max": 1.089442253112793,
+      "objective/train/weights_min": 0.6256026029586792,
+      "theoretical_loss": 3.710474227179893,
+      "tokens_seen": 910950400
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007312630396405071,
+      "loss": 0.0742,
+      "theoretical_loss": 3.710474227179893,
+      "tokens_seen": 910950400
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007311827956989248,
+      "loss": 0.0742,
+      "theoretical_loss": 3.710371664484547,
+      "tokens_seen": 911212544
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007311025517573423,
+      "loss": 0.0715,
+      "theoretical_loss": 3.710269139549869,
+      "tokens_seen": 911474688
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.00073102230781576,
+      "loss": 0.0725,
+      "theoretical_loss": 3.7101666523511034,
+      "tokens_seen": 911736832
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007309420638741775,
+      "loss": 0.0749,
+      "theoretical_loss": 3.710064202863517,
+      "tokens_seen": 911998976
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007308618199325952,
+      "loss": 0.077,
+      "theoretical_loss": 3.7099617910623994,
+      "tokens_seen": 912261120
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007307815759910127,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7098594169230648,
+      "tokens_seen": 912523264
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007307013320494302,
+      "loss": 0.0731,
+      "theoretical_loss": 3.7097570804208497,
+      "tokens_seen": 912785408
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007306210881078479,
+      "loss": 0.0741,
+      "theoretical_loss": 3.709654781531113,
+      "tokens_seen": 913047552
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007305408441662654,
+      "loss": 0.0752,
+      "theoretical_loss": 3.709552520229239,
+      "tokens_seen": 913309696
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000730460600224683,
+      "loss": 0.0742,
+      "theoretical_loss": 3.7094502964906337,
+      "tokens_seen": 913571840
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007303803562831007,
+      "loss": 0.0742,
+      "theoretical_loss": 3.709348110290726,
+      "tokens_seen": 913833984
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007303001123415183,
+      "loss": 0.0744,
+      "theoretical_loss": 3.7092459616049682,
+      "tokens_seen": 914096128
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": 0.0010315380059182644,
+      "objective/train/docs_used": 337261,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.495456576347351,
+      "objective/train/original_loss": 1.4954564571380615,
+      "objective/train/theoretical_loss": 3.709194901322231,
+      "objective/train/tokens_used": 934687200,
+      "objective/train/value_avg": -0.01190185546875,
+      "objective/train/value_loss": 0.00031099331681616604,
+      "objective/train/value_max": -0.0002453327178955078,
+      "objective/train/value_min": -0.70556640625,
+      "objective/train/value_reward_corr": 0.7287494601771468,
+      "objective/train/value_std": 0.0191650390625,
+      "objective/train/weight_avg": 1.0011770725250244,
+      "objective/train/weighted_lm_loss": 1.498216986656189,
+      "objective/train/weights_max": 1.606416940689087,
+      "objective/train/weights_min": 0.3689524829387665,
+      "theoretical_loss": 3.709194901322231,
+      "tokens_seen": 914227200
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007302198683999358,
+      "loss": 0.0747,
+      "theoretical_loss": 3.709143850408837,
+      "tokens_seen": 914358272
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007301396244583534,
+      "loss": 0.0742,
+      "theoretical_loss": 3.7090417766778305,
+      "tokens_seen": 914620416
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000730059380516771,
+      "loss": 0.0738,
+      "theoretical_loss": 3.7089397403874704,
+      "tokens_seen": 914882560
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007299791365751885,
+      "loss": 0.0742,
+      "theoretical_loss": 3.708837741513301,
+      "tokens_seen": 915144704
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007298988926336062,
+      "loss": 0.0752,
+      "theoretical_loss": 3.7087357800308904,
+      "tokens_seen": 915406848
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007298186486920237,
+      "loss": 0.0735,
+      "theoretical_loss": 3.708633855915829,
+      "tokens_seen": 915668992
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007297384047504413,
+      "loss": 0.0742,
+      "theoretical_loss": 3.708531969143731,
+      "tokens_seen": 915931136
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000729658160808859,
+      "loss": 0.0769,
+      "theoretical_loss": 3.708430119690232,
+      "tokens_seen": 916193280
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007295779168672765,
+      "loss": 0.075,
+      "theoretical_loss": 3.708328307530991,
+      "tokens_seen": 916455424
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007294976729256942,
+      "loss": 0.0736,
+      "theoretical_loss": 3.7082265326416914,
+      "tokens_seen": 916717568
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007294174289841117,
+      "loss": 0.0733,
+      "theoretical_loss": 3.708124794998037,
+      "tokens_seen": 916979712
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007293371850425293,
+      "loss": 0.0724,
+      "theoretical_loss": 3.708023094575756,
+      "tokens_seen": 917241856
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": 0.0011083170538768172,
+      "objective/train/docs_used": 338425,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5612324476242065,
+      "objective/train/original_loss": 1.561232328414917,
+      "objective/train/theoretical_loss": 3.7079214313505986,
+      "objective/train/tokens_used": 937964000,
+      "objective/train/value_avg": -0.0082550048828125,
+      "objective/train/value_loss": 0.0003109085373580456,
+      "objective/train/value_max": -0.00016868114471435547,
+      "objective/train/value_min": -0.955078125,
+      "objective/train/value_reward_corr": 0.687470457987905,
+      "objective/train/value_std": 0.0196990966796875,
+      "objective/train/weight_avg": 1.0012561082839966,
+      "objective/train/weighted_lm_loss": 1.562901258468628,
+      "objective/train/weights_max": 2.4972524642944336,
+      "objective/train/weights_min": 0.38558143377304077,
+      "theoretical_loss": 3.7079214313505986,
+      "tokens_seen": 917504000
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007292569411009469,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7079214313505986,
+      "tokens_seen": 917504000
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007291766971593645,
+      "loss": 0.073,
+      "theoretical_loss": 3.707819805298338,
+      "tokens_seen": 917766144
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000729096453217782,
+      "loss": 0.0754,
+      "theoretical_loss": 3.7077182163947704,
+      "tokens_seen": 918028288
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007290162092761998,
+      "loss": 0.0749,
+      "theoretical_loss": 3.7076166646157134,
+      "tokens_seen": 918290432
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007289359653346173,
+      "loss": 0.075,
+      "theoretical_loss": 3.7075151499370094,
+      "tokens_seen": 918552576
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007288557213930348,
+      "loss": 0.0769,
+      "theoretical_loss": 3.7074136723345212,
+      "tokens_seen": 918814720
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007287754774514525,
+      "loss": 0.0771,
+      "theoretical_loss": 3.707312231784136,
+      "tokens_seen": 919076864
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.00072869523350987,
+      "loss": 0.0746,
+      "theoretical_loss": 3.7072108282617617,
+      "tokens_seen": 919339008
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007286149895682876,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7071094617433307,
+      "tokens_seen": 919601152
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007285347456267052,
+      "loss": 0.075,
+      "theoretical_loss": 3.707008132204796,
+      "tokens_seen": 919863296
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007284545016851228,
+      "loss": 0.0733,
+      "theoretical_loss": 3.7069068396221345,
+      "tokens_seen": 920125440
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007283742577435403,
+      "loss": 0.0733,
+      "theoretical_loss": 3.706805583971345,
+      "tokens_seen": 920387584
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007282940138019579,
+      "loss": 0.0757,
+      "theoretical_loss": 3.7067043652284495,
+      "tokens_seen": 920649728
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": -0.00018246793479193002,
+      "objective/train/docs_used": 339576,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.468998908996582,
+      "objective/train/original_loss": 1.468998908996582,
+      "objective/train/theoretical_loss": 3.706653769689974,
+      "objective/train/tokens_used": 941240800,
+      "objective/train/value_avg": -0.0093231201171875,
+      "objective/train/value_loss": 0.0003314831992611289,
+      "objective/train/value_max": -0.00017261505126953125,
+      "objective/train/value_min": -0.313720703125,
+      "objective/train/value_reward_corr": 0.6134654432248161,
+      "objective/train/value_std": 0.0128936767578125,
+      "objective/train/weight_avg": 0.9999735355377197,
+      "objective/train/weighted_lm_loss": 1.4704101085662842,
+      "objective/train/weights_max": 1.2405765056610107,
+      "objective/train/weights_min": 0.38339895009994507,
+      "theoretical_loss": 3.706653769689974,
+      "tokens_seen": 920780800
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007282137698603756,
+      "loss": 0.0812,
+      "theoretical_loss": 3.7066031833694906,
+      "tokens_seen": 920911872
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007281335259187932,
+      "loss": 0.0766,
+      "theoretical_loss": 3.7065020383705347,
+      "tokens_seen": 921174016
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007280532819772108,
+      "loss": 0.072,
+      "theoretical_loss": 3.70640093020767,
+      "tokens_seen": 921436160
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007279730380356283,
+      "loss": 0.0763,
+      "theoretical_loss": 3.7062998588570073,
+      "tokens_seen": 921698304
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000727892794094046,
+      "loss": 0.0741,
+      "theoretical_loss": 3.7061988242946793,
+      "tokens_seen": 921960448
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007278125501524635,
+      "loss": 0.0745,
+      "theoretical_loss": 3.7060978264968423,
+      "tokens_seen": 922222592
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000727732306210881,
+      "loss": 0.0754,
+      "theoretical_loss": 3.705996865439672,
+      "tokens_seen": 922484736
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007276520622692987,
+      "loss": 0.0728,
+      "theoretical_loss": 3.7058959410993695,
+      "tokens_seen": 922746880
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007275718183277162,
+      "loss": 0.0744,
+      "theoretical_loss": 3.7057950534521558,
+      "tokens_seen": 923009024
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007274915743861338,
+      "loss": 0.0764,
+      "theoretical_loss": 3.705694202474275,
+      "tokens_seen": 923271168
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007274113304445515,
+      "loss": 0.0743,
+      "theoretical_loss": 3.7055933881419936,
+      "tokens_seen": 923533312
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007273310865029691,
+      "loss": 0.0708,
+      "theoretical_loss": 3.7054926104315995,
+      "tokens_seen": 923795456
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": 0.0013274255907163024,
+      "objective/train/docs_used": 340768,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6183089017868042,
+      "objective/train/original_loss": 1.6183090209960938,
+      "objective/train/theoretical_loss": 3.705391869319403,
+      "objective/train/tokens_used": 944517600,
+      "objective/train/value_avg": -0.00782012939453125,
+      "objective/train/value_loss": 0.00026030809385702014,
+      "objective/train/value_max": -0.0001398324966430664,
+      "objective/train/value_min": -0.3779296875,
+      "objective/train/value_reward_corr": 0.6056620533963706,
+      "objective/train/value_std": 0.0130767822265625,
+      "objective/train/weight_avg": 1.0014489889144897,
+      "objective/train/weighted_lm_loss": 1.620178461074829,
+      "objective/train/weights_max": 1.4343163967132568,
+      "objective/train/weights_min": 0.37169942259788513,
+      "theoretical_loss": 3.705391869319403,
+      "tokens_seen": 924057600
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007272508425613866,
+      "loss": 0.0733,
+      "theoretical_loss": 3.705391869319403,
+      "tokens_seen": 924057600
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007271705986198042,
+      "loss": 0.0717,
+      "theoretical_loss": 3.7052911647817357,
+      "tokens_seen": 924319744
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007270903546782218,
+      "loss": 0.0754,
+      "theoretical_loss": 3.7051904967949527,
+      "tokens_seen": 924581888
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007270101107366394,
+      "loss": 0.0712,
+      "theoretical_loss": 3.7050898653354296,
+      "tokens_seen": 924844032
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000726929866795057,
+      "loss": 0.0747,
+      "theoretical_loss": 3.7049892703795653,
+      "tokens_seen": 925106176
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007268496228534745,
+      "loss": 0.0742,
+      "theoretical_loss": 3.70488871190378,
+      "tokens_seen": 925368320
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007267693789118923,
+      "loss": 0.0738,
+      "theoretical_loss": 3.704788189884515,
+      "tokens_seen": 925630464
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007266891349703098,
+      "loss": 0.0755,
+      "theoretical_loss": 3.7046877042982347,
+      "tokens_seen": 925892608
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007266088910287273,
+      "loss": 0.076,
+      "theoretical_loss": 3.7045872551214254,
+      "tokens_seen": 926154752
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000726528647087145,
+      "loss": 0.072,
+      "theoretical_loss": 3.704486842330594,
+      "tokens_seen": 926416896
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007264484031455625,
+      "loss": 0.074,
+      "theoretical_loss": 3.7043864659022696,
+      "tokens_seen": 926679040
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007263681592039801,
+      "loss": 0.0717,
+      "theoretical_loss": 3.704286125813004,
+      "tokens_seen": 926941184
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007262879152623977,
+      "loss": 0.074,
+      "theoretical_loss": 3.7041858220393706,
+      "tokens_seen": 927203328
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": -0.0009031386580318213,
+      "objective/train/docs_used": 341815,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.391129970550537,
+      "objective/train/original_loss": 1.391129970550537,
+      "objective/train/theoretical_loss": 3.7041356837636013,
+      "objective/train/tokens_used": 947794400,
+      "objective/train/value_avg": -0.0078125,
+      "objective/train/value_loss": 0.0002524466544855386,
+      "objective/train/value_max": -8.481740951538086e-05,
+      "objective/train/value_min": -0.256591796875,
+      "objective/train/value_reward_corr": 0.7173105114970126,
+      "objective/train/value_std": 0.012786865234375,
+      "objective/train/weight_avg": 0.9992140531539917,
+      "objective/train/weighted_lm_loss": 1.3906961679458618,
+      "objective/train/weights_max": 1.1406426429748535,
+      "objective/train/weights_min": 0.3694460988044739,
+      "theoretical_loss": 3.7041356837636013,
+      "tokens_seen": 927334400
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007262076713208153,
+      "loss": 0.0731,
+      "theoretical_loss": 3.704085554557964,
+      "tokens_seen": 927465472
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007261274273792328,
+      "loss": 0.0729,
+      "theoretical_loss": 3.703985323345399,
+      "tokens_seen": 927727616
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007260471834376505,
+      "loss": 0.0707,
+      "theoretical_loss": 3.7038851283783156,
+      "tokens_seen": 927989760
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000725966939496068,
+      "loss": 0.0748,
+      "theoretical_loss": 3.7037849696333724,
+      "tokens_seen": 928251904
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007258866955544856,
+      "loss": 0.0767,
+      "theoretical_loss": 3.703684847087251,
+      "tokens_seen": 928514048
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007258064516129033,
+      "loss": 0.0762,
+      "theoretical_loss": 3.7035847607166534,
+      "tokens_seen": 928776192
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007257262076713208,
+      "loss": 0.0735,
+      "theoretical_loss": 3.703484710498306,
+      "tokens_seen": 929038336
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007256459637297385,
+      "loss": 0.0802,
+      "theoretical_loss": 3.703384696408953,
+      "tokens_seen": 929300480
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000725565719788156,
+      "loss": 0.0735,
+      "theoretical_loss": 3.7032847184253628,
+      "tokens_seen": 929562624
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007254854758465736,
+      "loss": 0.0732,
+      "theoretical_loss": 3.7031847765243233,
+      "tokens_seen": 929824768
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007254052319049912,
+      "loss": 0.0713,
+      "theoretical_loss": 3.7030848706826465,
+      "tokens_seen": 930086912
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007253249879634087,
+      "loss": 0.0751,
+      "theoretical_loss": 3.702985000877163,
+      "tokens_seen": 930349056
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": 0.002034248784184456,
+      "objective/train/docs_used": 343003,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.581547737121582,
+      "objective/train/original_loss": 1.581547737121582,
+      "objective/train/theoretical_loss": 3.7028851670847267,
+      "objective/train/tokens_used": 951071200,
+      "objective/train/value_avg": -0.007381439208984375,
+      "objective/train/value_loss": 0.0002921258274000138,
+      "objective/train/value_max": -0.00011324882507324219,
+      "objective/train/value_min": -0.54833984375,
+      "objective/train/value_reward_corr": 0.569236282685358,
+      "objective/train/value_std": 0.01422882080078125,
+      "objective/train/weight_avg": 1.002168893814087,
+      "objective/train/weighted_lm_loss": 1.5849435329437256,
+      "objective/train/weights_max": 1.5144890546798706,
+      "objective/train/weights_min": 0.36928337812423706,
+      "theoretical_loss": 3.7028851670847267,
+      "tokens_seen": 930611200
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007252447440218264,
+      "loss": 0.0746,
+      "theoretical_loss": 3.7028851670847267,
+      "tokens_seen": 930611200
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000725164500080244,
+      "loss": 0.0729,
+      "theoretical_loss": 3.7027853692822124,
+      "tokens_seen": 930873344
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007250842561386616,
+      "loss": 0.076,
+      "theoretical_loss": 3.702685607446516,
+      "tokens_seen": 931135488
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007250040121970791,
+      "loss": 0.0761,
+      "theoretical_loss": 3.7025858815545543,
+      "tokens_seen": 931397632
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007249237682554968,
+      "loss": 0.0746,
+      "theoretical_loss": 3.7024861915832665,
+      "tokens_seen": 931659776
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007248435243139143,
+      "loss": 0.0739,
+      "theoretical_loss": 3.7023865375096126,
+      "tokens_seen": 931921920
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007247632803723318,
+      "loss": 0.0758,
+      "theoretical_loss": 3.7022869193105734,
+      "tokens_seen": 932184064
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007246830364307495,
+      "loss": 0.073,
+      "theoretical_loss": 3.702187336963151,
+      "tokens_seen": 932446208
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000724602792489167,
+      "loss": 0.0742,
+      "theoretical_loss": 3.70208779044437,
+      "tokens_seen": 932708352
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007245225485475848,
+      "loss": 0.0719,
+      "theoretical_loss": 3.7019882797312746,
+      "tokens_seen": 932970496
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007244423046060023,
+      "loss": 0.0795,
+      "theoretical_loss": 3.701888804800931,
+      "tokens_seen": 933232640
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007243620606644199,
+      "loss": 0.0709,
+      "theoretical_loss": 3.701789365630426,
+      "tokens_seen": 933494784
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007242818167228375,
+      "loss": 0.0728,
+      "theoretical_loss": 3.701689962196868,
+      "tokens_seen": 933756928
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": 0.0019412703113630414,
+      "objective/train/docs_used": 344002,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6019786596298218,
+      "objective/train/original_loss": 1.6019787788391113,
+      "objective/train/theoretical_loss": 3.7016402738742964,
+      "objective/train/tokens_used": 954348000,
+      "objective/train/value_avg": -0.00753021240234375,
+      "objective/train/value_loss": 0.00015285314293578267,
+      "objective/train/value_max": -0.00015234947204589844,
+      "objective/train/value_min": -0.267822265625,
+      "objective/train/value_reward_corr": 0.5820011602921469,
+      "objective/train/value_std": 0.0103759765625,
+      "objective/train/weight_avg": 1.0020129680633545,
+      "objective/train/weighted_lm_loss": 1.6059612035751343,
+      "objective/train/weights_max": 1.2121343612670898,
+      "objective/train/weights_min": 0.3773167133331299,
+      "theoretical_loss": 3.7016402738742964,
+      "tokens_seen": 933888000
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000724201572781255,
+      "loss": 0.0768,
+      "theoretical_loss": 3.701590594477387,
+      "tokens_seen": 934019072
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007241213288396726,
+      "loss": 0.0769,
+      "theoretical_loss": 3.701491262449131,
+      "tokens_seen": 934281216
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007240410848980902,
+      "loss": 0.0765,
+      "theoretical_loss": 3.7013919660892736,
+      "tokens_seen": 934543360
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007239608409565078,
+      "loss": 0.0742,
+      "theoretical_loss": 3.701292705375006,
+      "tokens_seen": 934805504
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007238805970149253,
+      "loss": 0.0755,
+      "theoretical_loss": 3.701193480283542,
+      "tokens_seen": 935067648
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000723800353073343,
+      "loss": 0.0718,
+      "theoretical_loss": 3.701094290792116,
+      "tokens_seen": 935329792
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007237201091317606,
+      "loss": 0.0754,
+      "theoretical_loss": 3.7009951368779825,
+      "tokens_seen": 935591936
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007236398651901781,
+      "loss": 0.073,
+      "theoretical_loss": 3.700896018518418,
+      "tokens_seen": 935854080
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007235596212485958,
+      "loss": 0.0756,
+      "theoretical_loss": 3.700796935690719,
+      "tokens_seen": 936116224
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007234793773070133,
+      "loss": 0.0737,
+      "theoretical_loss": 3.700697888372204,
+      "tokens_seen": 936378368
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007233991333654309,
+      "loss": 0.0746,
+      "theoretical_loss": 3.7005988765402114,
+      "tokens_seen": 936640512
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007233188894238485,
+      "loss": 0.0737,
+      "theoretical_loss": 3.700499900172101,
+      "tokens_seen": 936902656
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": 0.0009387860773131251,
+      "objective/train/docs_used": 344763,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4261451959609985,
+      "objective/train/original_loss": 1.426145315170288,
+      "objective/train/theoretical_loss": 3.700400959245252,
+      "objective/train/tokens_used": 957624800,
+      "objective/train/value_avg": -0.01166534423828125,
+      "objective/train/value_loss": 0.0004122898681089282,
+      "objective/train/value_max": -0.0001398324966430664,
+      "objective/train/value_min": -0.81689453125,
+      "objective/train/value_reward_corr": 0.8483639387700215,
+      "objective/train/value_std": 0.03436279296875,
+      "objective/train/weight_avg": 1.0011343955993652,
+      "objective/train/weighted_lm_loss": 1.4269537925720215,
+      "objective/train/weights_max": 1.7235952615737915,
+      "objective/train/weights_min": 0.3695659041404724,
+      "theoretical_loss": 3.700400959245252,
+      "tokens_seen": 937164800
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007232386454822661,
+      "loss": 0.0716,
+      "theoretical_loss": 3.700400959245252,
+      "tokens_seen": 937164800
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007231584015406837,
+      "loss": 0.0732,
+      "theoretical_loss": 3.7003020537370657,
+      "tokens_seen": 937426944
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007230781575991013,
+      "loss": 0.0736,
+      "theoretical_loss": 3.7002031836249643,
+      "tokens_seen": 937689088
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007229979136575189,
+      "loss": 0.0756,
+      "theoretical_loss": 3.7001043488863896,
+      "tokens_seen": 937951232
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007229176697159365,
+      "loss": 0.0753,
+      "theoretical_loss": 3.7000055494988047,
+      "tokens_seen": 938213376
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007228374257743541,
+      "loss": 0.0772,
+      "theoretical_loss": 3.6999067854396936,
+      "tokens_seen": 938475520
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007227571818327716,
+      "loss": 0.0715,
+      "theoretical_loss": 3.6998080566865608,
+      "tokens_seen": 938737664
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007226769378911893,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6997093632169307,
+      "tokens_seen": 938999808
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007225966939496068,
+      "loss": 0.0768,
+      "theoretical_loss": 3.699610705008349,
+      "tokens_seen": 939261952
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007225164500080244,
+      "loss": 0.0707,
+      "theoretical_loss": 3.6995120820383818,
+      "tokens_seen": 939524096
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.000722436206066442,
+      "loss": 0.074,
+      "theoretical_loss": 3.6994134942846157,
+      "tokens_seen": 939786240
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007223559621248595,
+      "loss": 0.0758,
+      "theoretical_loss": 3.6993149417246576,
+      "tokens_seen": 940048384
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 0.0007222757181832772,
+      "loss": 0.0737,
+      "theoretical_loss": 3.699216424336135,
+      "tokens_seen": 940310528
+    },
+    {
+      "epoch": 0.28,
+      "objective/train/advantage_avg": 0.0008543147123418748,
+      "objective/train/docs_used": 345855,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.430222988128662,
+      "objective/train/original_loss": 1.430222988128662,
+      "objective/train/theoretical_loss": 3.6991671788241764,
+      "objective/train/tokens_used": 960901600,
+      "objective/train/value_avg": -0.00884246826171875,
+      "objective/train/value_loss": 0.00023561967827845365,
+      "objective/train/value_max": -0.00021660327911376953,
+      "objective/train/value_min": -0.6904296875,
+      "objective/train/value_reward_corr": 0.6824104864233338,
+      "objective/train/value_std": 0.0145416259765625,
+      "objective/train/weight_avg": 1.0009649991989136,
+      "objective/train/weighted_lm_loss": 1.4308393001556396,
+      "objective/train/weights_max": 1.3335462808609009,
+      "objective/train/weights_min": 0.3730488717556,
+      "theoretical_loss": 3.6991671788241764,
+      "tokens_seen": 940441600
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007221954742416948,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6991179420966964,
+      "tokens_seen": 940572672
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007221152303001124,
+      "loss": 0.0729,
+      "theoretical_loss": 3.69901949498401,
+      "tokens_seen": 940834816
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007220349863585299,
+      "loss": 0.0759,
+      "theoretical_loss": 3.6989210829757644,
+      "tokens_seen": 941096960
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007219547424169476,
+      "loss": 0.0758,
+      "theoretical_loss": 3.6988227060496692,
+      "tokens_seen": 941359104
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007218744984753651,
+      "loss": 0.0761,
+      "theoretical_loss": 3.6987243641834535,
+      "tokens_seen": 941621248
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007217942545337827,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6986260573548675,
+      "tokens_seen": 941883392
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007217140105922003,
+      "loss": 0.0746,
+      "theoretical_loss": 3.698527785541682,
+      "tokens_seen": 942145536
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007216337666506178,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6984295487216867,
+      "tokens_seen": 942407680
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007215535227090356,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6983313468726924,
+      "tokens_seen": 942669824
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007214732787674531,
+      "loss": 0.0773,
+      "theoretical_loss": 3.6982331799725303,
+      "tokens_seen": 942931968
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007213930348258707,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6981350479990525,
+      "tokens_seen": 943194112
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007213127908842883,
+      "loss": 0.0781,
+      "theoretical_loss": 3.6980369509301285,
+      "tokens_seen": 943456256
+    },
+    {
+      "epoch": 0.29,
+      "objective/train/advantage_avg": 0.0011019601952284575,
+      "objective/train/docs_used": 346959,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3961070775985718,
+      "objective/train/original_loss": 1.3961069583892822,
+      "objective/train/theoretical_loss": 3.6979388887436517,
+      "objective/train/tokens_used": 964178400,
+      "objective/train/value_avg": -0.00862884521484375,
+      "objective/train/value_loss": 0.0001508643908891827,
+      "objective/train/value_max": -0.00017535686492919922,
+      "objective/train/value_min": -0.24072265625,
+      "objective/train/value_reward_corr": 0.7373741548355937,
+      "objective/train/value_std": 0.012786865234375,
+      "objective/train/weight_avg": 1.0011765956878662,
+      "objective/train/weighted_lm_loss": 1.3976231813430786,
+      "objective/train/weights_max": 1.2201510667800903,
+      "objective/train/weights_min": 0.8218066692352295,
+      "theoretical_loss": 3.6979388887436517,
+      "tokens_seen": 943718400
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007212325469427058,
+      "loss": 0.0758,
+      "theoretical_loss": 3.6979388887436517,
+      "tokens_seen": 943718400
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007211523030011234,
+      "loss": 0.0759,
+      "theoretical_loss": 3.697840861417533,
+      "tokens_seen": 943980544
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.000721072059059541,
+      "loss": 0.0755,
+      "theoretical_loss": 3.697742868929704,
+      "tokens_seen": 944242688
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007209918151179586,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6976449112581173,
+      "tokens_seen": 944504832
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007209115711763761,
+      "loss": 0.0746,
+      "theoretical_loss": 3.697546988380744,
+      "tokens_seen": 944766976
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007208313272347939,
+      "loss": 0.076,
+      "theoretical_loss": 3.697449100275577,
+      "tokens_seen": 945029120
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007207510832932114,
+      "loss": 0.0726,
+      "theoretical_loss": 3.6973512469206278,
+      "tokens_seen": 945291264
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.000720670839351629,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6972534282939282,
+      "tokens_seen": 945553408
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007205905954100466,
+      "loss": 0.0726,
+      "theoretical_loss": 3.6971556443735314,
+      "tokens_seen": 945815552
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007205103514684641,
+      "loss": 0.0711,
+      "theoretical_loss": 3.697057895137508,
+      "tokens_seen": 946077696
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007204301075268818,
+      "loss": 0.0772,
+      "theoretical_loss": 3.696960180563951,
+      "tokens_seen": 946339840
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007203498635852993,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6968625006309717,
+      "tokens_seen": 946601984
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007202696196437169,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6967648553167014,
+      "tokens_seen": 946864128
+    },
+    {
+      "epoch": 0.29,
+      "objective/train/advantage_avg": -8.704046194907278e-05,
+      "objective/train/docs_used": 348186,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.480729579925537,
+      "objective/train/original_loss": 1.4807298183441162,
+      "objective/train/theoretical_loss": 3.696716045634754,
+      "objective/train/tokens_used": 967455200,
+      "objective/train/value_avg": -0.0102691650390625,
+      "objective/train/value_loss": 0.0004358619044069201,
+      "objective/train/value_max": -0.00021660327911376953,
+      "objective/train/value_min": -0.783203125,
+      "objective/train/value_reward_corr": 0.7756683391643069,
+      "objective/train/value_std": 0.0248260498046875,
+      "objective/train/weight_avg": 1.0001178979873657,
+      "objective/train/weighted_lm_loss": 1.4807584285736084,
+      "objective/train/weights_max": 2.014146089553833,
+      "objective/train/weights_min": 0.3976996839046478,
+      "theoretical_loss": 3.696716045634754,
+      "tokens_seen": 946995200
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007201893757021345,
+      "loss": 0.0738,
+      "theoretical_loss": 3.696667244599292,
+      "tokens_seen": 947126272
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.000720109131760552,
+      "loss": 0.0759,
+      "theoretical_loss": 3.6965696684569154,
+      "tokens_seen": 947388416
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007200288878189697,
+      "loss": 0.0763,
+      "theoretical_loss": 3.6964721268677616,
+      "tokens_seen": 947650560
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007199486438773873,
+      "loss": 0.0744,
+      "theoretical_loss": 3.696374619810043,
+      "tokens_seen": 947912704
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007198683999358049,
+      "loss": 0.0715,
+      "theoretical_loss": 3.6962771472619886,
+      "tokens_seen": 948174848
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007197881559942224,
+      "loss": 0.0743,
+      "theoretical_loss": 3.69617970920185,
+      "tokens_seen": 948436992
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007197079120526401,
+      "loss": 0.0766,
+      "theoretical_loss": 3.6960823056078973,
+      "tokens_seen": 948699136
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007196276681110576,
+      "loss": 0.0765,
+      "theoretical_loss": 3.6959849364584203,
+      "tokens_seen": 948961280
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007195474241694752,
+      "loss": 0.0751,
+      "theoretical_loss": 3.695887601731728,
+      "tokens_seen": 949223424
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007194671802278928,
+      "loss": 0.0765,
+      "theoretical_loss": 3.69579030140615,
+      "tokens_seen": 949485568
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007193869362863103,
+      "loss": 0.0751,
+      "theoretical_loss": 3.6956930354600352,
+      "tokens_seen": 949747712
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007193066923447281,
+      "loss": 0.075,
+      "theoretical_loss": 3.6955958038717522,
+      "tokens_seen": 950009856
+    },
+    {
+      "epoch": 0.29,
+      "objective/train/advantage_avg": 0.0015586912631988525,
+      "objective/train/docs_used": 349336,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4084975719451904,
+      "objective/train/original_loss": 1.4084975719451904,
+      "objective/train/theoretical_loss": 3.695498606619688,
+      "objective/train/tokens_used": 970732000,
+      "objective/train/value_avg": -0.00797271728515625,
+      "objective/train/value_loss": 0.0002061313862213865,
+      "objective/train/value_max": -0.00015115737915039062,
+      "objective/train/value_min": -0.53125,
+      "objective/train/value_reward_corr": 0.6561778993169798,
+      "objective/train/value_std": 0.0132598876953125,
+      "objective/train/weight_avg": 1.0016520023345947,
+      "objective/train/weighted_lm_loss": 1.411009669303894,
+      "objective/train/weights_max": 1.2708643674850464,
+      "objective/train/weights_min": 0.36876535415649414,
+      "theoretical_loss": 3.695498606619688,
+      "tokens_seen": 950272000
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007192264484031456,
+      "loss": 0.0747,
+      "theoretical_loss": 3.695498606619688,
+      "tokens_seen": 950272000
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007191462044615632,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6954014436822513,
+      "tokens_seen": 950534144
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007190659605199808,
+      "loss": 0.0749,
+      "theoretical_loss": 3.695304315037868,
+      "tokens_seen": 950796288
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007189857165783984,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6952072206649857,
+      "tokens_seen": 951058432
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007189054726368159,
+      "loss": 0.075,
+      "theoretical_loss": 3.695110160542069,
+      "tokens_seen": 951320576
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007188252286952335,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6950131346476054,
+      "tokens_seen": 951582720
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007187449847536511,
+      "loss": 0.0755,
+      "theoretical_loss": 3.694916142960098,
+      "tokens_seen": 951844864
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007186647408120686,
+      "loss": 0.0723,
+      "theoretical_loss": 3.6948191854580728,
+      "tokens_seen": 952107008
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007185844968704864,
+      "loss": 0.0738,
+      "theoretical_loss": 3.694722262120072,
+      "tokens_seen": 952369152
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007185042529289039,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6946253729246594,
+      "tokens_seen": 952631296
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007184240089873215,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6945285178504172,
+      "tokens_seen": 952893440
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007183437650457391,
+      "loss": 0.0729,
+      "theoretical_loss": 3.694431696875948,
+      "tokens_seen": 953155584
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007182635211041566,
+      "loss": 0.0778,
+      "theoretical_loss": 3.6943349099798715,
+      "tokens_seen": 953417728
+    },
+    {
+      "epoch": 0.29,
+      "objective/train/advantage_avg": 0.0013693609507754445,
+      "objective/train/docs_used": 350619,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3620078563690186,
+      "objective/train/original_loss": 1.3620076179504395,
+      "objective/train/theoretical_loss": 3.694286529304555,
+      "objective/train/tokens_used": 974008800,
+      "objective/train/value_avg": -0.005817413330078125,
+      "objective/train/value_loss": 0.00011372018343536183,
+      "objective/train/value_max": -0.0002065896987915039,
+      "objective/train/value_min": -0.55126953125,
+      "objective/train/value_reward_corr": 0.6999231970698898,
+      "objective/train/value_std": 0.009918212890625,
+      "objective/train/weight_avg": 1.0014216899871826,
+      "objective/train/weighted_lm_loss": 1.3646951913833618,
+      "objective/train/weights_max": 1.1521174907684326,
+      "objective/train/weights_min": 0.3827209174633026,
+      "theoretical_loss": 3.694286529304555,
+      "tokens_seen": 953548800
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007181832771625743,
+      "loss": 0.0708,
+      "theoretical_loss": 3.6942381571408287,
+      "tokens_seen": 953679872
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007181030332209918,
+      "loss": 0.0761,
+      "theoretical_loss": 3.6941414383374793,
+      "tokens_seen": 953942016
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007180227892794094,
+      "loss": 0.0726,
+      "theoretical_loss": 3.6940447535485026,
+      "tokens_seen": 954204160
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.000717942545337827,
+      "loss": 0.0736,
+      "theoretical_loss": 3.6939481027525956,
+      "tokens_seen": 954466304
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007178623013962447,
+      "loss": 0.074,
+      "theoretical_loss": 3.6938514859284766,
+      "tokens_seen": 954728448
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007177820574546622,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6937549030548813,
+      "tokens_seen": 954990592
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007177018135130798,
+      "loss": 0.0749,
+      "theoretical_loss": 3.693658354110565,
+      "tokens_seen": 955252736
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007176215695714974,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6935618390743032,
+      "tokens_seen": 955514880
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007175413256299149,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6934653579248886,
+      "tokens_seen": 955777024
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007174610816883326,
+      "loss": 0.0729,
+      "theoretical_loss": 3.693368910641135,
+      "tokens_seen": 956039168
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007173808377467501,
+      "loss": 0.0761,
+      "theoretical_loss": 3.693272497201874,
+      "tokens_seen": 956301312
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007173005938051677,
+      "loss": 0.0777,
+      "theoretical_loss": 3.6931761175859554,
+      "tokens_seen": 956563456
+    },
+    {
+      "epoch": 0.29,
+      "objective/train/advantage_avg": 0.0013215347426012158,
+      "objective/train/docs_used": 351696,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4953937530517578,
+      "objective/train/original_loss": 1.4953936338424683,
+      "objective/train/theoretical_loss": 3.693079771772251,
+      "objective/train/tokens_used": 977285600,
+      "objective/train/value_avg": -0.005706787109375,
+      "objective/train/value_loss": 0.00014413167082238942,
+      "objective/train/value_max": -0.00010889768600463867,
+      "objective/train/value_min": -0.2381591796875,
+      "objective/train/value_reward_corr": 0.5217617356406479,
+      "objective/train/value_std": 0.00839996337890625,
+      "objective/train/weight_avg": 1.0013859272003174,
+      "objective/train/weighted_lm_loss": 1.4972920417785645,
+      "objective/train/weights_max": 1.1233699321746826,
+      "objective/train/weights_min": 0.3683270514011383,
+      "theoretical_loss": 3.693079771772251,
+      "tokens_seen": 956825600
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007172203498635853,
+      "loss": 0.0747,
+      "theoretical_loss": 3.693079771772251,
+      "tokens_seen": 956825600
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007171401059220028,
+      "loss": 0.0763,
+      "theoretical_loss": 3.692983459739649,
+      "tokens_seen": 957087744
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007170598619804205,
+      "loss": 0.0776,
+      "theoretical_loss": 3.6928871814670563,
+      "tokens_seen": 957349888
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007169796180388381,
+      "loss": 0.0763,
+      "theoretical_loss": 3.6927909369334007,
+      "tokens_seen": 957612032
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007168993740972557,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6926947261176277,
+      "tokens_seen": 957874176
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007168191301556733,
+      "loss": 0.0762,
+      "theoretical_loss": 3.692598548998702,
+      "tokens_seen": 958136320
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007167388862140909,
+      "loss": 0.0773,
+      "theoretical_loss": 3.692502405555606,
+      "tokens_seen": 958398464
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007166586422725084,
+      "loss": 0.077,
+      "theoretical_loss": 3.692406295767344,
+      "tokens_seen": 958660608
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007165783983309261,
+      "loss": 0.0753,
+      "theoretical_loss": 3.692310219612936,
+      "tokens_seen": 958922752
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007164981543893436,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6922141770714214,
+      "tokens_seen": 959184896
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007164179104477611,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6921181681218602,
+      "tokens_seen": 959447040
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007163376665061789,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6920221927433294,
+      "tokens_seen": 959709184
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007162574225645964,
+      "loss": 0.0772,
+      "theoretical_loss": 3.691926250914925,
+      "tokens_seen": 959971328
+    },
+    {
+      "epoch": 0.29,
+      "objective/train/advantage_avg": 0.0015116332797333598,
+      "objective/train/docs_used": 352874,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4265995025634766,
+      "objective/train/original_loss": 1.4265992641448975,
+      "objective/train/theoretical_loss": 3.6918782925754936,
+      "objective/train/tokens_used": 980562400,
+      "objective/train/value_avg": -0.0121917724609375,
+      "objective/train/value_loss": 0.00023754734138492495,
+      "objective/train/value_max": -0.00017130374908447266,
+      "objective/train/value_min": -0.31884765625,
+      "objective/train/value_reward_corr": 0.9000245293805224,
+      "objective/train/value_std": 0.03515625,
+      "objective/train/weight_avg": 1.0016270875930786,
+      "objective/train/weighted_lm_loss": 1.4295847415924072,
+      "objective/train/weights_max": 1.215097427368164,
+      "objective/train/weights_min": 0.3761727213859558,
+      "theoretical_loss": 3.6918782925754936,
+      "tokens_seen": 960102400
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.000716177178623014,
+      "loss": 0.0754,
+      "theoretical_loss": 3.691830342615763,
+      "tokens_seen": 960233472
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007160969346814316,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6917344678249755,
+      "tokens_seen": 960495616
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007160166907398492,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6916386265217156,
+      "tokens_seen": 960757760
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007159364467982667,
+      "loss": 0.0726,
+      "theoretical_loss": 3.6915428186851553,
+      "tokens_seen": 961019904
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007158562028566843,
+      "loss": 0.075,
+      "theoretical_loss": 3.6914470442944824,
+      "tokens_seen": 961282048
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007157759589151019,
+      "loss": 0.0722,
+      "theoretical_loss": 3.691351303328907,
+      "tokens_seen": 961544192
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007156957149735194,
+      "loss": 0.0754,
+      "theoretical_loss": 3.691255595767654,
+      "tokens_seen": 961806336
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007156154710319372,
+      "loss": 0.0766,
+      "theoretical_loss": 3.6911599215899704,
+      "tokens_seen": 962068480
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007155352270903547,
+      "loss": 0.0759,
+      "theoretical_loss": 3.6910642807751195,
+      "tokens_seen": 962330624
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007154549831487724,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6909686733023843,
+      "tokens_seen": 962592768
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007153747392071899,
+      "loss": 0.0722,
+      "theoretical_loss": 3.690873099151065,
+      "tokens_seen": 962854912
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007152944952656074,
+      "loss": 0.0729,
+      "theoretical_loss": 3.690777558300482,
+      "tokens_seen": 963117056
+    },
+    {
+      "epoch": 0.29,
+      "objective/train/advantage_avg": 0.000455680739833042,
+      "objective/train/docs_used": 353969,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4282293319702148,
+      "objective/train/original_loss": 1.428229570388794,
+      "objective/train/theoretical_loss": 3.690682050729972,
+      "objective/train/tokens_used": 983839200,
+      "objective/train/value_avg": -0.010162353515625,
+      "objective/train/value_loss": 0.0002770853752736002,
+      "objective/train/value_max": -0.00014543533325195312,
+      "objective/train/value_min": -0.2489013671875,
+      "objective/train/value_reward_corr": 0.6144828349212833,
+      "objective/train/value_std": 0.01318359375,
+      "objective/train/weight_avg": 1.0005788803100586,
+      "objective/train/weighted_lm_loss": 1.4285677671432495,
+      "objective/train/weights_max": 1.145114541053772,
+      "objective/train/weights_min": 0.37042829394340515,
+      "theoretical_loss": 3.690682050729972,
+      "tokens_seen": 963379200
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007152142513240251,
+      "loss": 0.0734,
+      "theoretical_loss": 3.690682050729972,
+      "tokens_seen": 963379200
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007151340073824426,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6905865764188923,
+      "tokens_seen": 963641344
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007150537634408602,
+      "loss": 0.0722,
+      "theoretical_loss": 3.6904911353466177,
+      "tokens_seen": 963903488
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007149735194992778,
+      "loss": 0.0754,
+      "theoretical_loss": 3.690395727492541,
+      "tokens_seen": 964165632
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007148932755576955,
+      "loss": 0.0733,
+      "theoretical_loss": 3.690300352836074,
+      "tokens_seen": 964427776
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.000714813031616113,
+      "loss": 0.077,
+      "theoretical_loss": 3.690205011356646,
+      "tokens_seen": 964689920
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007147327876745306,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6901097030337056,
+      "tokens_seen": 964952064
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007146525437329482,
+      "loss": 0.0727,
+      "theoretical_loss": 3.6900144278467204,
+      "tokens_seen": 965214208
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007145722997913657,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6899191857751736,
+      "tokens_seen": 965476352
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007144920558497834,
+      "loss": 0.0723,
+      "theoretical_loss": 3.6898239767985688,
+      "tokens_seen": 965738496
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007144118119082009,
+      "loss": 0.0754,
+      "theoretical_loss": 3.689728800896428,
+      "tokens_seen": 966000640
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007143315679666186,
+      "loss": 0.0735,
+      "theoretical_loss": 3.68963365804829,
+      "tokens_seen": 966262784
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007142513240250361,
+      "loss": 0.076,
+      "theoretical_loss": 3.689538548233713,
+      "tokens_seen": 966524928
+    },
+    {
+      "epoch": 0.29,
+      "objective/train/advantage_avg": 0.00043955291039310396,
+      "objective/train/docs_used": 355190,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5631569623947144,
+      "objective/train/original_loss": 1.5631568431854248,
+      "objective/train/theoretical_loss": 3.6894910057076267,
+      "objective/train/tokens_used": 987116000,
+      "objective/train/value_avg": -0.00716400146484375,
+      "objective/train/value_loss": 0.0002913126372732222,
+      "objective/train/value_max": -0.00010150671005249023,
+      "objective/train/value_min": -0.6953125,
+      "objective/train/value_reward_corr": 0.7267076806923944,
+      "objective/train/value_std": 0.0150299072265625,
+      "objective/train/weight_avg": 1.0005711317062378,
+      "objective/train/weighted_lm_loss": 1.5634570121765137,
+      "objective/train/weights_max": 1.3262733221054077,
+      "objective/train/weights_min": 0.3728724420070648,
+      "theoretical_loss": 3.6894910057076267,
+      "tokens_seen": 966656000
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007141710800834536,
+      "loss": 0.0774,
+      "theoretical_loss": 3.6894434714322726,
+      "tokens_seen": 966787072
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007140908361418714,
+      "loss": 0.0731,
+      "theoretical_loss": 3.689348427623563,
+      "tokens_seen": 967049216
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007140105922002889,
+      "loss": 0.0726,
+      "theoretical_loss": 3.689253416787197,
+      "tokens_seen": 967311360
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007139303482587065,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6891584389028047,
+      "tokens_seen": 967573504
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007138501043171241,
+      "loss": 0.0746,
+      "theoretical_loss": 3.689063493950034,
+      "tokens_seen": 967835648
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007137698603755417,
+      "loss": 0.0736,
+      "theoretical_loss": 3.6889685819085525,
+      "tokens_seen": 968097792
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007136896164339592,
+      "loss": 0.0744,
+      "theoretical_loss": 3.688873702758044,
+      "tokens_seen": 968359936
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007136093724923768,
+      "loss": 0.073,
+      "theoretical_loss": 3.688778856478211,
+      "tokens_seen": 968622080
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007135291285507944,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6886840430487746,
+      "tokens_seen": 968884224
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007134488846092119,
+      "loss": 0.0726,
+      "theoretical_loss": 3.688589262449474,
+      "tokens_seen": 969146368
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007133686406676297,
+      "loss": 0.0754,
+      "theoretical_loss": 3.6884945146600643,
+      "tokens_seen": 969408512
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007132883967260472,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6883997996603215,
+      "tokens_seen": 969670656
+    },
+    {
+      "epoch": 0.29,
+      "objective/train/advantage_avg": 0.0019111091969534755,
+      "objective/train/docs_used": 356486,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4429503679275513,
+      "objective/train/original_loss": 1.4429501295089722,
+      "objective/train/theoretical_loss": 3.688305117430038,
+      "objective/train/tokens_used": 990392800,
+      "objective/train/value_avg": -0.0103302001953125,
+      "objective/train/value_loss": 0.0004549300647340715,
+      "objective/train/value_max": -8.749961853027344e-05,
+      "objective/train/value_min": -0.64794921875,
+      "objective/train/value_reward_corr": 0.6670057748618838,
+      "objective/train/value_std": 0.0211181640625,
+      "objective/train/weight_avg": 1.0021229982376099,
+      "objective/train/weighted_lm_loss": 1.446812391281128,
+      "objective/train/weights_max": 1.8136069774627686,
+      "objective/train/weights_min": 0.3739864230155945,
+      "theoretical_loss": 3.688305117430038,
+      "tokens_seen": 969932800
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007132081527844648,
+      "loss": 0.0744,
+      "theoretical_loss": 3.688305117430038,
+      "tokens_seen": 969932800
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007131279088428824,
+      "loss": 0.0744,
+      "theoretical_loss": 3.688210467949023,
+      "tokens_seen": 970194944
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007130476649013,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6881158511971055,
+      "tokens_seen": 970457088
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007129674209597176,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6880212671541326,
+      "tokens_seen": 970719232
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007128871770181351,
+      "loss": 0.0752,
+      "theoretical_loss": 3.687926715799967,
+      "tokens_seen": 970981376
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007128069330765527,
+      "loss": 0.0746,
+      "theoretical_loss": 3.687832197114491,
+      "tokens_seen": 971243520
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007127266891349703,
+      "loss": 0.0762,
+      "theoretical_loss": 3.687737711077605,
+      "tokens_seen": 971505664
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.000712646445193388,
+      "loss": 0.0739,
+      "theoretical_loss": 3.687643257669225,
+      "tokens_seen": 971767808
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007125662012518055,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6875488368692877,
+      "tokens_seen": 972029952
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007124859573102232,
+      "loss": 0.0736,
+      "theoretical_loss": 3.687454448657745,
+      "tokens_seen": 972292096
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007124057133686407,
+      "loss": 0.0741,
+      "theoretical_loss": 3.687360093014568,
+      "tokens_seen": 972554240
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007123254694270582,
+      "loss": 0.0744,
+      "theoretical_loss": 3.687265769919745,
+      "tokens_seen": 972816384
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007122452254854759,
+      "loss": 0.0753,
+      "theoretical_loss": 3.6871714793532826,
+      "tokens_seen": 973078528
+    },
+    {
+      "epoch": 0.29,
+      "objective/train/advantage_avg": 0.000875156547408551,
+      "objective/train/docs_used": 357717,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.565974235534668,
+      "objective/train/original_loss": 1.5659743547439575,
+      "objective/train/theoretical_loss": 3.6871243462619425,
+      "objective/train/tokens_used": 993669600,
+      "objective/train/value_avg": -0.00701141357421875,
+      "objective/train/value_loss": 0.000197466419194825,
+      "objective/train/value_max": -8.958578109741211e-05,
+      "objective/train/value_min": -0.229248046875,
+      "objective/train/value_reward_corr": 0.6185040343790946,
+      "objective/train/value_std": 0.0106201171875,
+      "objective/train/weight_avg": 1.0009616613388062,
+      "objective/train/weighted_lm_loss": 1.5670831203460693,
+      "objective/train/weights_max": 1.1158901453018188,
+      "objective/train/weights_min": 0.3683288097381592,
+      "theoretical_loss": 3.6871243462619425,
+      "tokens_seen": 973209600
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 0.0007121649815438934,
+      "loss": 0.0768,
+      "theoretical_loss": 3.687077221295203,
+      "tokens_seen": 973340672
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000712084737602311,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6869829957255496,
+      "tokens_seen": 973602816
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007120044936607286,
+      "loss": 0.0734,
+      "theoretical_loss": 3.68688880262438,
+      "tokens_seen": 973864960
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007119242497191462,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6867946419717716,
+      "tokens_seen": 974127104
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007118440057775638,
+      "loss": 0.0726,
+      "theoretical_loss": 3.6867005137478177,
+      "tokens_seen": 974389248
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007117637618359814,
+      "loss": 0.0768,
+      "theoretical_loss": 3.686606417932631,
+      "tokens_seen": 974651392
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000711683517894399,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6865123545063403,
+      "tokens_seen": 974913536
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007116032739528166,
+      "loss": 0.0768,
+      "theoretical_loss": 3.686418323449093,
+      "tokens_seen": 975175680
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007115230300112342,
+      "loss": 0.0765,
+      "theoretical_loss": 3.6863243247410526,
+      "tokens_seen": 975437824
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007114427860696517,
+      "loss": 0.0735,
+      "theoretical_loss": 3.686230358362401,
+      "tokens_seen": 975699968
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007113625421280694,
+      "loss": 0.0739,
+      "theoretical_loss": 3.686136424293338,
+      "tokens_seen": 975962112
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007112822981864869,
+      "loss": 0.074,
+      "theoretical_loss": 3.68604252251408,
+      "tokens_seen": 976224256
+    },
+    {
+      "epoch": 0.3,
+      "objective/train/advantage_avg": 0.001195325399748981,
+      "objective/train/docs_used": 358882,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3014249801635742,
+      "objective/train/original_loss": 1.3014247417449951,
+      "objective/train/theoretical_loss": 3.6859486530048615,
+      "objective/train/tokens_used": 996946400,
+      "objective/train/value_avg": -0.00737762451171875,
+      "objective/train/value_loss": 0.00010647853196132928,
+      "objective/train/value_max": -0.00010889768600463867,
+      "objective/train/value_min": -0.2327880859375,
+      "objective/train/value_reward_corr": 0.708244856375411,
+      "objective/train/value_std": 0.009979248046875,
+      "objective/train/weight_avg": 1.0012476444244385,
+      "objective/train/weighted_lm_loss": 1.3023656606674194,
+      "objective/train/weights_max": 1.2456159591674805,
+      "objective/train/weights_min": 0.611983060836792,
+      "theoretical_loss": 3.6859486530048615,
+      "tokens_seen": 976486400
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007112020542449044,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6859486530048615,
+      "tokens_seen": 976486400
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007111218103033222,
+      "loss": 0.0736,
+      "theoretical_loss": 3.685854815745933,
+      "tokens_seen": 976748544
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007110415663617397,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6857610107175645,
+      "tokens_seen": 977010688
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007109613224201573,
+      "loss": 0.0771,
+      "theoretical_loss": 3.6856672379000415,
+      "tokens_seen": 977272832
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007108810784785749,
+      "loss": 0.0763,
+      "theoretical_loss": 3.6855734972736682,
+      "tokens_seen": 977534976
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007108008345369925,
+      "loss": 0.0721,
+      "theoretical_loss": 3.685479788818766,
+      "tokens_seen": 977797120
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.00071072059059541,
+      "loss": 0.0757,
+      "theoretical_loss": 3.6853861125156717,
+      "tokens_seen": 978059264
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007106403466538276,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6852924683447412,
+      "tokens_seen": 978321408
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007105601027122452,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6851988562863482,
+      "tokens_seen": 978583552
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007104798587706628,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6851052763208823,
+      "tokens_seen": 978845696
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007103996148290805,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6850117284287505,
+      "tokens_seen": 979107840
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000710319370887498,
+      "loss": 0.0708,
+      "theoretical_loss": 3.6849182125903774,
+      "tokens_seen": 979369984
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007102391269459157,
+      "loss": 0.0766,
+      "theoretical_loss": 3.6848247287862046,
+      "tokens_seen": 979632128
+    },
+    {
+      "epoch": 0.3,
+      "objective/train/advantage_avg": 0.0012934327824041247,
+      "objective/train/docs_used": 360128,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5772310495376587,
+      "objective/train/original_loss": 1.5772309303283691,
+      "objective/train/theoretical_loss": 3.6847779988908362,
+      "objective/train/tokens_used": 1000223200,
+      "objective/train/value_avg": -0.00775909423828125,
+      "objective/train/value_loss": 0.0002463175624143332,
+      "objective/train/value_max": -0.0002065896987915039,
+      "objective/train/value_min": -0.6826171875,
+      "objective/train/value_reward_corr": 0.5791152585073471,
+      "objective/train/value_std": 0.01091766357421875,
+      "objective/train/weight_avg": 1.0014005899429321,
+      "objective/train/weighted_lm_loss": 1.5800875425338745,
+      "objective/train/weights_max": 1.209107518196106,
+      "objective/train/weights_min": 0.36841312050819397,
+      "theoretical_loss": 3.6847779988908362,
+      "tokens_seen": 979763200
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007101588830043332,
+      "loss": 0.0759,
+      "theoretical_loss": 3.684731276996691,
+      "tokens_seen": 979894272
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007100786390627508,
+      "loss": 0.0749,
+      "theoretical_loss": 3.684637857202312,
+      "tokens_seen": 980156416
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007099983951211684,
+      "loss": 0.0742,
+      "theoretical_loss": 3.684544469383562,
+      "tokens_seen": 980418560
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007099181511795859,
+      "loss": 0.0726,
+      "theoretical_loss": 3.6844511135209497,
+      "tokens_seen": 980680704
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007098379072380035,
+      "loss": 0.0733,
+      "theoretical_loss": 3.684357789595003,
+      "tokens_seen": 980942848
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007097576632964211,
+      "loss": 0.0756,
+      "theoretical_loss": 3.684264497586266,
+      "tokens_seen": 981204992
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007096774193548388,
+      "loss": 0.0751,
+      "theoretical_loss": 3.684171237475301,
+      "tokens_seen": 981467136
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007095971754132563,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6840780092426852,
+      "tokens_seen": 981729280
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000709516931471674,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6839848128690145,
+      "tokens_seen": 981991424
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007094366875300915,
+      "loss": 0.0762,
+      "theoretical_loss": 3.683891648334901,
+      "tokens_seen": 982253568
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000709356443588509,
+      "loss": 0.0751,
+      "theoretical_loss": 3.6837985156209743,
+      "tokens_seen": 982515712
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007092761996469267,
+      "loss": 0.0774,
+      "theoretical_loss": 3.683705414707881,
+      "tokens_seen": 982777856
+    },
+    {
+      "debugging/Compilability": 1.0,
+      "debugging/distinct-1-grams": 0.7488293247094114,
+      "debugging/entropy-1-grams": 5.349564433264487,
+      "debugging/length": 462.46153846153845,
+      "debugging/num_segments": 13,
+      "debugging/raw_token_scores_avg": 0.00979544036090374,
+      "debugging/raw_token_scores_std": 0.022818773984909058,
+      "debugging/score": 0.0073770168052465215,
+      "debugging/score_std": 0.0054551366535427,
+      "epoch": 0.3,
+      "objective/train/advantage_avg": 0.0004159576492384076,
+      "objective/train/docs_used": 361383,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.380717158317566,
+      "objective/train/original_loss": 1.3807172775268555,
+      "objective/train/theoretical_loss": 3.6836123455762837,
+      "objective/train/tokens_used": 1003500000,
+      "objective/train/value_avg": -0.0102081298828125,
+      "objective/train/value_loss": 0.00025552461738698184,
+      "objective/train/value_max": -0.00016605854034423828,
+      "objective/train/value_min": -0.74853515625,
+      "objective/train/value_reward_corr": 0.7139572076837593,
+      "objective/train/value_std": 0.0162353515625,
+      "objective/train/weight_avg": 1.000531554222107,
+      "objective/train/weighted_lm_loss": 1.3806174993515015,
+      "objective/train/weights_max": 1.1966832876205444,
+      "objective/train/weights_min": 0.3853108584880829,
+      "theoretical_loss": 3.6836123455762837,
+      "tokens_seen": 983040000
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007091959557053442,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6836123455762837,
+      "tokens_seen": 983040000
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007091157117637619,
+      "loss": 0.075,
+      "theoretical_loss": 3.683519308206863,
+      "tokens_seen": 983302144
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007090354678221794,
+      "loss": 0.0749,
+      "theoretical_loss": 3.683426302580316,
+      "tokens_seen": 983564288
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000708955223880597,
+      "loss": 0.0761,
+      "theoretical_loss": 3.683333328677356,
+      "tokens_seen": 983826432
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007088749799390147,
+      "loss": 0.0707,
+      "theoretical_loss": 3.6832403864787144,
+      "tokens_seen": 984088576
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007087947359974322,
+      "loss": 0.0749,
+      "theoretical_loss": 3.683147475965139,
+      "tokens_seen": 984350720
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007087144920558498,
+      "loss": 0.076,
+      "theoretical_loss": 3.683054597117393,
+      "tokens_seen": 984612864
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007086342481142674,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6829617499162595,
+      "tokens_seen": 984875008
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000708554004172685,
+      "loss": 0.0751,
+      "theoretical_loss": 3.6828689343425345,
+      "tokens_seen": 985137152
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007084737602311025,
+      "loss": 0.0741,
+      "theoretical_loss": 3.682776150377034,
+      "tokens_seen": 985399296
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007083935162895202,
+      "loss": 0.0765,
+      "theoretical_loss": 3.682683398000589,
+      "tokens_seen": 985661440
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007083132723479377,
+      "loss": 0.0773,
+      "theoretical_loss": 3.6825906771940478,
+      "tokens_seen": 985923584
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007082330284063552,
+      "loss": 0.0738,
+      "theoretical_loss": 3.682497987938275,
+      "tokens_seen": 986185728
+    },
+    {
+      "epoch": 0.3,
+      "objective/train/advantage_avg": 0.002384291496127844,
+      "objective/train/docs_used": 362627,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.497067928314209,
+      "objective/train/original_loss": 1.4970680475234985,
+      "objective/train/theoretical_loss": 3.682451655135952,
+      "objective/train/tokens_used": 1006776800,
+      "objective/train/value_avg": -0.007381439208984375,
+      "objective/train/value_loss": 0.00016254279762506485,
+      "objective/train/value_max": -0.0001442432403564453,
+      "objective/train/value_min": -0.274169921875,
+      "objective/train/value_reward_corr": 0.5673249772676263,
+      "objective/train/value_std": 0.008941650390625,
+      "objective/train/weight_avg": 1.00246000289917,
+      "objective/train/weighted_lm_loss": 1.5009002685546875,
+      "objective/train/weights_max": 1.1678251028060913,
+      "objective/train/weights_min": 0.3721363842487335,
+      "theoretical_loss": 3.682451655135952,
+      "tokens_seen": 986316800
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000708152784464773,
+      "loss": 0.0731,
+      "theoretical_loss": 3.682405330214153,
+      "tokens_seen": 986447872
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007080725405231905,
+      "loss": 0.0717,
+      "theoretical_loss": 3.682312704002579,
+      "tokens_seen": 986710016
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007079922965816082,
+      "loss": 0.0762,
+      "theoretical_loss": 3.6822201092844686,
+      "tokens_seen": 986972160
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007079120526400257,
+      "loss": 0.0735,
+      "theoretical_loss": 3.682127546040753,
+      "tokens_seen": 987234304
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007078318086984433,
+      "loss": 0.0753,
+      "theoretical_loss": 3.6820350142523806,
+      "tokens_seen": 987496448
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007077515647568609,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6819425139003155,
+      "tokens_seen": 987758592
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007076713208152784,
+      "loss": 0.0763,
+      "theoretical_loss": 3.6818500449655396,
+      "tokens_seen": 988020736
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000707591076873696,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6817576074290503,
+      "tokens_seen": 988282880
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007075108329321136,
+      "loss": 0.0724,
+      "theoretical_loss": 3.681665201271862,
+      "tokens_seen": 988545024
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007074305889905313,
+      "loss": 0.0774,
+      "theoretical_loss": 3.681572826475006,
+      "tokens_seen": 988807168
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007073503450489488,
+      "loss": 0.0752,
+      "theoretical_loss": 3.681480483019529,
+      "tokens_seen": 989069312
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007072701011073665,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6813881708864953,
+      "tokens_seen": 989331456
+    },
+    {
+      "epoch": 0.3,
+      "objective/train/advantage_avg": 0.000814878090750426,
+      "objective/train/docs_used": 363941,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4229824542999268,
+      "objective/train/original_loss": 1.4229824542999268,
+      "objective/train/theoretical_loss": 3.681295890056985,
+      "objective/train/tokens_used": 1010053600,
+      "objective/train/value_avg": -0.007781982421875,
+      "objective/train/value_loss": 0.00012201992649352178,
+      "objective/train/value_max": -0.0001080632209777832,
+      "objective/train/value_min": -0.1939697265625,
+      "objective/train/value_reward_corr": 0.701736020129122,
+      "objective/train/value_std": 0.01049041748046875,
+      "objective/train/weight_avg": 1.000870704650879,
+      "objective/train/weighted_lm_loss": 1.4249759912490845,
+      "objective/train/weights_max": 1.2072614431381226,
+      "objective/train/weights_min": 0.3682592809200287,
+      "theoretical_loss": 3.681295890056985,
+      "tokens_seen": 989593600
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000707189857165784,
+      "loss": 0.0741,
+      "theoretical_loss": 3.681295890056985,
+      "tokens_seen": 989593600
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007071096132242016,
+      "loss": 0.0718,
+      "theoretical_loss": 3.681203640512095,
+      "tokens_seen": 989855744
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007070293692826192,
+      "loss": 0.076,
+      "theoretical_loss": 3.681111422232937,
+      "tokens_seen": 990117888
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007069491253410367,
+      "loss": 0.0765,
+      "theoretical_loss": 3.681019235200643,
+      "tokens_seen": 990380032
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007068688813994543,
+      "loss": 0.0747,
+      "theoretical_loss": 3.680927079396357,
+      "tokens_seen": 990642176
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007067886374578719,
+      "loss": 0.0736,
+      "theoretical_loss": 3.680834954801242,
+      "tokens_seen": 990904320
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007067083935162896,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6807428613964763,
+      "tokens_seen": 991166464
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007066281495747072,
+      "loss": 0.0778,
+      "theoretical_loss": 3.6806507991632555,
+      "tokens_seen": 991428608
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007065479056331248,
+      "loss": 0.0746,
+      "theoretical_loss": 3.68055876808279,
+      "tokens_seen": 991690752
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007064676616915423,
+      "loss": 0.0791,
+      "theoretical_loss": 3.680466768136308,
+      "tokens_seen": 991952896
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007063874177499599,
+      "loss": 0.0758,
+      "theoretical_loss": 3.680374799305053,
+      "tokens_seen": 992215040
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007063071738083775,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6802828615702845,
+      "tokens_seen": 992477184
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000706226929866795,
+      "loss": 0.0765,
+      "theoretical_loss": 3.6801909549132796,
+      "tokens_seen": 992739328
+    },
+    {
+      "epoch": 0.3,
+      "objective/train/advantage_avg": 0.001581284566782415,
+      "objective/train/docs_used": 365105,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4962949752807617,
+      "objective/train/original_loss": 1.4962949752807617,
+      "objective/train/theoretical_loss": 3.6801450132330915,
+      "objective/train/tokens_used": 1013330400,
+      "objective/train/value_avg": -0.00997161865234375,
+      "objective/train/value_loss": 0.00031275878427550197,
+      "objective/train/value_max": -7.486343383789062e-05,
+      "objective/train/value_min": -0.4453125,
+      "objective/train/value_reward_corr": 0.7353213483699139,
+      "objective/train/value_std": 0.020538330078125,
+      "objective/train/weight_avg": 1.0017253160476685,
+      "objective/train/weighted_lm_loss": 1.498146414756775,
+      "objective/train/weights_max": 1.244823932647705,
+      "objective/train/weights_min": 0.4185446798801422,
+      "theoretical_loss": 3.6801450132330915,
+      "tokens_seen": 992870400
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007061466859252127,
+      "loss": 0.0758,
+      "theoretical_loss": 3.6800990793153305,
+      "tokens_seen": 993001472
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007060664419836302,
+      "loss": 0.075,
+      "theoretical_loss": 3.6800072347577455,
+      "tokens_seen": 993263616
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007059861980420478,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6799154212218506,
+      "tokens_seen": 993525760
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007059059541004655,
+      "loss": 0.0726,
+      "theoretical_loss": 3.679823638688985,
+      "tokens_seen": 993787904
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000705825710158883,
+      "loss": 0.0742,
+      "theoretical_loss": 3.679731887140508,
+      "tokens_seen": 994050048
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007057454662173006,
+      "loss": 0.0712,
+      "theoretical_loss": 3.6796401665577916,
+      "tokens_seen": 994312192
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007056652222757182,
+      "loss": 0.0768,
+      "theoretical_loss": 3.679548476922225,
+      "tokens_seen": 994574336
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007055849783341358,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6794568182152143,
+      "tokens_seen": 994836480
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007055047343925533,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6793651904181806,
+      "tokens_seen": 995098624
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000705424490450971,
+      "loss": 0.0749,
+      "theoretical_loss": 3.679273593512563,
+      "tokens_seen": 995360768
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007053442465093885,
+      "loss": 0.0767,
+      "theoretical_loss": 3.679182027479812,
+      "tokens_seen": 995622912
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007052640025678061,
+      "loss": 0.0765,
+      "theoretical_loss": 3.6790904923014005,
+      "tokens_seen": 995885056
+    },
+    {
+      "epoch": 0.3,
+      "objective/train/advantage_avg": 0.0009565390646457672,
+      "objective/train/docs_used": 366329,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5922820568084717,
+      "objective/train/original_loss": 1.5922820568084717,
+      "objective/train/theoretical_loss": 3.6789989879588125,
+      "objective/train/tokens_used": 1016607200,
+      "objective/train/value_avg": -0.010162353515625,
+      "objective/train/value_loss": 0.0004036914324387908,
+      "objective/train/value_max": -8.028745651245117e-05,
+      "objective/train/value_min": -0.489990234375,
+      "objective/train/value_reward_corr": 0.6579596576861813,
+      "objective/train/value_std": 0.017059326171875,
+      "objective/train/weight_avg": 1.0011204481124878,
+      "objective/train/weighted_lm_loss": 1.594104528427124,
+      "objective/train/weights_max": 1.1680172681808472,
+      "objective/train/weights_min": 0.2262655794620514,
+      "theoretical_loss": 3.6789989879588125,
+      "tokens_seen": 996147200
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007051837586262238,
+      "loss": 0.074,
+      "theoretical_loss": 3.6789989879588125,
+      "tokens_seen": 996147200
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007051035146846413,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6789075144335497,
+      "tokens_seen": 996409344
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000705023270743059,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6788160717071303,
+      "tokens_seen": 996671488
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007049430268014765,
+      "loss": 0.0771,
+      "theoretical_loss": 3.678724659761087,
+      "tokens_seen": 996933632
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007048627828598941,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6786332785769695,
+      "tokens_seen": 997195776
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007047825389183117,
+      "loss": 0.0736,
+      "theoretical_loss": 3.678541928136344,
+      "tokens_seen": 997457920
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007047022949767292,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6784506084207904,
+      "tokens_seen": 997720064
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007046220510351468,
+      "loss": 0.0779,
+      "theoretical_loss": 3.6783593194119066,
+      "tokens_seen": 997982208
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007045418070935644,
+      "loss": 0.0761,
+      "theoretical_loss": 3.6782680610913054,
+      "tokens_seen": 998244352
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007044615631519821,
+      "loss": 0.079,
+      "theoretical_loss": 3.6781768334406157,
+      "tokens_seen": 998506496
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007043813192103996,
+      "loss": 0.0736,
+      "theoretical_loss": 3.678085636441482,
+      "tokens_seen": 998768640
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007043010752688173,
+      "loss": 0.0783,
+      "theoretical_loss": 3.677994470075565,
+      "tokens_seen": 999030784
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007042208313272348,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6779033343245406,
+      "tokens_seen": 999292928
+    },
+    {
+      "epoch": 0.3,
+      "objective/train/advantage_avg": 0.0017308670794591308,
+      "objective/train/docs_used": 367579,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6493194103240967,
+      "objective/train/original_loss": 1.6493196487426758,
+      "objective/train/theoretical_loss": 3.677857777923891,
+      "objective/train/tokens_used": 1019884000,
+      "objective/train/value_avg": -0.006999969482421875,
+      "objective/train/value_loss": 0.00011768531840061769,
+      "objective/train/value_max": -0.00017130374908447266,
+      "objective/train/value_min": -0.298828125,
+      "objective/train/value_reward_corr": 0.617122584170849,
+      "objective/train/value_std": 0.00861358642578125,
+      "objective/train/weight_avg": 1.0017890930175781,
+      "objective/train/weighted_lm_loss": 1.6527791023254395,
+      "objective/train/weights_max": 1.1782879829406738,
+      "objective/train/weights_min": 0.7371587753295898,
+      "theoretical_loss": 3.677857777923891,
+      "tokens_seen": 999424000
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007041405873856524,
+      "loss": 0.0756,
+      "theoretical_loss": 3.677812229170101,
+      "tokens_seen": 999555072
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.00070406034344407,
+      "loss": 0.0732,
+      "theoretical_loss": 3.677721154593953,
+      "tokens_seen": 999817216
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007039800995024875,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6776301105778213,
+      "tokens_seen": 1000079360
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007038998555609052,
+      "loss": 0.076,
+      "theoretical_loss": 3.6775390971034447,
+      "tokens_seen": 1000341504
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007038196116193227,
+      "loss": 0.0742,
+      "theoretical_loss": 3.6774481141525777,
+      "tokens_seen": 1000603648
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007037393676777404,
+      "loss": 0.075,
+      "theoretical_loss": 3.6773571617069907,
+      "tokens_seen": 1000865792
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000703659123736158,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6772662397484703,
+      "tokens_seen": 1001127936
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007035788797945756,
+      "loss": 0.0704,
+      "theoretical_loss": 3.6771753482588183,
+      "tokens_seen": 1001390080
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007034986358529931,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6770844872198523,
+      "tokens_seen": 1001652224
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007034183919114107,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6769936566134045,
+      "tokens_seen": 1001914368
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007033381479698283,
+      "loss": 0.0755,
+      "theoretical_loss": 3.676902856421324,
+      "tokens_seen": 1002176512
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007032579040282458,
+      "loss": 0.076,
+      "theoretical_loss": 3.6768120866254757,
+      "tokens_seen": 1002438656
+    },
+    {
+      "epoch": 0.3,
+      "objective/train/advantage_avg": -0.0004565780109260231,
+      "objective/train/docs_used": 368716,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.602346420288086,
+      "objective/train/original_loss": 1.6023463010787964,
+      "objective/train/theoretical_loss": 3.6767213472077387,
+      "objective/train/tokens_used": 1023160800,
+      "objective/train/value_avg": -0.0089874267578125,
+      "objective/train/value_loss": 0.0004929814604111016,
+      "objective/train/value_max": -9.459257125854492e-05,
+      "objective/train/value_min": -0.70947265625,
+      "objective/train/value_reward_corr": 0.6152421290672049,
+      "objective/train/value_std": 0.01763916015625,
+      "objective/train/weight_avg": 0.9997656345367432,
+      "objective/train/weighted_lm_loss": 1.6013120412826538,
+      "objective/train/weights_max": 1.700226902961731,
+      "objective/train/weights_min": 0.3712657690048218,
+      "theoretical_loss": 3.6767213472077387,
+      "tokens_seen": 1002700800
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007031776600866635,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6767213472077387,
+      "tokens_seen": 1002700800
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000703097416145081,
+      "loss": 0.073,
+      "theoretical_loss": 3.676630638150008,
+      "tokens_seen": 1002962944
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007030171722034986,
+      "loss": 0.0722,
+      "theoretical_loss": 3.6765399594341943,
+      "tokens_seen": 1003225088
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007029369282619163,
+      "loss": 0.0758,
+      "theoretical_loss": 3.676449311042225,
+      "tokens_seen": 1003487232
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007028566843203338,
+      "loss": 0.0789,
+      "theoretical_loss": 3.6763586929560415,
+      "tokens_seen": 1003749376
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007027764403787515,
+      "loss": 0.0758,
+      "theoretical_loss": 3.6762681051576003,
+      "tokens_seen": 1004011520
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.000702696196437169,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6761775476288747,
+      "tokens_seen": 1004273664
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007026159524955866,
+      "loss": 0.075,
+      "theoretical_loss": 3.6760870203518525,
+      "tokens_seen": 1004535808
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007025357085540042,
+      "loss": 0.0715,
+      "theoretical_loss": 3.6759965233085383,
+      "tokens_seen": 1004797952
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007024554646124218,
+      "loss": 0.0748,
+      "theoretical_loss": 3.6759060564809496,
+      "tokens_seen": 1005060096
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007023752206708393,
+      "loss": 0.0748,
+      "theoretical_loss": 3.6758156198511216,
+      "tokens_seen": 1005322240
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007022949767292569,
+      "loss": 0.076,
+      "theoretical_loss": 3.675725213401104,
+      "tokens_seen": 1005584384
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007022147327876746,
+      "loss": 0.0757,
+      "theoretical_loss": 3.6756348371129617,
+      "tokens_seen": 1005846528
+    },
+    {
+      "epoch": 0.3,
+      "objective/train/advantage_avg": 0.0010321858571842313,
+      "objective/train/docs_used": 369909,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.614028811454773,
+      "objective/train/original_loss": 1.6140289306640625,
+      "objective/train/theoretical_loss": 3.6755896602739933,
+      "objective/train/tokens_used": 1026437600,
+      "objective/train/value_avg": -0.00745391845703125,
+      "objective/train/value_loss": 0.00014535474474541843,
+      "objective/train/value_max": -8.153915405273438e-05,
+      "objective/train/value_min": -0.311767578125,
+      "objective/train/value_reward_corr": 0.76924190262928,
+      "objective/train/value_std": 0.01385498046875,
+      "objective/train/weight_avg": 1.001103401184082,
+      "objective/train/weighted_lm_loss": 1.6157536506652832,
+      "objective/train/weights_max": 1.166409969329834,
+      "objective/train/weights_min": 0.6640326976776123,
+      "theoretical_loss": 3.6755896602739933,
+      "tokens_seen": 1005977600
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007021344888460921,
+      "loss": 0.0776,
+      "theoretical_loss": 3.6755444909687744,
+      "tokens_seen": 1006108672
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 0.0007020542449045098,
+      "loss": 0.0745,
+      "theoretical_loss": 3.675454174950639,
+      "tokens_seen": 1006370816
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007019740009629273,
+      "loss": 0.0734,
+      "theoretical_loss": 3.675363889040666,
+      "tokens_seen": 1006632960
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007018937570213449,
+      "loss": 0.0748,
+      "theoretical_loss": 3.675273633220981,
+      "tokens_seen": 1006895104
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007018135130797625,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6751834074737264,
+      "tokens_seen": 1007157248
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.00070173326913818,
+      "loss": 0.0762,
+      "theoretical_loss": 3.675093211781059,
+      "tokens_seen": 1007419392
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007016530251965977,
+      "loss": 0.0773,
+      "theoretical_loss": 3.67500304612515,
+      "tokens_seen": 1007681536
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007015727812550152,
+      "loss": 0.0756,
+      "theoretical_loss": 3.674912910488187,
+      "tokens_seen": 1007943680
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007014925373134329,
+      "loss": 0.0759,
+      "theoretical_loss": 3.6748228048523726,
+      "tokens_seen": 1008205824
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007014122933718505,
+      "loss": 0.0783,
+      "theoretical_loss": 3.674732729199924,
+      "tokens_seen": 1008467968
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007013320494302681,
+      "loss": 0.0742,
+      "theoretical_loss": 3.674642683513074,
+      "tokens_seen": 1008730112
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007012518054886856,
+      "loss": 0.0765,
+      "theoretical_loss": 3.674552667774071,
+      "tokens_seen": 1008992256
+    },
+    {
+      "epoch": 0.31,
+      "objective/train/advantage_avg": 0.0011902566766366363,
+      "objective/train/docs_used": 370943,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5907751321792603,
+      "objective/train/original_loss": 1.5907747745513916,
+      "objective/train/theoretical_loss": 3.6744626819651773,
+      "objective/train/tokens_used": 1029714400,
+      "objective/train/value_avg": -0.00921630859375,
+      "objective/train/value_loss": 0.0002929775801021606,
+      "objective/train/value_max": -0.00011324882507324219,
+      "objective/train/value_min": -0.263427734375,
+      "objective/train/value_reward_corr": 0.6733462873979555,
+      "objective/train/value_std": 0.0144805908203125,
+      "objective/train/weight_avg": 1.0013235807418823,
+      "objective/train/weighted_lm_loss": 1.5915753841400146,
+      "objective/train/weights_max": 1.3013832569122314,
+      "objective/train/weights_min": 0.39872050285339355,
+      "theoretical_loss": 3.6744626819651773,
+      "tokens_seen": 1009254400
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007011715615471032,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6744626819651773,
+      "tokens_seen": 1009254400
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007010913176055208,
+      "loss": 0.0748,
+      "theoretical_loss": 3.674372726068671,
+      "tokens_seen": 1009516544
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007010110736639383,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6742828000668464,
+      "tokens_seen": 1009778688
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.000700930829722356,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6741929039420103,
+      "tokens_seen": 1010040832
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007008505857807735,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6741030376764865,
+      "tokens_seen": 1010302976
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007007703418391912,
+      "loss": 0.0764,
+      "theoretical_loss": 3.674013201252614,
+      "tokens_seen": 1010565120
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007006900978976088,
+      "loss": 0.0751,
+      "theoretical_loss": 3.6739233946527454,
+      "tokens_seen": 1010827264
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007006098539560264,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6738336178592492,
+      "tokens_seen": 1011089408
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007005296100144439,
+      "loss": 0.0791,
+      "theoretical_loss": 3.6737438708545094,
+      "tokens_seen": 1011351552
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007004493660728615,
+      "loss": 0.0751,
+      "theoretical_loss": 3.673654153620924,
+      "tokens_seen": 1011613696
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007003691221312791,
+      "loss": 0.0784,
+      "theoretical_loss": 3.673564466140906,
+      "tokens_seen": 1011875840
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007002888781896967,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6734748083968842,
+      "tokens_seen": 1012137984
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007002086342481143,
+      "loss": 0.0759,
+      "theoretical_loss": 3.6733851803713016,
+      "tokens_seen": 1012400128
+    },
+    {
+      "epoch": 0.31,
+      "objective/train/advantage_avg": -0.00044222682481631637,
+      "objective/train/docs_used": 372166,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5056043863296509,
+      "objective/train/original_loss": 1.5056045055389404,
+      "objective/train/theoretical_loss": 3.6733403774974427,
+      "objective/train/tokens_used": 1032991200,
+      "objective/train/value_avg": -0.01678466796875,
+      "objective/train/value_loss": 0.00021295166516210884,
+      "objective/train/value_max": -0.00011146068572998047,
+      "objective/train/value_min": -0.6337890625,
+      "objective/train/value_reward_corr": 0.9449292467966458,
+      "objective/train/value_std": 0.041168212890625,
+      "objective/train/weight_avg": 0.9996622800827026,
+      "objective/train/weighted_lm_loss": 1.5041799545288086,
+      "objective/train/weights_max": 1.2211940288543701,
+      "objective/train/weights_min": 0.6597907543182373,
+      "theoretical_loss": 3.6733403774974427,
+      "tokens_seen": 1012531200
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007001283903065318,
+      "loss": 0.076,
+      "theoretical_loss": 3.673295582046616,
+      "tokens_seen": 1012662272
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0007000481463649496,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6732060134053013,
+      "tokens_seen": 1012924416
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006999679024233671,
+      "loss": 0.0761,
+      "theoretical_loss": 3.673116474429844,
+      "tokens_seen": 1013186560
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006998876584817846,
+      "loss": 0.0775,
+      "theoretical_loss": 3.673026965102748,
+      "tokens_seen": 1013448704
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006998074145402023,
+      "loss": 0.0739,
+      "theoretical_loss": 3.67293748540653,
+      "tokens_seen": 1013710848
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006997271705986198,
+      "loss": 0.0738,
+      "theoretical_loss": 3.672848035323723,
+      "tokens_seen": 1013972992
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006996469266570374,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6727586148368743,
+      "tokens_seen": 1014235136
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.000699566682715455,
+      "loss": 0.0755,
+      "theoretical_loss": 3.672669223928545,
+      "tokens_seen": 1014497280
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006994864387738726,
+      "loss": 0.0745,
+      "theoretical_loss": 3.672579862581313,
+      "tokens_seen": 1014759424
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006994061948322901,
+      "loss": 0.0746,
+      "theoretical_loss": 3.672490530777769,
+      "tokens_seen": 1015021568
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006993259508907077,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6724012285005196,
+      "tokens_seen": 1015283712
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006992457069491254,
+      "loss": 0.0762,
+      "theoretical_loss": 3.6723119557321864,
+      "tokens_seen": 1015545856
+    },
+    {
+      "epoch": 0.31,
+      "objective/train/advantage_avg": -0.000864784000441432,
+      "objective/train/docs_used": 372984,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4644211530685425,
+      "objective/train/original_loss": 1.4644211530685425,
+      "objective/train/theoretical_loss": 3.6722227124554045,
+      "objective/train/tokens_used": 1036268000,
+      "objective/train/value_avg": -0.00815582275390625,
+      "objective/train/value_loss": 0.0006307647563517094,
+      "objective/train/value_max": -0.00012433528900146484,
+      "objective/train/value_min": -0.7392578125,
+      "objective/train/value_reward_corr": 0.6358148568839063,
+      "objective/train/value_std": 0.01560211181640625,
+      "objective/train/weight_avg": 0.999402642250061,
+      "objective/train/weighted_lm_loss": 1.4642397165298462,
+      "objective/train/weights_max": 2.048352003097534,
+      "objective/train/weights_min": 0.39648786187171936,
+      "theoretical_loss": 3.6722227124554045,
+      "tokens_seen": 1015808000
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006991654630075429,
+      "loss": 0.0757,
+      "theoretical_loss": 3.6722227124554045,
+      "tokens_seen": 1015808000
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006990852190659606,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6721334986528236,
+      "tokens_seen": 1016070144
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006990049751243781,
+      "loss": 0.0767,
+      "theoretical_loss": 3.6720443143071106,
+      "tokens_seen": 1016332288
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006989247311827958,
+      "loss": 0.0744,
+      "theoretical_loss": 3.671955159400943,
+      "tokens_seen": 1016594432
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006988444872412133,
+      "loss": 0.0764,
+      "theoretical_loss": 3.6718660339170173,
+      "tokens_seen": 1016856576
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006987642432996308,
+      "loss": 0.076,
+      "theoretical_loss": 3.6717769378380414,
+      "tokens_seen": 1017118720
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006986839993580485,
+      "loss": 0.075,
+      "theoretical_loss": 3.671687871146739,
+      "tokens_seen": 1017380864
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.000698603755416466,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6715988338258487,
+      "tokens_seen": 1017643008
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006985235114748837,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6715098258581236,
+      "tokens_seen": 1017905152
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006984432675333013,
+      "loss": 0.0762,
+      "theoretical_loss": 3.6714208472263303,
+      "tokens_seen": 1018167296
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006983630235917189,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6713318979132517,
+      "tokens_seen": 1018429440
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006982827796501364,
+      "loss": 0.0755,
+      "theoretical_loss": 3.671242977901683,
+      "tokens_seen": 1018691584
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.000698202535708554,
+      "loss": 0.0761,
+      "theoretical_loss": 3.671154087174436,
+      "tokens_seen": 1018953728
+    },
+    {
+      "epoch": 0.31,
+      "objective/train/advantage_avg": 0.0015296280616894364,
+      "objective/train/docs_used": 374190,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4388104677200317,
+      "objective/train/original_loss": 1.4388103485107422,
+      "objective/train/theoretical_loss": 3.6711096527870657,
+      "objective/train/tokens_used": 1039544800,
+      "objective/train/value_avg": -0.005893707275390625,
+      "objective/train/value_loss": 0.00029690880910493433,
+      "objective/train/value_max": -4.297494888305664e-05,
+      "objective/train/value_min": -0.744140625,
+      "objective/train/value_reward_corr": 0.5134253858490913,
+      "objective/train/value_std": 0.0106964111328125,
+      "objective/train/weight_avg": 1.0016546249389648,
+      "objective/train/weighted_lm_loss": 1.4419684410095215,
+      "objective/train/weights_max": 1.183031678199768,
+      "objective/train/weights_min": 0.3685206472873688,
+      "theoretical_loss": 3.6711096527870657,
+      "tokens_seen": 1019084800
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006981222917669716,
+      "loss": 0.0761,
+      "theoretical_loss": 3.6710652257143366,
+      "tokens_seen": 1019215872
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006980420478253891,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6709763935042243,
+      "tokens_seen": 1019478016
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006979618038838068,
+      "loss": 0.0759,
+      "theoretical_loss": 3.670887590526953,
+      "tokens_seen": 1019740160
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006978815599422243,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6707988167653927,
+      "tokens_seen": 1020002304
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006978013160006421,
+      "loss": 0.075,
+      "theoretical_loss": 3.670710072202426,
+      "tokens_seen": 1020264448
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006977210720590596,
+      "loss": 0.0763,
+      "theoretical_loss": 3.670621356820951,
+      "tokens_seen": 1020526592
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006976408281174772,
+      "loss": 0.0737,
+      "theoretical_loss": 3.67053267060388,
+      "tokens_seen": 1020788736
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006975605841758948,
+      "loss": 0.0761,
+      "theoretical_loss": 3.6704440135341394,
+      "tokens_seen": 1021050880
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006974803402343123,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6703553855946702,
+      "tokens_seen": 1021313024
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006974000962927299,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6702667867684275,
+      "tokens_seen": 1021575168
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006973198523511475,
+      "loss": 0.0727,
+      "theoretical_loss": 3.670178217038381,
+      "tokens_seen": 1021837312
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006972396084095651,
+      "loss": 0.0739,
+      "theoretical_loss": 3.670089676387515,
+      "tokens_seen": 1022099456
+    },
+    {
+      "epoch": 0.31,
+      "objective/train/advantage_avg": 0.0019567436538636684,
+      "objective/train/docs_used": 375283,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4512615203857422,
+      "objective/train/original_loss": 1.451261281967163,
+      "objective/train/theoretical_loss": 3.6700011647988275,
+      "objective/train/tokens_used": 1042821600,
+      "objective/train/value_avg": -0.01641845703125,
+      "objective/train/value_loss": 0.00019754536333493888,
+      "objective/train/value_max": -9.5367431640625e-05,
+      "objective/train/value_min": -0.307861328125,
+      "objective/train/value_reward_corr": 0.9487556089623391,
+      "objective/train/value_std": 0.040863037109375,
+      "objective/train/weight_avg": 1.0020544528961182,
+      "objective/train/weighted_lm_loss": 1.4532920122146606,
+      "objective/train/weights_max": 1.1703647375106812,
+      "objective/train/weights_min": 0.7569119334220886,
+      "theoretical_loss": 3.6700011647988275,
+      "tokens_seen": 1022361600
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006971593644679826,
+      "loss": 0.0772,
+      "theoretical_loss": 3.6700011647988275,
+      "tokens_seen": 1022361600
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006970791205264004,
+      "loss": 0.0726,
+      "theoretical_loss": 3.6699126822553314,
+      "tokens_seen": 1022623744
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006969988765848179,
+      "loss": 0.0735,
+      "theoretical_loss": 3.669824228740053,
+      "tokens_seen": 1022885888
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006969186326432354,
+      "loss": 0.0772,
+      "theoretical_loss": 3.6697358042360344,
+      "tokens_seen": 1023148032
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006968383887016531,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6696474087263296,
+      "tokens_seen": 1023410176
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006967581447600706,
+      "loss": 0.0774,
+      "theoretical_loss": 3.6695590421940096,
+      "tokens_seen": 1023672320
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006966779008184882,
+      "loss": 0.0762,
+      "theoretical_loss": 3.6694707046221575,
+      "tokens_seen": 1023934464
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006965976568769058,
+      "loss": 0.0742,
+      "theoretical_loss": 3.669382395993871,
+      "tokens_seen": 1024196608
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006965174129353234,
+      "loss": 0.0754,
+      "theoretical_loss": 3.669294116292263,
+      "tokens_seen": 1024458752
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.000696437168993741,
+      "loss": 0.075,
+      "theoretical_loss": 3.6692058655004605,
+      "tokens_seen": 1024720896
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006963569250521585,
+      "loss": 0.0766,
+      "theoretical_loss": 3.669117643601602,
+      "tokens_seen": 1024983040
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006962766811105762,
+      "loss": 0.0718,
+      "theoretical_loss": 3.6690294505788446,
+      "tokens_seen": 1025245184
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006961964371689938,
+      "loss": 0.0738,
+      "theoretical_loss": 3.668941286415355,
+      "tokens_seen": 1025507328
+    },
+    {
+      "epoch": 0.31,
+      "objective/train/advantage_avg": 0.001169724389910698,
+      "objective/train/docs_used": 376516,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2908331155776978,
+      "objective/train/original_loss": 1.2908332347869873,
+      "objective/train/theoretical_loss": 3.6688972151505803,
+      "objective/train/tokens_used": 1046098400,
+      "objective/train/value_avg": -0.006412506103515625,
+      "objective/train/value_loss": 0.00014659865701105446,
+      "objective/train/value_max": -6.204843521118164e-05,
+      "objective/train/value_min": -0.332275390625,
+      "objective/train/value_reward_corr": 0.7111525298676997,
+      "objective/train/value_std": 0.012237548828125,
+      "objective/train/weight_avg": 1.0012383460998535,
+      "objective/train/weighted_lm_loss": 1.2924283742904663,
+      "objective/train/weights_max": 1.180111289024353,
+      "objective/train/weights_min": 0.3988421857357025,
+      "theoretical_loss": 3.6688972151505803,
+      "tokens_seen": 1025638400
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006961161932274114,
+      "loss": 0.0731,
+      "theoretical_loss": 3.668853151094318,
+      "tokens_seen": 1025769472
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006960359492858289,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6687650445989295,
+      "tokens_seen": 1026031616
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006959557053442466,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6686769669124004,
+      "tokens_seen": 1026293760
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006958754614026641,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6685889180179565,
+      "tokens_seen": 1026555904
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006957952174610816,
+      "loss": 0.0741,
+      "theoretical_loss": 3.668500897898837,
+      "tokens_seen": 1026818048
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006957149735194993,
+      "loss": 0.0754,
+      "theoretical_loss": 3.668412906538295,
+      "tokens_seen": 1027080192
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006956347295779168,
+      "loss": 0.0759,
+      "theoretical_loss": 3.6683249439195977,
+      "tokens_seen": 1027342336
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006955544856363345,
+      "loss": 0.0751,
+      "theoretical_loss": 3.668237010026026,
+      "tokens_seen": 1027604480
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006954742416947521,
+      "loss": 0.0754,
+      "theoretical_loss": 3.668149104840876,
+      "tokens_seen": 1027866624
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006953939977531697,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6680612283474567,
+      "tokens_seen": 1028128768
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006953137538115872,
+      "loss": 0.0743,
+      "theoretical_loss": 3.667973380529091,
+      "tokens_seen": 1028390912
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006952335098700048,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6678855613691157,
+      "tokens_seen": 1028653056
+    },
+    {
+      "epoch": 0.31,
+      "objective/train/advantage_avg": 0.0010213935747742653,
+      "objective/train/docs_used": 377682,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4129959344863892,
+      "objective/train/original_loss": 1.4129958152770996,
+      "objective/train/theoretical_loss": 3.667797770850883,
+      "objective/train/tokens_used": 1049375200,
+      "objective/train/value_avg": -0.007488250732421875,
+      "objective/train/value_loss": 0.00019148027058690786,
+      "objective/train/value_max": -0.0001055598258972168,
+      "objective/train/value_min": -0.326904296875,
+      "objective/train/value_reward_corr": 0.5655283244055973,
+      "objective/train/value_std": 0.01061248779296875,
+      "objective/train/weight_avg": 1.0011082887649536,
+      "objective/train/weighted_lm_loss": 1.4145474433898926,
+      "objective/train/weights_max": 1.0799387693405151,
+      "objective/train/weights_min": 0.3737297058105469,
+      "theoretical_loss": 3.667797770850883,
+      "tokens_seen": 1028915200
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006951532659284224,
+      "loss": 0.0734,
+      "theoretical_loss": 3.667797770850883,
+      "tokens_seen": 1028915200
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.00069507302198684,
+      "loss": 0.0748,
+      "theoretical_loss": 3.667710008957756,
+      "tokens_seen": 1029177344
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006949927780452576,
+      "loss": 0.0776,
+      "theoretical_loss": 3.667622275673115,
+      "tokens_seen": 1029439488
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006949125341036751,
+      "loss": 0.0747,
+      "theoretical_loss": 3.667534570980353,
+      "tokens_seen": 1029701632
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006948322901620929,
+      "loss": 0.076,
+      "theoretical_loss": 3.667446894862876,
+      "tokens_seen": 1029963776
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006947520462205104,
+      "loss": 0.0731,
+      "theoretical_loss": 3.667359247304104,
+      "tokens_seen": 1030225920
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006946718022789279,
+      "loss": 0.074,
+      "theoretical_loss": 3.667271628287472,
+      "tokens_seen": 1030488064
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006945915583373456,
+      "loss": 0.0754,
+      "theoretical_loss": 3.6671840377964275,
+      "tokens_seen": 1030750208
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006945113143957631,
+      "loss": 0.0756,
+      "theoretical_loss": 3.667096475814433,
+      "tokens_seen": 1031012352
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006944310704541807,
+      "loss": 0.0751,
+      "theoretical_loss": 3.6670089423249643,
+      "tokens_seen": 1031274496
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006943508265125983,
+      "loss": 0.0785,
+      "theoretical_loss": 3.6669214373115104,
+      "tokens_seen": 1031536640
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006942705825710159,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6668339607575744,
+      "tokens_seen": 1031798784
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006941903386294334,
+      "loss": 0.0726,
+      "theoretical_loss": 3.6667465126466743,
+      "tokens_seen": 1032060928
+    },
+    {
+      "epoch": 0.31,
+      "objective/train/advantage_avg": 0.0018603794742375612,
+      "objective/train/docs_used": 378946,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4675285816192627,
+      "objective/train/original_loss": 1.4675285816192627,
+      "objective/train/theoretical_loss": 3.6667027992522145,
+      "objective/train/tokens_used": 1052652000,
+      "objective/train/value_avg": -0.00775909423828125,
+      "objective/train/value_loss": 0.00012807230814360082,
+      "objective/train/value_max": -0.0001150369644165039,
+      "objective/train/value_min": -0.2120361328125,
+      "objective/train/value_reward_corr": 0.680010560022861,
+      "objective/train/value_std": 0.01026153564453125,
+      "objective/train/weight_avg": 1.0019227266311646,
+      "objective/train/weighted_lm_loss": 1.4709603786468506,
+      "objective/train/weights_max": 1.1424355506896973,
+      "objective/train/weights_min": 0.5569349527359009,
+      "theoretical_loss": 3.6667027992522145,
+      "tokens_seen": 1032192000
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006941100946878512,
+      "loss": 0.0707,
+      "theoretical_loss": 3.6666590929623393,
+      "tokens_seen": 1032323072
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006940298507462687,
+      "loss": 0.0717,
+      "theoretical_loss": 3.666571701688115,
+      "tokens_seen": 1032585216
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006939496068046863,
+      "loss": 0.0769,
+      "theoretical_loss": 3.6664843388075594,
+      "tokens_seen": 1032847360
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006938693628631039,
+      "loss": 0.0707,
+      "theoretical_loss": 3.6663970043042435,
+      "tokens_seen": 1033109504
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006937891189215214,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6663096981617533,
+      "tokens_seen": 1033371648
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006937088749799391,
+      "loss": 0.0717,
+      "theoretical_loss": 3.6662224203636886,
+      "tokens_seen": 1033633792
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006936286310383566,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6661351708936616,
+      "tokens_seen": 1033895936
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006935483870967742,
+      "loss": 0.0694,
+      "theoretical_loss": 3.6660479497352982,
+      "tokens_seen": 1034158080
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006934681431551918,
+      "loss": 0.0754,
+      "theoretical_loss": 3.665960756872239,
+      "tokens_seen": 1034420224
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006933878992136093,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6658735922881376,
+      "tokens_seen": 1034682368
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.000693307655272027,
+      "loss": 0.0745,
+      "theoretical_loss": 3.665786455966661,
+      "tokens_seen": 1034944512
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006932274113304446,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6656993478914903,
+      "tokens_seen": 1035206656
+    },
+    {
+      "epoch": 0.31,
+      "objective/train/advantage_avg": 0.00034525172668509185,
+      "objective/train/docs_used": 379908,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.507829189300537,
+      "objective/train/original_loss": 1.507829189300537,
+      "objective/train/theoretical_loss": 3.6656122680463197,
+      "objective/train/tokens_used": 1055928800,
+      "objective/train/value_avg": -0.00595855712890625,
+      "objective/train/value_loss": 9.688719001132995e-05,
+      "objective/train/value_max": -0.00011682510375976562,
+      "objective/train/value_min": -0.22119140625,
+      "objective/train/value_reward_corr": 0.7160243405342503,
+      "objective/train/value_std": 0.00960540771484375,
+      "objective/train/weight_avg": 1.000393033027649,
+      "objective/train/weighted_lm_loss": 1.5080686807632446,
+      "objective/train/weights_max": 1.0766479969024658,
+      "objective/train/weights_min": 0.820656955242157,
+      "theoretical_loss": 3.6656122680463197,
+      "tokens_seen": 1035468800
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006931471673888622,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6656122680463197,
+      "tokens_seen": 1035468800
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006930669234472797,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6655252164148564,
+      "tokens_seen": 1035730944
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006929866795056974,
+      "loss": 0.0761,
+      "theoretical_loss": 3.6654381929808233,
+      "tokens_seen": 1035993088
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006929064355641149,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6653511977279534,
+      "tokens_seen": 1036255232
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006928261916225324,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6652642306399965,
+      "tokens_seen": 1036517376
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006927459476809501,
+      "loss": 0.075,
+      "theoretical_loss": 3.6651772917007137,
+      "tokens_seen": 1036779520
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006926657037393676,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6650903808938806,
+      "tokens_seen": 1037041664
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006925854597977854,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6650034982032857,
+      "tokens_seen": 1037303808
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006925052158562029,
+      "loss": 0.0735,
+      "theoretical_loss": 3.664916643612732,
+      "tokens_seen": 1037565952
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006924249719146205,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6648298171060345,
+      "tokens_seen": 1037828096
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006923447279730381,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6647430186670222,
+      "tokens_seen": 1038090240
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006922644840314556,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6646562482795373,
+      "tokens_seen": 1038352384
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006921842400898732,
+      "loss": 0.0722,
+      "theoretical_loss": 3.664569505927436,
+      "tokens_seen": 1038614528
+    },
+    {
+      "epoch": 0.31,
+      "objective/train/advantage_avg": -0.00035626679891720414,
+      "objective/train/docs_used": 381080,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3043032884597778,
+      "objective/train/original_loss": 1.3043031692504883,
+      "objective/train/theoretical_loss": 3.6645261452596136,
+      "objective/train/tokens_used": 1059205600,
+      "objective/train/value_avg": -0.00853729248046875,
+      "objective/train/value_loss": 0.00025545063544996083,
+      "objective/train/value_max": -7.724761962890625e-05,
+      "objective/train/value_min": -0.296142578125,
+      "objective/train/value_reward_corr": 0.7234297240140802,
+      "objective/train/value_std": 0.01416015625,
+      "objective/train/weight_avg": 0.9997609257698059,
+      "objective/train/weighted_lm_loss": 1.3044672012329102,
+      "objective/train/weights_max": 1.3083919286727905,
+      "objective/train/weights_min": 0.3739407956600189,
+      "theoretical_loss": 3.6645261452596136,
+      "tokens_seen": 1038745600
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006921039961482908,
+      "loss": 0.0737,
+      "theoretical_loss": 3.664482791594588,
+      "tokens_seen": 1038876672
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006920237522067084,
+      "loss": 0.0738,
+      "theoretical_loss": 3.664396105264875,
+      "tokens_seen": 1039138816
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 0.0006919435082651259,
+      "loss": 0.0736,
+      "theoretical_loss": 3.6643094469221933,
+      "tokens_seen": 1039400960
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006918632643235437,
+      "loss": 0.0734,
+      "theoretical_loss": 3.664222816550452,
+      "tokens_seen": 1039663104
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006917830203819612,
+      "loss": 0.0722,
+      "theoretical_loss": 3.6641362141335727,
+      "tokens_seen": 1039925248
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006917027764403787,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6640496396554925,
+      "tokens_seen": 1040187392
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006916225324987964,
+      "loss": 0.0753,
+      "theoretical_loss": 3.6639630931001594,
+      "tokens_seen": 1040449536
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006915422885572139,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6638765744515367,
+      "tokens_seen": 1040711680
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006914620446156316,
+      "loss": 0.0761,
+      "theoretical_loss": 3.6637900836935993,
+      "tokens_seen": 1040973824
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006913818006740491,
+      "loss": 0.0736,
+      "theoretical_loss": 3.6637036208103364,
+      "tokens_seen": 1041235968
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006913015567324667,
+      "loss": 0.0712,
+      "theoretical_loss": 3.663617185785749,
+      "tokens_seen": 1041498112
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006912213127908843,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6635307786038536,
+      "tokens_seen": 1041760256
+    },
+    {
+      "epoch": 0.32,
+      "objective/train/advantage_avg": 0.0007461439236067235,
+      "objective/train/docs_used": 382133,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4432910680770874,
+      "objective/train/original_loss": 1.4432913064956665,
+      "objective/train/theoretical_loss": 3.663444399248678,
+      "objective/train/tokens_used": 1062482400,
+      "objective/train/value_avg": -0.006084442138671875,
+      "objective/train/value_loss": 0.00022422554320655763,
+      "objective/train/value_max": -8.219480514526367e-05,
+      "objective/train/value_min": -0.318603515625,
+      "objective/train/value_reward_corr": 0.5543261672765586,
+      "objective/train/value_std": 0.00926971435546875,
+      "objective/train/weight_avg": 1.0008444786071777,
+      "objective/train/weighted_lm_loss": 1.443787693977356,
+      "objective/train/weights_max": 1.2419401407241821,
+      "objective/train/weights_min": 0.3867599070072174,
+      "theoretical_loss": 3.663444399248678,
+      "tokens_seen": 1042022400
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000691141068849302,
+      "loss": 0.072,
+      "theoretical_loss": 3.663444399248678,
+      "tokens_seen": 1042022400
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006910608249077195,
+      "loss": 0.0754,
+      "theoretical_loss": 3.6633580477042633,
+      "tokens_seen": 1042284544
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006909805809661371,
+      "loss": 0.0734,
+      "theoretical_loss": 3.663271723954665,
+      "tokens_seen": 1042546688
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006909003370245547,
+      "loss": 0.074,
+      "theoretical_loss": 3.6631854279839513,
+      "tokens_seen": 1042808832
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006908200930829722,
+      "loss": 0.0703,
+      "theoretical_loss": 3.6630991597762024,
+      "tokens_seen": 1043070976
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006907398491413899,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6630129193155128,
+      "tokens_seen": 1043333120
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006906596051998074,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6629267065859894,
+      "tokens_seen": 1043595264
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000690579361258225,
+      "loss": 0.0756,
+      "theoretical_loss": 3.662840521571753,
+      "tokens_seen": 1043857408
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006904991173166426,
+      "loss": 0.0741,
+      "theoretical_loss": 3.662754364256937,
+      "tokens_seen": 1044119552
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006904188733750601,
+      "loss": 0.0773,
+      "theoretical_loss": 3.662668234625688,
+      "tokens_seen": 1044381696
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006903386294334778,
+      "loss": 0.072,
+      "theoretical_loss": 3.6625821326621653,
+      "tokens_seen": 1044643840
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006902583854918954,
+      "loss": 0.0761,
+      "theoretical_loss": 3.6624960583505404,
+      "tokens_seen": 1044905984
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000690178141550313,
+      "loss": 0.0717,
+      "theoretical_loss": 3.662410011675001,
+      "tokens_seen": 1045168128
+    },
+    {
+      "epoch": 0.32,
+      "objective/train/advantage_avg": 0.0007527482812292874,
+      "objective/train/docs_used": 383342,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4821735620498657,
+      "objective/train/original_loss": 1.4821735620498657,
+      "objective/train/theoretical_loss": 3.6623669986958247,
+      "objective/train/tokens_used": 1065759200,
+      "objective/train/value_avg": -0.008392333984375,
+      "objective/train/value_loss": 0.00048586874618195,
+      "objective/train/value_max": -0.00010722875595092773,
+      "objective/train/value_min": -0.705078125,
+      "objective/train/value_reward_corr": 0.709452779874576,
+      "objective/train/value_std": 0.0186614990234375,
+      "objective/train/weight_avg": 1.000962257385254,
+      "objective/train/weighted_lm_loss": 1.482825517654419,
+      "objective/train/weights_max": 1.404705286026001,
+      "objective/train/weights_min": 0.36978310346603394,
+      "theoretical_loss": 3.6623669986958247,
+      "tokens_seen": 1045299200
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006900978976087306,
+      "loss": 0.0748,
+      "theoretical_loss": 3.6623239926197444,
+      "tokens_seen": 1045430272
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006900176536671482,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6622380011689826,
+      "tokens_seen": 1045692416
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006899374097255657,
+      "loss": 0.0736,
+      "theoretical_loss": 3.66215203730694,
+      "tokens_seen": 1045954560
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006898571657839833,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6620661010178543,
+      "tokens_seen": 1046216704
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006897769218424009,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6619801922859763,
+      "tokens_seen": 1046478848
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006896966779008184,
+      "loss": 0.0728,
+      "theoretical_loss": 3.661894311095568,
+      "tokens_seen": 1046740992
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006896164339592362,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6618084574309075,
+      "tokens_seen": 1047003136
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006895361900176537,
+      "loss": 0.0727,
+      "theoretical_loss": 3.6617226312762834,
+      "tokens_seen": 1047265280
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006894559460760713,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6616368326159976,
+      "tokens_seen": 1047527424
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006893757021344889,
+      "loss": 0.0748,
+      "theoretical_loss": 3.6615510614343654,
+      "tokens_seen": 1047789568
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006892954581929064,
+      "loss": 0.0714,
+      "theoretical_loss": 3.661465317715715,
+      "tokens_seen": 1048051712
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000689215214251324,
+      "loss": 0.073,
+      "theoretical_loss": 3.6613796014443865,
+      "tokens_seen": 1048313856
+    },
+    {
+      "epoch": 0.32,
+      "objective/train/advantage_avg": 0.0009373921202495694,
+      "objective/train/docs_used": 384505,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5765043497085571,
+      "objective/train/original_loss": 1.5765044689178467,
+      "objective/train/theoretical_loss": 3.661293912604734,
+      "objective/train/tokens_used": 1069036000,
+      "objective/train/value_avg": -0.007293701171875,
+      "objective/train/value_loss": 0.00017708793166093528,
+      "objective/train/value_max": -5.561113357543945e-05,
+      "objective/train/value_min": -0.2237548828125,
+      "objective/train/value_reward_corr": 0.6360596224023429,
+      "objective/train/value_std": 0.0101165771484375,
+      "objective/train/weight_avg": 1.001020908355713,
+      "objective/train/weighted_lm_loss": 1.5784999132156372,
+      "objective/train/weights_max": 1.112996220588684,
+      "objective/train/weights_min": 0.3697379231452942,
+      "theoretical_loss": 3.661293912604734,
+      "tokens_seen": 1048576000
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006891349703097416,
+      "loss": 0.0727,
+      "theoretical_loss": 3.661293912604734,
+      "tokens_seen": 1048576000
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006890547263681592,
+      "loss": 0.0737,
+      "theoretical_loss": 3.661208251181124,
+      "tokens_seen": 1048838144
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006889744824265767,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6611226171579356,
+      "tokens_seen": 1049100288
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006888942384849945,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6610370105195607,
+      "tokens_seen": 1049362432
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000688813994543412,
+      "loss": 0.0753,
+      "theoretical_loss": 3.660951431250405,
+      "tokens_seen": 1049624576
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006887337506018296,
+      "loss": 0.0712,
+      "theoretical_loss": 3.6608658793348847,
+      "tokens_seen": 1049886720
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006886535066602472,
+      "loss": 0.074,
+      "theoretical_loss": 3.6607803547574314,
+      "tokens_seen": 1050148864
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006885732627186647,
+      "loss": 0.0755,
+      "theoretical_loss": 3.660694857502487,
+      "tokens_seen": 1050411008
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006884930187770824,
+      "loss": 0.073,
+      "theoretical_loss": 3.660609387554509,
+      "tokens_seen": 1050673152
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006884127748354999,
+      "loss": 0.0763,
+      "theoretical_loss": 3.6605239448979647,
+      "tokens_seen": 1050935296
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006883325308939175,
+      "loss": 0.0751,
+      "theoretical_loss": 3.660438529517336,
+      "tokens_seen": 1051197440
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006882522869523351,
+      "loss": 0.0762,
+      "theoretical_loss": 3.660353141397116,
+      "tokens_seen": 1051459584
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006881720430107528,
+      "loss": 0.0766,
+      "theoretical_loss": 3.6602677805218127,
+      "tokens_seen": 1051721728
+    },
+    {
+      "epoch": 0.32,
+      "objective/train/advantage_avg": 0.00259937415830791,
+      "objective/train/docs_used": 385698,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4460582733154297,
+      "objective/train/original_loss": 1.4460582733154297,
+      "objective/train/theoretical_loss": 3.660225110296166,
+      "objective/train/tokens_used": 1072312800,
+      "objective/train/value_avg": -0.0083160400390625,
+      "objective/train/value_loss": 0.0003496213466860354,
+      "objective/train/value_max": -9.387731552124023e-05,
+      "objective/train/value_min": -0.708984375,
+      "objective/train/value_reward_corr": 0.6992441106554197,
+      "objective/train/value_std": 0.0217742919921875,
+      "objective/train/weight_avg": 1.0027629137039185,
+      "objective/train/weighted_lm_loss": 1.450356125831604,
+      "objective/train/weights_max": 1.4854079484939575,
+      "objective/train/weights_min": 0.3811357617378235,
+      "theoretical_loss": 3.660225110296166,
+      "tokens_seen": 1051852800
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006880917990691703,
+      "loss": 0.0753,
+      "theoretical_loss": 3.660182446875944,
+      "tokens_seen": 1051983872
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006880115551275879,
+      "loss": 0.0699,
+      "theoretical_loss": 3.6600971404440434,
+      "tokens_seen": 1052246016
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006879313111860055,
+      "loss": 0.0714,
+      "theoretical_loss": 3.660011861210654,
+      "tokens_seen": 1052508160
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000687851067244423,
+      "loss": 0.0713,
+      "theoretical_loss": 3.659926609160334,
+      "tokens_seen": 1052770304
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006877708233028407,
+      "loss": 0.0699,
+      "theoretical_loss": 3.6598413842776534,
+      "tokens_seen": 1053032448
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006876905793612582,
+      "loss": 0.073,
+      "theoretical_loss": 3.6597561865471935,
+      "tokens_seen": 1053294592
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006876103354196759,
+      "loss": 0.0773,
+      "theoretical_loss": 3.6596710159535504,
+      "tokens_seen": 1053556736
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006875300914780934,
+      "loss": 0.0752,
+      "theoretical_loss": 3.659585872481331,
+      "tokens_seen": 1053818880
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006874498475365109,
+      "loss": 0.075,
+      "theoretical_loss": 3.659500756115156,
+      "tokens_seen": 1054081024
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006873696035949287,
+      "loss": 0.0736,
+      "theoretical_loss": 3.659415666839658,
+      "tokens_seen": 1054343168
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006872893596533462,
+      "loss": 0.0723,
+      "theoretical_loss": 3.6593306046394813,
+      "tokens_seen": 1054605312
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006872091157117638,
+      "loss": 0.0748,
+      "theoretical_loss": 3.6592455694992854,
+      "tokens_seen": 1054867456
+    },
+    {
+      "epoch": 0.32,
+      "objective/train/advantage_avg": 0.0018493332900106907,
+      "objective/train/docs_used": 386814,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4244455099105835,
+      "objective/train/original_loss": 1.424445390701294,
+      "objective/train/theoretical_loss": 3.659160561403739,
+      "objective/train/tokens_used": 1075589600,
+      "objective/train/value_avg": -0.00870513916015625,
+      "objective/train/value_loss": 0.00016267823230009526,
+      "objective/train/value_max": -0.00012636184692382812,
+      "objective/train/value_min": -0.32421875,
+      "objective/train/value_reward_corr": 0.6824163394532532,
+      "objective/train/value_std": 0.01215362548828125,
+      "objective/train/weight_avg": 1.0019264221191406,
+      "objective/train/weighted_lm_loss": 1.426944375038147,
+      "objective/train/weights_max": 1.2696858644485474,
+      "objective/train/weights_min": 0.3791867196559906,
+      "theoretical_loss": 3.659160561403739,
+      "tokens_seen": 1055129600
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006871288717701814,
+      "loss": 0.0698,
+      "theoretical_loss": 3.659160561403739,
+      "tokens_seen": 1055129600
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000687048627828599,
+      "loss": 0.0717,
+      "theoretical_loss": 3.6590755803375252,
+      "tokens_seen": 1055391744
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006869683838870165,
+      "loss": 0.0734,
+      "theoretical_loss": 3.65899062628534,
+      "tokens_seen": 1055653888
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006868881399454341,
+      "loss": 0.0726,
+      "theoretical_loss": 3.65890569923189,
+      "tokens_seen": 1055916032
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006868078960038517,
+      "loss": 0.0742,
+      "theoretical_loss": 3.658820799161896,
+      "tokens_seen": 1056178176
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006867276520622692,
+      "loss": 0.0748,
+      "theoretical_loss": 3.65873592606009,
+      "tokens_seen": 1056440320
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000686647408120687,
+      "loss": 0.0707,
+      "theoretical_loss": 3.658651079911218,
+      "tokens_seen": 1056702464
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006865671641791045,
+      "loss": 0.0736,
+      "theoretical_loss": 3.658566260700036,
+      "tokens_seen": 1056964608
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006864869202375221,
+      "loss": 0.0734,
+      "theoretical_loss": 3.658481468411315,
+      "tokens_seen": 1057226752
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006864066762959397,
+      "loss": 0.0714,
+      "theoretical_loss": 3.6583967030298368,
+      "tokens_seen": 1057488896
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006863264323543572,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6583119645403954,
+      "tokens_seen": 1057751040
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006862461884127749,
+      "loss": 0.071,
+      "theoretical_loss": 3.658227252927799,
+      "tokens_seen": 1058013184
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006861659444711924,
+      "loss": 0.0704,
+      "theoretical_loss": 3.6581425681768653,
+      "tokens_seen": 1058275328
+    },
+    {
+      "epoch": 0.32,
+      "objective/train/advantage_avg": 0.0014182066079229116,
+      "objective/train/docs_used": 388023,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3999180793762207,
+      "objective/train/original_loss": 1.3999180793762207,
+      "objective/train/theoretical_loss": 3.6581002358697816,
+      "objective/train/tokens_used": 1078866400,
+      "objective/train/value_avg": -0.0078582763671875,
+      "objective/train/value_loss": 0.00017249694792553782,
+      "objective/train/value_max": -0.00011771917343139648,
+      "objective/train/value_min": -0.32958984375,
+      "objective/train/value_reward_corr": 0.6715197510150517,
+      "objective/train/value_std": 0.01299285888671875,
+      "objective/train/weight_avg": 1.001496434211731,
+      "objective/train/weighted_lm_loss": 1.4018974304199219,
+      "objective/train/weights_max": 1.1797270774841309,
+      "objective/train/weights_min": 0.3719121813774109,
+      "theoretical_loss": 3.6581002358697816,
+      "tokens_seen": 1058406400
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.00068608570052961,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6580579102724267,
+      "tokens_seen": 1058537472
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006860054565880276,
+      "loss": 0.0734,
+      "theoretical_loss": 3.657973279199327,
+      "tokens_seen": 1058799616
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006859252126464453,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6578886749424226,
+      "tokens_seen": 1059061760
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006858449687048628,
+      "loss": 0.077,
+      "theoretical_loss": 3.657804097486581,
+      "tokens_seen": 1059323904
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006857647247632804,
+      "loss": 0.0704,
+      "theoretical_loss": 3.657719546816685,
+      "tokens_seen": 1059586048
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000685684480821698,
+      "loss": 0.071,
+      "theoretical_loss": 3.657635022917626,
+      "tokens_seen": 1059848192
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006856042368801155,
+      "loss": 0.0714,
+      "theoretical_loss": 3.65755052577431,
+      "tokens_seen": 1060110336
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006855239929385332,
+      "loss": 0.071,
+      "theoretical_loss": 3.657466055371654,
+      "tokens_seen": 1060372480
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006854437489969507,
+      "loss": 0.0722,
+      "theoretical_loss": 3.657381611694588,
+      "tokens_seen": 1060634624
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006853635050553683,
+      "loss": 0.073,
+      "theoretical_loss": 3.6572971947280544,
+      "tokens_seen": 1060896768
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006852832611137859,
+      "loss": 0.0763,
+      "theoretical_loss": 3.6572128044570067,
+      "tokens_seen": 1061158912
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006852030171722034,
+      "loss": 0.0725,
+      "theoretical_loss": 3.657128440866412,
+      "tokens_seen": 1061421056
+    },
+    {
+      "epoch": 0.32,
+      "objective/train/advantage_avg": 0.0011737276799976826,
+      "objective/train/docs_used": 389244,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5271493196487427,
+      "objective/train/original_loss": 1.5271493196487427,
+      "objective/train/theoretical_loss": 3.6570441039412485,
+      "objective/train/tokens_used": 1082143200,
+      "objective/train/value_avg": -0.00862884521484375,
+      "objective/train/value_loss": 0.0003618821792770177,
+      "objective/train/value_max": -0.00017261505126953125,
+      "objective/train/value_min": -0.403076171875,
+      "objective/train/value_reward_corr": 0.6260921546392633,
+      "objective/train/value_std": 0.01381683349609375,
+      "objective/train/weight_avg": 1.0013278722763062,
+      "objective/train/weighted_lm_loss": 1.529421329498291,
+      "objective/train/weights_max": 1.1920899152755737,
+      "objective/train/weights_min": 0.37519246339797974,
+      "theoretical_loss": 3.6570441039412485,
+      "tokens_seen": 1061683200
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006851227732306211,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6570441039412485,
+      "tokens_seen": 1061683200
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006850425292890387,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6569597936665064,
+      "tokens_seen": 1061945344
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006849622853474563,
+      "loss": 0.0754,
+      "theoretical_loss": 3.6568755100271897,
+      "tokens_seen": 1062207488
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006848820414058739,
+      "loss": 0.0727,
+      "theoretical_loss": 3.656791253008313,
+      "tokens_seen": 1062469632
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006848017974642915,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6567070225949028,
+      "tokens_seen": 1062731776
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000684721553522709,
+      "loss": 0.0727,
+      "theoretical_loss": 3.656622818771999,
+      "tokens_seen": 1062993920
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006846413095811267,
+      "loss": 0.0711,
+      "theoretical_loss": 3.6565386415246524,
+      "tokens_seen": 1063256064
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006845610656395442,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6564544908379273,
+      "tokens_seen": 1063518208
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006844808216979617,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6563703666968985,
+      "tokens_seen": 1063780352
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006844005777563795,
+      "loss": 0.0709,
+      "theoretical_loss": 3.656286269086653,
+      "tokens_seen": 1064042496
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000684320333814797,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6562021979922923,
+      "tokens_seen": 1064304640
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006842400898732146,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6561181533989267,
+      "tokens_seen": 1064566784
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006841598459316322,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6560341352916796,
+      "tokens_seen": 1064828928
+    },
+    {
+      "epoch": 0.32,
+      "objective/train/advantage_avg": 0.001128622330725193,
+      "objective/train/docs_used": 390462,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3764036893844604,
+      "objective/train/original_loss": 1.376403570175171,
+      "objective/train/theoretical_loss": 3.6559921361657057,
+      "objective/train/tokens_used": 1085420000,
+      "objective/train/value_avg": -0.01018524169921875,
+      "objective/train/value_loss": 0.00021231910795904696,
+      "objective/train/value_max": -6.35385513305664e-05,
+      "objective/train/value_min": -0.358642578125,
+      "objective/train/value_reward_corr": 0.7759159261494163,
+      "objective/train/value_std": 0.0171661376953125,
+      "objective/train/weight_avg": 1.0012286901474,
+      "objective/train/weighted_lm_loss": 1.3785223960876465,
+      "objective/train/weights_max": 1.2096216678619385,
+      "objective/train/weights_min": 0.37203988432884216,
+      "theoretical_loss": 3.6559921361657057,
+      "tokens_seen": 1064960000
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006840796019900498,
+      "loss": 0.073,
+      "theoretical_loss": 3.655950143655688,
+      "tokens_seen": 1065091072
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006839993580484673,
+      "loss": 0.0734,
+      "theoretical_loss": 3.655866178476098,
+      "tokens_seen": 1065353216
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006839191141068849,
+      "loss": 0.0728,
+      "theoretical_loss": 3.65578223973807,
+      "tokens_seen": 1065615360
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006838388701653025,
+      "loss": 0.077,
+      "theoretical_loss": 3.6556983274267765,
+      "tokens_seen": 1065877504
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006837586262237201,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6556144415273994,
+      "tokens_seen": 1066139648
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006836783822821378,
+      "loss": 0.0734,
+      "theoretical_loss": 3.655530582025136,
+      "tokens_seen": 1066401792
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006835981383405553,
+      "loss": 0.0718,
+      "theoretical_loss": 3.6554467489051925,
+      "tokens_seen": 1066663936
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000683517894398973,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6553629421527885,
+      "tokens_seen": 1066926080
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006834376504573905,
+      "loss": 0.0743,
+      "theoretical_loss": 3.655279161753156,
+      "tokens_seen": 1067188224
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000683357406515808,
+      "loss": 0.0712,
+      "theoretical_loss": 3.6551954076915374,
+      "tokens_seen": 1067450368
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006832771625742257,
+      "loss": 0.0727,
+      "theoretical_loss": 3.655111679953188,
+      "tokens_seen": 1067712512
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006831969186326432,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6550279785233757,
+      "tokens_seen": 1067974656
+    },
+    {
+      "epoch": 0.32,
+      "objective/train/advantage_avg": 0.0003532489645294845,
+      "objective/train/docs_used": 391774,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5375131368637085,
+      "objective/train/original_loss": 1.5375131368637085,
+      "objective/train/theoretical_loss": 3.654944303387378,
+      "objective/train/tokens_used": 1088696800,
+      "objective/train/value_avg": -0.00994873046875,
+      "objective/train/value_loss": 0.00020713380945380777,
+      "objective/train/value_max": -7.724761962890625e-05,
+      "objective/train/value_min": -0.334228515625,
+      "objective/train/value_reward_corr": 0.7750881540426269,
+      "objective/train/value_std": 0.0160369873046875,
+      "objective/train/weight_avg": 1.000451683998108,
+      "objective/train/weighted_lm_loss": 1.5384892225265503,
+      "objective/train/weights_max": 1.286349892616272,
+      "objective/train/weights_min": 0.3698126971721649,
+      "theoretical_loss": 3.654944303387378,
+      "tokens_seen": 1068236800
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006831166746910608,
+      "loss": 0.0735,
+      "theoretical_loss": 3.654944303387378,
+      "tokens_seen": 1068236800
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006830364307494784,
+      "loss": 0.073,
+      "theoretical_loss": 3.654860654530486,
+      "tokens_seen": 1068498944
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006829561868078961,
+      "loss": 0.0742,
+      "theoretical_loss": 3.6547770319380026,
+      "tokens_seen": 1068761088
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006828759428663136,
+      "loss": 0.0712,
+      "theoretical_loss": 3.6546934355952425,
+      "tokens_seen": 1069023232
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006827956989247312,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6546098654875303,
+      "tokens_seen": 1069285376
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006827154549831488,
+      "loss": 0.0743,
+      "theoretical_loss": 3.654526321600205,
+      "tokens_seen": 1069547520
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006826352110415663,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6544428039186165,
+      "tokens_seen": 1069809664
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000682554967099984,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6543593124281264,
+      "tokens_seen": 1070071808
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006824747231584015,
+      "loss": 0.074,
+      "theoretical_loss": 3.654275847114107,
+      "tokens_seen": 1070333952
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006823944792168192,
+      "loss": 0.0762,
+      "theoretical_loss": 3.6541924079619443,
+      "tokens_seen": 1070596096
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006823142352752367,
+      "loss": 0.0741,
+      "theoretical_loss": 3.654108994957034,
+      "tokens_seen": 1070858240
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006822339913336542,
+      "loss": 0.0751,
+      "theoretical_loss": 3.654025608084786,
+      "tokens_seen": 1071120384
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.000682153747392072,
+      "loss": 0.0744,
+      "theoretical_loss": 3.653942247330619,
+      "tokens_seen": 1071382528
+    },
+    {
+      "epoch": 0.32,
+      "objective/train/advantage_avg": 0.0008207072969526052,
+      "objective/train/docs_used": 393037,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5273302793502808,
+      "objective/train/original_loss": 1.5273303985595703,
+      "objective/train/theoretical_loss": 3.6539005767432635,
+      "objective/train/tokens_used": 1091973600,
+      "objective/train/value_avg": -0.0079193115234375,
+      "objective/train/value_loss": 0.00018726506095845252,
+      "objective/train/value_max": -4.792213439941406e-05,
+      "objective/train/value_min": -0.328125,
+      "objective/train/value_reward_corr": 0.6683278593111599,
+      "objective/train/value_std": 0.0133209228515625,
+      "objective/train/weight_avg": 1.0009064674377441,
+      "objective/train/weighted_lm_loss": 1.5285539627075195,
+      "objective/train/weights_max": 1.301619052886963,
+      "objective/train/weights_min": 0.3697463870048523,
+      "theoretical_loss": 3.6539005767432635,
+      "tokens_seen": 1071513600
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006820735034504895,
+      "loss": 0.0727,
+      "theoretical_loss": 3.653858912679966,
+      "tokens_seen": 1071644672
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006819932595089071,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6537756041182696,
+      "tokens_seen": 1071906816
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006819130155673247,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6536923216309862,
+      "tokens_seen": 1072168960
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 0.0006818327716257423,
+      "loss": 0.0744,
+      "theoretical_loss": 3.653609065203582,
+      "tokens_seen": 1072431104
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006817525276841598,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6535258348215356,
+      "tokens_seen": 1072693248
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006816722837425775,
+      "loss": 0.0748,
+      "theoretical_loss": 3.653442630470337,
+      "tokens_seen": 1072955392
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000681592039800995,
+      "loss": 0.074,
+      "theoretical_loss": 3.653359452135488,
+      "tokens_seen": 1073217536
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006815117958594125,
+      "loss": 0.0737,
+      "theoretical_loss": 3.653276299802503,
+      "tokens_seen": 1073479680
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006814315519178303,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6531931734569056,
+      "tokens_seen": 1073741824
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006813513079762478,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6531100730842336,
+      "tokens_seen": 1074003968
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006812710640346655,
+      "loss": 0.0747,
+      "theoretical_loss": 3.653026998670035,
+      "tokens_seen": 1074266112
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000681190820093083,
+      "loss": 0.0734,
+      "theoretical_loss": 3.652943950199869,
+      "tokens_seen": 1074528256
+    },
+    {
+      "epoch": 0.33,
+      "objective/train/advantage_avg": 0.0013173749903216958,
+      "objective/train/docs_used": 394219,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3774545192718506,
+      "objective/train/original_loss": 1.3774546384811401,
+      "objective/train/theoretical_loss": 3.652860927659307,
+      "objective/train/tokens_used": 1095250400,
+      "objective/train/value_avg": -0.009002685546875,
+      "objective/train/value_loss": 0.000231547121074982,
+      "objective/train/value_max": -6.300210952758789e-05,
+      "objective/train/value_min": -0.359130859375,
+      "objective/train/value_reward_corr": 0.6883898461921806,
+      "objective/train/value_std": 0.01419830322265625,
+      "objective/train/weight_avg": 1.0014238357543945,
+      "objective/train/weighted_lm_loss": 1.3791875839233398,
+      "objective/train/weights_max": 1.1214268207550049,
+      "objective/train/weights_min": 0.3686865568161011,
+      "theoretical_loss": 3.652860927659307,
+      "tokens_seen": 1074790400
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006811105761515006,
+      "loss": 0.0735,
+      "theoretical_loss": 3.652860927659307,
+      "tokens_seen": 1074790400
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006810303322099182,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6527779310339326,
+      "tokens_seen": 1075052544
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006809500882683357,
+      "loss": 0.0741,
+      "theoretical_loss": 3.652694960309339,
+      "tokens_seen": 1075314688
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006808698443267533,
+      "loss": 0.075,
+      "theoretical_loss": 3.6526120154711332,
+      "tokens_seen": 1075576832
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006807896003851709,
+      "loss": 0.0721,
+      "theoretical_loss": 3.6525290965049324,
+      "tokens_seen": 1075838976
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006807093564435886,
+      "loss": 0.0738,
+      "theoretical_loss": 3.652446203396365,
+      "tokens_seen": 1076101120
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006806291125020061,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6523633361310717,
+      "tokens_seen": 1076363264
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006805488685604238,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6522804946947045,
+      "tokens_seen": 1076625408
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006804686246188413,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6521976790729265,
+      "tokens_seen": 1076887552
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006803883806772588,
+      "loss": 0.0727,
+      "theoretical_loss": 3.652114889251412,
+      "tokens_seen": 1077149696
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006803081367356765,
+      "loss": 0.0764,
+      "theoretical_loss": 3.6520321252158485,
+      "tokens_seen": 1077411840
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000680227892794094,
+      "loss": 0.0723,
+      "theoretical_loss": 3.651949386951933,
+      "tokens_seen": 1077673984
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006801476488525116,
+      "loss": 0.073,
+      "theoretical_loss": 3.6518666744453734,
+      "tokens_seen": 1077936128
+    },
+    {
+      "epoch": 0.33,
+      "objective/train/advantage_avg": 0.0019057122990489006,
+      "objective/train/docs_used": 395298,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3716213703155518,
+      "objective/train/original_loss": 1.3716213703155518,
+      "objective/train/theoretical_loss": 3.6518253278466397,
+      "objective/train/tokens_used": 1098527200,
+      "objective/train/value_avg": -0.007762908935546875,
+      "objective/train/value_loss": 0.00023904572299215943,
+      "objective/train/value_max": -8.219480514526367e-05,
+      "objective/train/value_min": -0.199462890625,
+      "objective/train/value_reward_corr": 0.600056819574208,
+      "objective/train/value_std": 0.0146636962890625,
+      "objective/train/weight_avg": 1.0020171403884888,
+      "objective/train/weighted_lm_loss": 1.3752607107162476,
+      "objective/train/weights_max": 1.1991831064224243,
+      "objective/train/weights_min": 0.3681375980377197,
+      "theoretical_loss": 3.6518253278466397,
+      "tokens_seen": 1078067200
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006800674049109292,
+      "loss": 0.0731,
+      "theoretical_loss": 3.651783987681892,
+      "tokens_seen": 1078198272
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006799871609693469,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6517013266472187,
+      "tokens_seen": 1078460416
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006799069170277645,
+      "loss": 0.0735,
+      "theoretical_loss": 3.651618691327098,
+      "tokens_seen": 1078722560
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000679826673086182,
+      "loss": 0.075,
+      "theoretical_loss": 3.651536081707284,
+      "tokens_seen": 1078984704
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006797464291445996,
+      "loss": 0.0759,
+      "theoretical_loss": 3.651453497773543,
+      "tokens_seen": 1079246848
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006796661852030172,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6513709395116516,
+      "tokens_seen": 1079508992
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006795859412614348,
+      "loss": 0.0716,
+      "theoretical_loss": 3.651288406907399,
+      "tokens_seen": 1079771136
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006795056973198523,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6512058999465844,
+      "tokens_seen": 1080033280
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.00067942545337827,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6511234186150197,
+      "tokens_seen": 1080295424
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006793452094366875,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6510409628985263,
+      "tokens_seen": 1080557568
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000679264965495105,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6509585327829392,
+      "tokens_seen": 1080819712
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006791847215535228,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6508761282541027,
+      "tokens_seen": 1081081856
+    },
+    {
+      "epoch": 0.33,
+      "objective/train/advantage_avg": 0.0018546562641859055,
+      "objective/train/docs_used": 396489,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.54283607006073,
+      "objective/train/original_loss": 1.5428359508514404,
+      "objective/train/theoretical_loss": 3.6507937492978733,
+      "objective/train/tokens_used": 1101804000,
+      "objective/train/value_avg": -0.007076263427734375,
+      "objective/train/value_loss": 0.0001660433190409094,
+      "objective/train/value_max": -0.00012242794036865234,
+      "objective/train/value_min": -0.2880859375,
+      "objective/train/value_reward_corr": 0.6533642274011282,
+      "objective/train/value_std": 0.01088714599609375,
+      "objective/train/weight_avg": 1.0019302368164062,
+      "objective/train/weighted_lm_loss": 1.5454530715942383,
+      "objective/train/weights_max": 1.150571584701538,
+      "objective/train/weights_min": 0.37533560395240784,
+      "theoretical_loss": 3.6507937492978733,
+      "tokens_seen": 1081344000
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006791044776119403,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6507937492978733,
+      "tokens_seen": 1081344000
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006790242336703579,
+      "loss": 0.0713,
+      "theoretical_loss": 3.6507113959001183,
+      "tokens_seen": 1081606144
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006789439897287755,
+      "loss": 0.0771,
+      "theoretical_loss": 3.6506290680467166,
+      "tokens_seen": 1081868288
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006788637457871931,
+      "loss": 0.0739,
+      "theoretical_loss": 3.650546765723558,
+      "tokens_seen": 1082130432
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006787835018456106,
+      "loss": 0.0729,
+      "theoretical_loss": 3.650464488916544,
+      "tokens_seen": 1082392576
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006787032579040283,
+      "loss": 0.0746,
+      "theoretical_loss": 3.650382237611587,
+      "tokens_seen": 1082654720
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006786230139624458,
+      "loss": 0.0764,
+      "theoretical_loss": 3.65030001179461,
+      "tokens_seen": 1082916864
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006785427700208634,
+      "loss": 0.0735,
+      "theoretical_loss": 3.650217811451548,
+      "tokens_seen": 1083179008
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006784625260792811,
+      "loss": 0.0728,
+      "theoretical_loss": 3.650135636568347,
+      "tokens_seen": 1083441152
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006783822821376986,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6500534871309642,
+      "tokens_seen": 1083703296
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006783020381961163,
+      "loss": 0.0737,
+      "theoretical_loss": 3.649971363125368,
+      "tokens_seen": 1083965440
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006782217942545338,
+      "loss": 0.074,
+      "theoretical_loss": 3.6498892645375367,
+      "tokens_seen": 1084227584
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006781415503129514,
+      "loss": 0.0727,
+      "theoretical_loss": 3.649807191353462,
+      "tokens_seen": 1084489728
+    },
+    {
+      "epoch": 0.33,
+      "objective/train/advantage_avg": 0.000662712671328336,
+      "objective/train/docs_used": 397685,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3404814004898071,
+      "objective/train/original_loss": 1.3404817581176758,
+      "objective/train/theoretical_loss": 3.649766164283458,
+      "objective/train/tokens_used": 1105080800,
+      "objective/train/value_avg": -0.0102081298828125,
+      "objective/train/value_loss": 0.00020532849885057658,
+      "objective/train/value_max": -0.00010472536087036133,
+      "objective/train/value_min": -0.33837890625,
+      "objective/train/value_reward_corr": 0.7515684867138178,
+      "objective/train/value_std": 0.01517486572265625,
+      "objective/train/weight_avg": 1.0007604360580444,
+      "objective/train/weighted_lm_loss": 1.3414093255996704,
+      "objective/train/weights_max": 1.1722419261932373,
+      "objective/train/weights_min": 0.4255511462688446,
+      "theoretical_loss": 3.649766164283458,
+      "tokens_seen": 1084620800
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000678061306371369,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6497251435591442,
+      "tokens_seen": 1084751872
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006779810624297865,
+      "loss": 0.0775,
+      "theoretical_loss": 3.6496431211405973,
+      "tokens_seen": 1085014016
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006779008184882041,
+      "loss": 0.0714,
+      "theoretical_loss": 3.649561124083844,
+      "tokens_seen": 1085276160
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006778205745466217,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6494791523749193,
+      "tokens_seen": 1085538304
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006777403306050394,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6493972059998696,
+      "tokens_seen": 1085800448
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006776600866634569,
+      "loss": 0.0735,
+      "theoretical_loss": 3.649315284944751,
+      "tokens_seen": 1086062592
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006775798427218746,
+      "loss": 0.072,
+      "theoretical_loss": 3.649233389195632,
+      "tokens_seen": 1086324736
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006774995987802921,
+      "loss": 0.0736,
+      "theoretical_loss": 3.6491515187385914,
+      "tokens_seen": 1086586880
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006774193548387097,
+      "loss": 0.071,
+      "theoretical_loss": 3.649069673559719,
+      "tokens_seen": 1086849024
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006773391108971273,
+      "loss": 0.0729,
+      "theoretical_loss": 3.648987853645116,
+      "tokens_seen": 1087111168
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006772588669555448,
+      "loss": 0.0715,
+      "theoretical_loss": 3.648906058980894,
+      "tokens_seen": 1087373312
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006771786230139625,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6488242895531764,
+      "tokens_seen": 1087635456
+    },
+    {
+      "epoch": 0.33,
+      "objective/train/advantage_avg": -0.0011196710402145982,
+      "objective/train/docs_used": 398862,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.379757046699524,
+      "objective/train/original_loss": 1.3797568082809448,
+      "objective/train/theoretical_loss": 3.6487425453480973,
+      "objective/train/tokens_used": 1108357600,
+      "objective/train/value_avg": -0.0118865966796875,
+      "objective/train/value_loss": 0.0003890149819198996,
+      "objective/train/value_max": -0.00010073184967041016,
+      "objective/train/value_min": -0.65087890625,
+      "objective/train/value_reward_corr": 0.7589761731426594,
+      "objective/train/value_std": 0.0198974609375,
+      "objective/train/weight_avg": 0.9990611672401428,
+      "objective/train/weighted_lm_loss": 1.3779809474945068,
+      "objective/train/weights_max": 1.7288053035736084,
+      "objective/train/weights_min": 0.3692087233066559,
+      "theoretical_loss": 3.6487425453480973,
+      "tokens_seen": 1087897600
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.00067709837907238,
+      "loss": 0.0722,
+      "theoretical_loss": 3.6487425453480973,
+      "tokens_seen": 1087897600
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006770181351307977,
+      "loss": 0.0712,
+      "theoretical_loss": 3.648660826351801,
+      "tokens_seen": 1088159744
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006769378911892153,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6485791325504437,
+      "tokens_seen": 1088421888
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006768576472476328,
+      "loss": 0.073,
+      "theoretical_loss": 3.648497463930192,
+      "tokens_seen": 1088684032
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006767774033060504,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6484158204772235,
+      "tokens_seen": 1088946176
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000676697159364468,
+      "loss": 0.0749,
+      "theoretical_loss": 3.648334202177727,
+      "tokens_seen": 1089208320
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006766169154228856,
+      "loss": 0.0764,
+      "theoretical_loss": 3.648252609017902,
+      "tokens_seen": 1089470464
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006765366714813031,
+      "loss": 0.0741,
+      "theoretical_loss": 3.648171040983959,
+      "tokens_seen": 1089732608
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006764564275397208,
+      "loss": 0.0763,
+      "theoretical_loss": 3.648089498062119,
+      "tokens_seen": 1089994752
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006763761835981383,
+      "loss": 0.0753,
+      "theoretical_loss": 3.648007980238614,
+      "tokens_seen": 1090256896
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006762959396565558,
+      "loss": 0.0723,
+      "theoretical_loss": 3.6479264874996877,
+      "tokens_seen": 1090519040
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006762156957149736,
+      "loss": 0.0727,
+      "theoretical_loss": 3.6478450198315926,
+      "tokens_seen": 1090781184
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006761354517733911,
+      "loss": 0.0754,
+      "theoretical_loss": 3.6477635772205947,
+      "tokens_seen": 1091043328
+    },
+    {
+      "epoch": 0.33,
+      "objective/train/advantage_avg": 0.0015444932505488396,
+      "objective/train/docs_used": 400125,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.572818636894226,
+      "objective/train/original_loss": 1.572818636894226,
+      "objective/train/theoretical_loss": 3.647722865307218,
+      "objective/train/tokens_used": 1111634400,
+      "objective/train/value_avg": -0.007625579833984375,
+      "objective/train/value_loss": 0.00013872672570869327,
+      "objective/train/value_max": -0.00011146068572998047,
+      "objective/train/value_min": -0.3896484375,
+      "objective/train/value_reward_corr": 0.6624823334319214,
+      "objective/train/value_std": 0.0112152099609375,
+      "objective/train/weight_avg": 1.0016119480133057,
+      "objective/train/weighted_lm_loss": 1.574569821357727,
+      "objective/train/weights_max": 1.1813123226165771,
+      "objective/train/weights_min": 0.6096436381340027,
+      "theoretical_loss": 3.647722865307218,
+      "tokens_seen": 1091174400
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006760552078318088,
+      "loss": 0.075,
+      "theoretical_loss": 3.647682159652969,
+      "tokens_seen": 1091305472
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006759749638902263,
+      "loss": 0.074,
+      "theoretical_loss": 3.647600767115002,
+      "tokens_seen": 1091567616
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006758947199486439,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6475193995929907,
+      "tokens_seen": 1091829760
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006758144760070615,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6474380570732423,
+      "tokens_seen": 1092091904
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000675734232065479,
+      "loss": 0.075,
+      "theoretical_loss": 3.6473567395420767,
+      "tokens_seen": 1092354048
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006756539881238966,
+      "loss": 0.0729,
+      "theoretical_loss": 3.647275446985822,
+      "tokens_seen": 1092616192
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006755737441823142,
+      "loss": 0.074,
+      "theoretical_loss": 3.64719417939082,
+      "tokens_seen": 1092878336
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006754935002407319,
+      "loss": 0.0757,
+      "theoretical_loss": 3.6471129367434205,
+      "tokens_seen": 1093140480
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006754132562991494,
+      "loss": 0.0715,
+      "theoretical_loss": 3.647031719029985,
+      "tokens_seen": 1093402624
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006753330123575671,
+      "loss": 0.0731,
+      "theoretical_loss": 3.646950526236887,
+      "tokens_seen": 1093664768
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006752527684159846,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6468693583505085,
+      "tokens_seen": 1093926912
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006751725244744022,
+      "loss": 0.0759,
+      "theoretical_loss": 3.646788215357244,
+      "tokens_seen": 1094189056
+    },
+    {
+      "epoch": 0.33,
+      "objective/train/advantage_avg": 0.0005579335847869515,
+      "objective/train/docs_used": 401268,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.393877625465393,
+      "objective/train/original_loss": 1.3938775062561035,
+      "objective/train/theoretical_loss": 3.646707097243498,
+      "objective/train/tokens_used": 1114911200,
+      "objective/train/value_avg": -0.007781982421875,
+      "objective/train/value_loss": 0.0001916129986057058,
+      "objective/train/value_max": -0.00014770030975341797,
+      "objective/train/value_min": -0.6875,
+      "objective/train/value_reward_corr": 0.7125330836958561,
+      "objective/train/value_std": 0.01245880126953125,
+      "objective/train/weight_avg": 1.000646948814392,
+      "objective/train/weighted_lm_loss": 1.3944873809814453,
+      "objective/train/weights_max": 1.2134755849838257,
+      "objective/train/weights_min": 0.3718355596065521,
+      "theoretical_loss": 3.646707097243498,
+      "tokens_seen": 1094451200
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006750922805328198,
+      "loss": 0.0733,
+      "theoretical_loss": 3.646707097243498,
+      "tokens_seen": 1094451200
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006750120365912373,
+      "loss": 0.0727,
+      "theoretical_loss": 3.646626003995685,
+      "tokens_seen": 1094713344
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000674931792649655,
+      "loss": 0.0757,
+      "theoretical_loss": 3.6465449356002315,
+      "tokens_seen": 1094975488
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006748515487080725,
+      "loss": 0.0713,
+      "theoretical_loss": 3.646463892043574,
+      "tokens_seen": 1095237632
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006747713047664902,
+      "loss": 0.0736,
+      "theoretical_loss": 3.6463828733121586,
+      "tokens_seen": 1095499776
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006746910608249078,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6463018793924453,
+      "tokens_seen": 1095761920
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006746108168833254,
+      "loss": 0.0713,
+      "theoretical_loss": 3.6462209102709,
+      "tokens_seen": 1096024064
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006745305729417429,
+      "loss": 0.0722,
+      "theoretical_loss": 3.6461399659340037,
+      "tokens_seen": 1096286208
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006744503290001605,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6460590463682454,
+      "tokens_seen": 1096548352
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006743700850585781,
+      "loss": 0.0727,
+      "theoretical_loss": 3.6459781515601244,
+      "tokens_seen": 1096810496
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006742898411169956,
+      "loss": 0.0748,
+      "theoretical_loss": 3.6458972814961528,
+      "tokens_seen": 1097072640
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006742095971754133,
+      "loss": 0.0723,
+      "theoretical_loss": 3.6458164361628516,
+      "tokens_seen": 1097334784
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006741293532338308,
+      "loss": 0.0723,
+      "theoretical_loss": 3.645735615546752,
+      "tokens_seen": 1097596928
+    },
+    {
+      "epoch": 0.33,
+      "objective/train/advantage_avg": 0.0027343458496034145,
+      "objective/train/docs_used": 402429,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3593459129333496,
+      "objective/train/original_loss": 1.35934579372406,
+      "objective/train/theoretical_loss": 3.645695214503448,
+      "objective/train/tokens_used": 1118188000,
+      "objective/train/value_avg": -0.00693511962890625,
+      "objective/train/value_loss": 0.00017142666911240667,
+      "objective/train/value_max": -9.316205978393555e-05,
+      "objective/train/value_min": -0.6669921875,
+      "objective/train/value_reward_corr": 0.6122719969080082,
+      "objective/train/value_std": 0.012725830078125,
+      "objective/train/weight_avg": 1.0028126239776611,
+      "objective/train/weighted_lm_loss": 1.363199234008789,
+      "objective/train/weights_max": 1.3812626600265503,
+      "objective/train/weights_min": 0.38981711864471436,
+      "theoretical_loss": 3.645695214503448,
+      "tokens_seen": 1097728000
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006740491092922485,
+      "loss": 0.0753,
+      "theoretical_loss": 3.645654819634397,
+      "tokens_seen": 1097859072
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006739688653506661,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6455740484123407,
+      "tokens_seen": 1098121216
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006738886214090836,
+      "loss": 0.0713,
+      "theoretical_loss": 3.645493301867145,
+      "tokens_seen": 1098383360
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006738083774675012,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6454125799853854,
+      "tokens_seen": 1098645504
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006737281335259188,
+      "loss": 0.0735,
+      "theoretical_loss": 3.645331882753645,
+      "tokens_seen": 1098907648
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006736478895843364,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6452512101585195,
+      "tokens_seen": 1099169792
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000673567645642754,
+      "loss": 0.0701,
+      "theoretical_loss": 3.645170562186615,
+      "tokens_seen": 1099431936
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006734874017011716,
+      "loss": 0.0742,
+      "theoretical_loss": 3.6450899388245466,
+      "tokens_seen": 1099694080
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006734071577595891,
+      "loss": 0.0747,
+      "theoretical_loss": 3.645009340058941,
+      "tokens_seen": 1099956224
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006733269138180067,
+      "loss": 0.0734,
+      "theoretical_loss": 3.644928765876436,
+      "tokens_seen": 1100218368
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006732466698764244,
+      "loss": 0.0722,
+      "theoretical_loss": 3.644848216263678,
+      "tokens_seen": 1100480512
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006731664259348419,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6447676912073255,
+      "tokens_seen": 1100742656
+    },
+    {
+      "epoch": 0.33,
+      "objective/train/advantage_avg": -0.0021009258925914764,
+      "objective/train/docs_used": 403530,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.643606185913086,
+      "objective/train/original_loss": 1.6436063051223755,
+      "objective/train/theoretical_loss": 3.6446871906940466,
+      "objective/train/tokens_used": 1121464800,
+      "objective/train/value_avg": -0.01171875,
+      "objective/train/value_loss": 0.00041994385537691414,
+      "objective/train/value_max": -0.00010311603546142578,
+      "objective/train/value_min": -0.2127685546875,
+      "objective/train/value_reward_corr": 0.8979253185065612,
+      "objective/train/value_std": 0.0230712890625,
+      "objective/train/weight_avg": 0.998100221157074,
+      "objective/train/weighted_lm_loss": 1.6399933099746704,
+      "objective/train/weights_max": 1.1952372789382935,
+      "objective/train/weights_min": 0.3681386411190033,
+      "theoretical_loss": 3.6446871906940466,
+      "tokens_seen": 1101004800
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006730861819932596,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6446871906940466,
+      "tokens_seen": 1101004800
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006730059380516771,
+      "loss": 0.072,
+      "theoretical_loss": 3.6446067147105197,
+      "tokens_seen": 1101266944
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006729256941100947,
+      "loss": 0.0712,
+      "theoretical_loss": 3.644526263243433,
+      "tokens_seen": 1101529088
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006728454501685123,
+      "loss": 0.0749,
+      "theoretical_loss": 3.644445836279488,
+      "tokens_seen": 1101791232
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006727652062269298,
+      "loss": 0.0697,
+      "theoretical_loss": 3.644365433805393,
+      "tokens_seen": 1102053376
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006726849622853474,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6442850558078685,
+      "tokens_seen": 1102315520
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000672604718343765,
+      "loss": 0.0766,
+      "theoretical_loss": 3.6442047022736452,
+      "tokens_seen": 1102577664
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006725244744021827,
+      "loss": 0.0758,
+      "theoretical_loss": 3.644124373189464,
+      "tokens_seen": 1102839808
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006724442304606002,
+      "loss": 0.0745,
+      "theoretical_loss": 3.644044068542076,
+      "tokens_seen": 1103101952
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006723639865190179,
+      "loss": 0.0728,
+      "theoretical_loss": 3.643963788318242,
+      "tokens_seen": 1103364096
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006722837425774354,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6438835325047356,
+      "tokens_seen": 1103626240
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006722034986358531,
+      "loss": 0.073,
+      "theoretical_loss": 3.6438033010883375,
+      "tokens_seen": 1103888384
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006721232546942706,
+      "loss": 0.0745,
+      "theoretical_loss": 3.643723094055841,
+      "tokens_seen": 1104150528
+    },
+    {
+      "epoch": 0.33,
+      "objective/train/advantage_avg": -0.0009628982516005635,
+      "objective/train/docs_used": 404699,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5864622592926025,
+      "objective/train/original_loss": 1.5864622592926025,
+      "objective/train/theoretical_loss": 3.643682999679431,
+      "objective/train/tokens_used": 1124741600,
+      "objective/train/value_avg": -0.00875091552734375,
+      "objective/train/value_loss": 0.0005373280146159232,
+      "objective/train/value_max": -6.502866744995117e-05,
+      "objective/train/value_min": -0.462158203125,
+      "objective/train/value_reward_corr": 0.6116719535059587,
+      "objective/train/value_std": 0.014617919921875,
+      "objective/train/weight_avg": 0.9992680549621582,
+      "objective/train/weighted_lm_loss": 1.5853101015090942,
+      "objective/train/weights_max": 1.250488042831421,
+      "objective/train/weights_min": 0.3689074218273163,
+      "theoretical_loss": 3.643682999679431,
+      "tokens_seen": 1104281600
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006720430107526881,
+      "loss": 0.0769,
+      "theoretical_loss": 3.643642911394048,
+      "tokens_seen": 1104412672
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006719627668111058,
+      "loss": 0.0735,
+      "theoretical_loss": 3.643562753089772,
+      "tokens_seen": 1104674816
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.0006718825228695233,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6434826191298364,
+      "tokens_seen": 1104936960
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 0.000671802278927941,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6434025095010747,
+      "tokens_seen": 1105199104
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006717220349863586,
+      "loss": 0.0764,
+      "theoretical_loss": 3.6433224241903304,
+      "tokens_seen": 1105461248
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006716417910447762,
+      "loss": 0.0749,
+      "theoretical_loss": 3.643242363184458,
+      "tokens_seen": 1105723392
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006715615471031937,
+      "loss": 0.074,
+      "theoretical_loss": 3.6431623264703212,
+      "tokens_seen": 1105985536
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006714813031616113,
+      "loss": 0.071,
+      "theoretical_loss": 3.6430823140347943,
+      "tokens_seen": 1106247680
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006714010592200289,
+      "loss": 0.0727,
+      "theoretical_loss": 3.643002325864763,
+      "tokens_seen": 1106509824
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006713208152784464,
+      "loss": 0.0727,
+      "theoretical_loss": 3.6429223619471207,
+      "tokens_seen": 1106771968
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006712405713368641,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6428424222687736,
+      "tokens_seen": 1107034112
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006711603273952816,
+      "loss": 0.0739,
+      "theoretical_loss": 3.642762506816636,
+      "tokens_seen": 1107296256
+    },
+    {
+      "epoch": 0.34,
+      "objective/train/advantage_avg": -0.0005870533641427755,
+      "objective/train/docs_used": 405946,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4835734367370605,
+      "objective/train/original_loss": 1.4835731983184814,
+      "objective/train/theoretical_loss": 3.642682615577634,
+      "objective/train/tokens_used": 1128018400,
+      "objective/train/value_avg": -0.007633209228515625,
+      "objective/train/value_loss": 0.00033086538314819336,
+      "objective/train/value_max": -3.916025161743164e-05,
+      "objective/train/value_min": -0.341796875,
+      "objective/train/value_reward_corr": 0.5667257952888192,
+      "objective/train/value_std": 0.01132965087890625,
+      "objective/train/weight_avg": 0.999566912651062,
+      "objective/train/weighted_lm_loss": 1.4834892749786377,
+      "objective/train/weights_max": 1.2339105606079102,
+      "objective/train/weights_min": 0.6143267154693604,
+      "theoretical_loss": 3.642682615577634,
+      "tokens_seen": 1107558400
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006710800834536994,
+      "loss": 0.0717,
+      "theoretical_loss": 3.642682615577634,
+      "tokens_seen": 1107558400
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006709998395121169,
+      "loss": 0.0757,
+      "theoretical_loss": 3.6426027485387023,
+      "tokens_seen": 1107820544
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006709195955705344,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6425229056867865,
+      "tokens_seen": 1108082688
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006708393516289521,
+      "loss": 0.0723,
+      "theoretical_loss": 3.642443087008844,
+      "tokens_seen": 1108344832
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006707591076873696,
+      "loss": 0.0718,
+      "theoretical_loss": 3.6423632924918383,
+      "tokens_seen": 1108606976
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006706788637457872,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6422835221227468,
+      "tokens_seen": 1108869120
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006705986198042048,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6422037758885555,
+      "tokens_seen": 1109131264
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006705183758626224,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6421240537762607,
+      "tokens_seen": 1109393408
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006704381319210399,
+      "loss": 0.0774,
+      "theoretical_loss": 3.6420443557728674,
+      "tokens_seen": 1109655552
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006703578879794575,
+      "loss": 0.0742,
+      "theoretical_loss": 3.6419646818653932,
+      "tokens_seen": 1109917696
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006702776440378752,
+      "loss": 0.076,
+      "theoretical_loss": 3.641885032040864,
+      "tokens_seen": 1110179840
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006701974000962927,
+      "loss": 0.0751,
+      "theoretical_loss": 3.6418054062863163,
+      "tokens_seen": 1110441984
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006701171561547104,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6417258045887966,
+      "tokens_seen": 1110704128
+    },
+    {
+      "epoch": 0.34,
+      "objective/train/advantage_avg": -0.0009894539834931493,
+      "objective/train/docs_used": 406993,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.475312352180481,
+      "objective/train/original_loss": 1.4753124713897705,
+      "objective/train/theoretical_loss": 3.6416860127573765,
+      "objective/train/tokens_used": 1131295200,
+      "objective/train/value_avg": -0.00717926025390625,
+      "objective/train/value_loss": 0.00043393525993451476,
+      "objective/train/value_max": -3.069639205932617e-05,
+      "objective/train/value_min": -0.8212890625,
+      "objective/train/value_reward_corr": 0.6151101633973011,
+      "objective/train/value_std": 0.01456451416015625,
+      "objective/train/weight_avg": 0.999183714389801,
+      "objective/train/weighted_lm_loss": 1.4731035232543945,
+      "objective/train/weights_max": 1.11464262008667,
+      "objective/train/weights_min": 0.22452190518379211,
+      "theoretical_loss": 3.6416860127573765,
+      "tokens_seen": 1110835200
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006700369122131279,
+      "loss": 0.0746,
+      "theoretical_loss": 3.641646226935361,
+      "tokens_seen": 1110966272
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006699566682715455,
+      "loss": 0.076,
+      "theoretical_loss": 3.641566673313076,
+      "tokens_seen": 1111228416
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006698764243299631,
+      "loss": 0.0776,
+      "theoretical_loss": 3.6414871437090186,
+      "tokens_seen": 1111490560
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006697961803883806,
+      "loss": 0.0752,
+      "theoretical_loss": 3.641407638110275,
+      "tokens_seen": 1111752704
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006697159364467983,
+      "loss": 0.0755,
+      "theoretical_loss": 3.641328156503942,
+      "tokens_seen": 1112014848
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006696356925052158,
+      "loss": 0.0757,
+      "theoretical_loss": 3.6412486988771255,
+      "tokens_seen": 1112276992
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006695554485636335,
+      "loss": 0.0727,
+      "theoretical_loss": 3.6411692652169423,
+      "tokens_seen": 1112539136
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006694752046220511,
+      "loss": 0.0749,
+      "theoretical_loss": 3.641089855510518,
+      "tokens_seen": 1112801280
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006693949606804687,
+      "loss": 0.0726,
+      "theoretical_loss": 3.64101046974499,
+      "tokens_seen": 1113063424
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006693147167388862,
+      "loss": 0.074,
+      "theoretical_loss": 3.640931107907504,
+      "tokens_seen": 1113325568
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006692344727973039,
+      "loss": 0.0768,
+      "theoretical_loss": 3.6408517699852165,
+      "tokens_seen": 1113587712
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006691542288557214,
+      "loss": 0.078,
+      "theoretical_loss": 3.640772455965293,
+      "tokens_seen": 1113849856
+    },
+    {
+      "epoch": 0.34,
+      "objective/train/advantage_avg": 0.0012157823657616973,
+      "objective/train/docs_used": 408181,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.420784592628479,
+      "objective/train/original_loss": 1.4207844734191895,
+      "objective/train/theoretical_loss": 3.64069316583491,
+      "objective/train/tokens_used": 1134572000,
+      "objective/train/value_avg": -0.0094451904296875,
+      "objective/train/value_loss": 0.00029014694155193865,
+      "objective/train/value_max": -9.387731552124023e-05,
+      "objective/train/value_min": -0.5224609375,
+      "objective/train/value_reward_corr": 0.6501879473840528,
+      "objective/train/value_std": 0.016510009765625,
+      "objective/train/weight_avg": 1.0013483762741089,
+      "objective/train/weighted_lm_loss": 1.4223439693450928,
+      "objective/train/weights_max": 1.2760461568832397,
+      "objective/train/weights_min": 0.36923545598983765,
+      "theoretical_loss": 3.64069316583491,
+      "tokens_seen": 1114112000
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006690739849141389,
+      "loss": 0.0726,
+      "theoretical_loss": 3.64069316583491,
+      "tokens_seen": 1114112000
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006689937409725566,
+      "loss": 0.0752,
+      "theoretical_loss": 3.640613899581253,
+      "tokens_seen": 1114374144
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006689134970309741,
+      "loss": 0.0777,
+      "theoretical_loss": 3.6405346571915187,
+      "tokens_seen": 1114636288
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006688332530893918,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6404554386529115,
+      "tokens_seen": 1114898432
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006687530091478094,
+      "loss": 0.072,
+      "theoretical_loss": 3.640376243952648,
+      "tokens_seen": 1115160576
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000668672765206227,
+      "loss": 0.072,
+      "theoretical_loss": 3.640297073077953,
+      "tokens_seen": 1115422720
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006685925212646445,
+      "loss": 0.0779,
+      "theoretical_loss": 3.640217926016061,
+      "tokens_seen": 1115684864
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006685122773230621,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6401388027542185,
+      "tokens_seen": 1115947008
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006684320333814797,
+      "loss": 0.0708,
+      "theoretical_loss": 3.6400597032796798,
+      "tokens_seen": 1116209152
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006683517894398973,
+      "loss": 0.0753,
+      "theoretical_loss": 3.6399806275797095,
+      "tokens_seen": 1116471296
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006682715454983149,
+      "loss": 0.0761,
+      "theoretical_loss": 3.639901575641582,
+      "tokens_seen": 1116733440
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006681913015567324,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6398225474525816,
+      "tokens_seen": 1116995584
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006681110576151502,
+      "loss": 0.0726,
+      "theoretical_loss": 3.639743543000003,
+      "tokens_seen": 1117257728
+    },
+    {
+      "epoch": 0.34,
+      "objective/train/advantage_avg": 0.000741979107260704,
+      "objective/train/docs_used": 409409,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5466253757476807,
+      "objective/train/original_loss": 1.5466253757476807,
+      "objective/train/theoretical_loss": 3.639704049670904,
+      "objective/train/tokens_used": 1137848800,
+      "objective/train/value_avg": -0.00763702392578125,
+      "objective/train/value_loss": 0.0003765631699934602,
+      "objective/train/value_max": -5.8770179748535156e-05,
+      "objective/train/value_min": -0.3564453125,
+      "objective/train/value_reward_corr": 0.5197531699733359,
+      "objective/train/value_std": 0.01153564453125,
+      "objective/train/weight_avg": 1.0008995532989502,
+      "objective/train/weighted_lm_loss": 1.5469353199005127,
+      "objective/train/weights_max": 1.3829063177108765,
+      "objective/train/weights_min": 0.22673387825489044,
+      "theoretical_loss": 3.639704049670904,
+      "tokens_seen": 1117388800
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006680308136735677,
+      "loss": 0.073,
+      "theoretical_loss": 3.6396645622711494,
+      "tokens_seen": 1117519872
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006679505697319852,
+      "loss": 0.0764,
+      "theoretical_loss": 3.639585605253335,
+      "tokens_seen": 1117782016
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006678703257904029,
+      "loss": 0.0738,
+      "theoretical_loss": 3.639506671933882,
+      "tokens_seen": 1118044160
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006677900818488204,
+      "loss": 0.0739,
+      "theoretical_loss": 3.639427762300125,
+      "tokens_seen": 1118306304
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000667709837907238,
+      "loss": 0.0748,
+      "theoretical_loss": 3.6393488763394064,
+      "tokens_seen": 1118568448
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006676295939656556,
+      "loss": 0.0738,
+      "theoretical_loss": 3.639270014039078,
+      "tokens_seen": 1118830592
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006675493500240732,
+      "loss": 0.074,
+      "theoretical_loss": 3.6391911753865034,
+      "tokens_seen": 1119092736
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006674691060824907,
+      "loss": 0.0754,
+      "theoretical_loss": 3.639112360369054,
+      "tokens_seen": 1119354880
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006673888621409083,
+      "loss": 0.0763,
+      "theoretical_loss": 3.6390335689741113,
+      "tokens_seen": 1119617024
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000667308618199326,
+      "loss": 0.075,
+      "theoretical_loss": 3.638954801189067,
+      "tokens_seen": 1119879168
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006672283742577436,
+      "loss": 0.076,
+      "theoretical_loss": 3.6388760570013226,
+      "tokens_seen": 1120141312
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006671481303161612,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6387973363982877,
+      "tokens_seen": 1120403456
+    },
+    {
+      "epoch": 0.34,
+      "objective/train/advantage_avg": 0.0013283737935125828,
+      "objective/train/docs_used": 410712,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4400638341903687,
+      "objective/train/original_loss": 1.440063714981079,
+      "objective/train/theoretical_loss": 3.638718639367384,
+      "objective/train/tokens_used": 1141125600,
+      "objective/train/value_avg": -0.007556915283203125,
+      "objective/train/value_loss": 0.0001608040911378339,
+      "objective/train/value_max": -6.300210952758789e-05,
+      "objective/train/value_min": -0.611328125,
+      "objective/train/value_reward_corr": 0.7982278408915517,
+      "objective/train/value_std": 0.01517486572265625,
+      "objective/train/weight_avg": 1.0014057159423828,
+      "objective/train/weighted_lm_loss": 1.4431976079940796,
+      "objective/train/weights_max": 1.3930864334106445,
+      "objective/train/weights_min": 0.5440833568572998,
+      "theoretical_loss": 3.638718639367384,
+      "tokens_seen": 1120665600
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006670678863745787,
+      "loss": 0.0728,
+      "theoretical_loss": 3.638718639367384,
+      "tokens_seen": 1120665600
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006669876424329964,
+      "loss": 0.0737,
+      "theoretical_loss": 3.638639965896041,
+      "tokens_seen": 1120927744
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006669073984914139,
+      "loss": 0.0766,
+      "theoretical_loss": 3.638561315971698,
+      "tokens_seen": 1121189888
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006668271545498314,
+      "loss": 0.0733,
+      "theoretical_loss": 3.638482689581805,
+      "tokens_seen": 1121452032
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006667469106082491,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6384040867138214,
+      "tokens_seen": 1121714176
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006666666666666666,
+      "loss": 0.0765,
+      "theoretical_loss": 3.6383255073552148,
+      "tokens_seen": 1121976320
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006665864227250843,
+      "loss": 0.0754,
+      "theoretical_loss": 3.638246951493463,
+      "tokens_seen": 1122238464
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006665061787835019,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6381684191160555,
+      "tokens_seen": 1122500608
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006664259348419195,
+      "loss": 0.0726,
+      "theoretical_loss": 3.638089910210488,
+      "tokens_seen": 1122762752
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000666345690900337,
+      "loss": 0.0762,
+      "theoretical_loss": 3.638011424764269,
+      "tokens_seen": 1123024896
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006662654469587546,
+      "loss": 0.0711,
+      "theoretical_loss": 3.6379329627649137,
+      "tokens_seen": 1123287040
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006661852030171722,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6378545241999487,
+      "tokens_seen": 1123549184
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006661049590755897,
+      "loss": 0.0745,
+      "theoretical_loss": 3.637776109056909,
+      "tokens_seen": 1123811328
+    },
+    {
+      "epoch": 0.34,
+      "objective/train/advantage_avg": -0.0026691153179854155,
+      "objective/train/docs_used": 411904,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2471657991409302,
+      "objective/train/original_loss": 1.2471659183502197,
+      "objective/train/theoretical_loss": 3.637736910264719,
+      "objective/train/tokens_used": 1144402400,
+      "objective/train/value_avg": -0.00782012939453125,
+      "objective/train/value_loss": 0.0007978877983987331,
+      "objective/train/value_max": -9.459257125854492e-05,
+      "objective/train/value_min": -0.71435546875,
+      "objective/train/value_reward_corr": 0.6231384912903344,
+      "objective/train/value_std": 0.0173797607421875,
+      "objective/train/weight_avg": 0.9976327419281006,
+      "objective/train/weighted_lm_loss": 1.243854284286499,
+      "objective/train/weights_max": 2.0428695678710938,
+      "objective/train/weights_min": 0.05494469031691551,
+      "theoretical_loss": 3.637736910264719,
+      "tokens_seen": 1123942400
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006660247151340074,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6376977173233405,
+      "tokens_seen": 1124073472
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006659444711924249,
+      "loss": 0.073,
+      "theoretical_loss": 3.6376193489867976,
+      "tokens_seen": 1124335616
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006658642272508427,
+      "loss": 0.0757,
+      "theoretical_loss": 3.6375410040348446,
+      "tokens_seen": 1124597760
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006657839833092602,
+      "loss": 0.0712,
+      "theoretical_loss": 3.637462682455055,
+      "tokens_seen": 1124859904
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006657037393676778,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6373843842350118,
+      "tokens_seen": 1125122048
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006656234954260954,
+      "loss": 0.0752,
+      "theoretical_loss": 3.637306109362308,
+      "tokens_seen": 1125384192
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006655432514845129,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6372278578245454,
+      "tokens_seen": 1125646336
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006654630075429305,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6371496296093357,
+      "tokens_seen": 1125908480
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006653827636013481,
+      "loss": 0.0748,
+      "theoretical_loss": 3.6370714247043003,
+      "tokens_seen": 1126170624
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006653025196597657,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6369932430970695,
+      "tokens_seen": 1126432768
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006652222757181832,
+      "loss": 0.0772,
+      "theoretical_loss": 3.6369150847752834,
+      "tokens_seen": 1126694912
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000665142031776601,
+      "loss": 0.0748,
+      "theoretical_loss": 3.6368369497265913,
+      "tokens_seen": 1126957056
+    },
+    {
+      "epoch": 0.34,
+      "objective/train/advantage_avg": -0.004586232826113701,
+      "objective/train/docs_used": 413086,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5389773845672607,
+      "objective/train/original_loss": 1.5389773845672607,
+      "objective/train/theoretical_loss": 3.6367588379386513,
+      "objective/train/tokens_used": 1147679200,
+      "objective/train/value_avg": -0.01276397705078125,
+      "objective/train/value_loss": 0.001206457382068038,
+      "objective/train/value_max": -8.285045623779297e-05,
+      "objective/train/value_min": -0.6474609375,
+      "objective/train/value_reward_corr": 0.7707702012782308,
+      "objective/train/value_std": 0.0240020751953125,
+      "objective/train/weight_avg": 0.9959734678268433,
+      "objective/train/weighted_lm_loss": 1.5339165925979614,
+      "objective/train/weights_max": 1.6746848821640015,
+      "objective/train/weights_min": 0.36978310346603394,
+      "theoretical_loss": 3.6367588379386513,
+      "tokens_seen": 1127219200
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006650617878350185,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6367588379386513,
+      "tokens_seen": 1127219200
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000664981543893436,
+      "loss": 0.0756,
+      "theoretical_loss": 3.636680749399133,
+      "tokens_seen": 1127481344
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006649012999518537,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6366026840957133,
+      "tokens_seen": 1127743488
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006648210560102712,
+      "loss": 0.076,
+      "theoretical_loss": 3.636524642016079,
+      "tokens_seen": 1128005632
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006647408120686889,
+      "loss": 0.0745,
+      "theoretical_loss": 3.636446623147927,
+      "tokens_seen": 1128267776
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006646605681271064,
+      "loss": 0.0763,
+      "theoretical_loss": 3.6363686274789626,
+      "tokens_seen": 1128529920
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000664580324185524,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6362906549969014,
+      "tokens_seen": 1128792064
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006645000802439416,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6362127056894673,
+      "tokens_seen": 1129054208
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006644198363023591,
+      "loss": 0.0753,
+      "theoretical_loss": 3.6361347795443955,
+      "tokens_seen": 1129316352
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006643395923607768,
+      "loss": 0.0739,
+      "theoretical_loss": 3.636056876549427,
+      "tokens_seen": 1129578496
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006642593484191944,
+      "loss": 0.074,
+      "theoretical_loss": 3.6359789966923164,
+      "tokens_seen": 1129840640
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000664179104477612,
+      "loss": 0.0773,
+      "theoretical_loss": 3.6359011399608243,
+      "tokens_seen": 1130102784
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006640988605360295,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6358233063427225,
+      "tokens_seen": 1130364928
+    },
+    {
+      "epoch": 0.34,
+      "objective/train/advantage_avg": 0.0018270047148689628,
+      "objective/train/docs_used": 414307,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.627175211906433,
+      "objective/train/original_loss": 1.6271753311157227,
+      "objective/train/theoretical_loss": 3.635784398197374,
+      "objective/train/tokens_used": 1150956000,
+      "objective/train/value_avg": -0.00772857666015625,
+      "objective/train/value_loss": 0.0002652402617968619,
+      "objective/train/value_max": -0.00010472536087036133,
+      "objective/train/value_min": -0.96875,
+      "objective/train/value_reward_corr": 0.6389814979319877,
+      "objective/train/value_std": 0.01456451416015625,
+      "objective/train/weight_avg": 1.0019452571868896,
+      "objective/train/weighted_lm_loss": 1.630281686782837,
+      "objective/train/weights_max": 1.5185024738311768,
+      "objective/train/weights_min": 0.3883684575557709,
+      "theoretical_loss": 3.635784398197374,
+      "tokens_seen": 1130496000
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006640186165944472,
+      "loss": 0.0716,
+      "theoretical_loss": 3.635745495825791,
+      "tokens_seen": 1130627072
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006639383726528647,
+      "loss": 0.0721,
+      "theoretical_loss": 3.63566770839782,
+      "tokens_seen": 1130889216
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006638581287112822,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6355899440466075,
+      "tokens_seen": 1131151360
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006637778847696999,
+      "loss": 0.0731,
+      "theoretical_loss": 3.635512202759964,
+      "tokens_seen": 1131413504
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006636976408281174,
+      "loss": 0.0738,
+      "theoretical_loss": 3.635434484525704,
+      "tokens_seen": 1131675648
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006636173968865351,
+      "loss": 0.076,
+      "theoretical_loss": 3.6353567893316567,
+      "tokens_seen": 1131937792
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006635371529449527,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6352791171656573,
+      "tokens_seen": 1132199936
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006634569090033703,
+      "loss": 0.0743,
+      "theoretical_loss": 3.635201468015551,
+      "tokens_seen": 1132462080
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006633766650617879,
+      "loss": 0.074,
+      "theoretical_loss": 3.635123841869193,
+      "tokens_seen": 1132724224
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006632964211202054,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6350462387144464,
+      "tokens_seen": 1132986368
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000663216177178623,
+      "loss": 0.0748,
+      "theoretical_loss": 3.634968658539184,
+      "tokens_seen": 1133248512
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006631359332370406,
+      "loss": 0.0768,
+      "theoretical_loss": 3.6348911013312883,
+      "tokens_seen": 1133510656
+    },
+    {
+      "epoch": 0.34,
+      "objective/train/advantage_avg": 0.0008693314739502966,
+      "objective/train/docs_used": 415442,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.477584958076477,
+      "objective/train/original_loss": 1.477584958076477,
+      "objective/train/theoretical_loss": 3.6348135670786506,
+      "objective/train/tokens_used": 1154232800,
+      "objective/train/value_avg": -0.007232666015625,
+      "objective/train/value_loss": 0.00015491771046072245,
+      "objective/train/value_max": -9.685754776000977e-05,
+      "objective/train/value_min": -0.66357421875,
+      "objective/train/value_reward_corr": 0.5982827764776568,
+      "objective/train/value_std": 0.01332855224609375,
+      "objective/train/weight_avg": 1.0009486675262451,
+      "objective/train/weighted_lm_loss": 1.47885262966156,
+      "objective/train/weights_max": 1.8655112981796265,
+      "objective/train/weights_min": 0.6109683513641357,
+      "theoretical_loss": 3.6348135670786506,
+      "tokens_seen": 1133772800
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006630556892954582,
+      "loss": 0.0736,
+      "theoretical_loss": 3.6348135670786506,
+      "tokens_seen": 1133772800
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006629754453538757,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6347360557691712,
+      "tokens_seen": 1134034944
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006628952014122935,
+      "loss": 0.0739,
+      "theoretical_loss": 3.63465856739076,
+      "tokens_seen": 1134297088
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000662814957470711,
+      "loss": 0.0747,
+      "theoretical_loss": 3.634581101931336,
+      "tokens_seen": 1134559232
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006627347135291286,
+      "loss": 0.0753,
+      "theoretical_loss": 3.6345036593788276,
+      "tokens_seen": 1134821376
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006626544695875462,
+      "loss": 0.0758,
+      "theoretical_loss": 3.6344262397211704,
+      "tokens_seen": 1135083520
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006625742256459637,
+      "loss": 0.0754,
+      "theoretical_loss": 3.6343488429463124,
+      "tokens_seen": 1135345664
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006624939817043813,
+      "loss": 0.076,
+      "theoretical_loss": 3.634271469042208,
+      "tokens_seen": 1135607808
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006624137377627989,
+      "loss": 0.0728,
+      "theoretical_loss": 3.634194117996822,
+      "tokens_seen": 1135869952
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006623334938212165,
+      "loss": 0.0736,
+      "theoretical_loss": 3.634116789798129,
+      "tokens_seen": 1136132096
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000662253249879634,
+      "loss": 0.077,
+      "theoretical_loss": 3.6340394844341097,
+      "tokens_seen": 1136394240
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006621730059380518,
+      "loss": 0.0771,
+      "theoretical_loss": 3.6339622018927575,
+      "tokens_seen": 1136656384
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006620927619964693,
+      "loss": 0.0741,
+      "theoretical_loss": 3.633884942162073,
+      "tokens_seen": 1136918528
+    },
+    {
+      "epoch": 0.34,
+      "objective/train/advantage_avg": 0.0013180155074223876,
+      "objective/train/docs_used": 416597,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.555419683456421,
+      "objective/train/original_loss": 1.555419683456421,
+      "objective/train/theoretical_loss": 3.633846320846984,
+      "objective/train/tokens_used": 1157509600,
+      "objective/train/value_avg": -0.006999969482421875,
+      "objective/train/value_loss": 0.00019526074174791574,
+      "objective/train/value_max": -0.00011771917343139648,
+      "objective/train/value_min": -0.3037109375,
+      "objective/train/value_reward_corr": 0.5462273072896175,
+      "objective/train/value_std": 0.009796142578125,
+      "objective/train/weight_avg": 1.0014064311981201,
+      "objective/train/weighted_lm_loss": 1.5573372840881348,
+      "objective/train/weights_max": 1.306922197341919,
+      "objective/train/weights_min": 0.36874493956565857,
+      "theoretical_loss": 3.633846320846984,
+      "tokens_seen": 1137049600
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006620125180548869,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6338077052300664,
+      "tokens_seen": 1137180672
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006619322741133045,
+      "loss": 0.0725,
+      "theoretical_loss": 3.633730491084756,
+      "tokens_seen": 1137442816
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.000661852030171722,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6336532997141706,
+      "tokens_seen": 1137704960
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006617717862301397,
+      "loss": 0.0781,
+      "theoretical_loss": 3.6335761311063473,
+      "tokens_seen": 1137967104
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 0.0006616915422885572,
+      "loss": 0.0764,
+      "theoretical_loss": 3.633498985249332,
+      "tokens_seen": 1138229248
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006616112983469748,
+      "loss": 0.074,
+      "theoretical_loss": 3.63342186213118,
+      "tokens_seen": 1138491392
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006615310544053924,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6333447617399557,
+      "tokens_seen": 1138753536
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.00066145081046381,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6332676840637324,
+      "tokens_seen": 1139015680
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006613705665222276,
+      "loss": 0.076,
+      "theoretical_loss": 3.633190629090592,
+      "tokens_seen": 1139277824
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006612903225806452,
+      "loss": 0.0751,
+      "theoretical_loss": 3.6331135968086263,
+      "tokens_seen": 1139539968
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006612100786390628,
+      "loss": 0.0739,
+      "theoretical_loss": 3.633036587205935,
+      "tokens_seen": 1139802112
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006611298346974803,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6329596002706275,
+      "tokens_seen": 1140064256
+    },
+    {
+      "epoch": 0.35,
+      "objective/train/advantage_avg": 0.0006165788508951664,
+      "objective/train/docs_used": 417664,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4508864879608154,
+      "objective/train/original_loss": 1.4508862495422363,
+      "objective/train/theoretical_loss": 3.632882635990822,
+      "objective/train/tokens_used": 1160786400,
+      "objective/train/value_avg": -0.006664276123046875,
+      "objective/train/value_loss": 0.00015237460320349783,
+      "objective/train/value_max": -8.285045623779297e-05,
+      "objective/train/value_min": -0.4951171875,
+      "objective/train/value_reward_corr": 0.7813708563816653,
+      "objective/train/value_std": 0.0169219970703125,
+      "objective/train/weight_avg": 1.000691533088684,
+      "objective/train/weighted_lm_loss": 1.4525622129440308,
+      "objective/train/weights_max": 1.1556388139724731,
+      "objective/train/weights_min": 0.6773775815963745,
+      "theoretical_loss": 3.632882635990822,
+      "tokens_seen": 1140326400
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000661049590755898,
+      "loss": 0.0745,
+      "theoretical_loss": 3.632882635990822,
+      "tokens_seen": 1140326400
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006609693468143155,
+      "loss": 0.0764,
+      "theoretical_loss": 3.632805694354646,
+      "tokens_seen": 1140588544
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006608891028727331,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6327287753502358,
+      "tokens_seen": 1140850688
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006608088589311507,
+      "loss": 0.0753,
+      "theoretical_loss": 3.632651878965735,
+      "tokens_seen": 1141112832
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006607286149895682,
+      "loss": 0.0734,
+      "theoretical_loss": 3.632575005189299,
+      "tokens_seen": 1141374976
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000660648371047986,
+      "loss": 0.0765,
+      "theoretical_loss": 3.6324981540090895,
+      "tokens_seen": 1141637120
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006605681271064035,
+      "loss": 0.074,
+      "theoretical_loss": 3.6324213254132793,
+      "tokens_seen": 1141899264
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006604878831648211,
+      "loss": 0.0758,
+      "theoretical_loss": 3.632344519390049,
+      "tokens_seen": 1142161408
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006604076392232387,
+      "loss": 0.0726,
+      "theoretical_loss": 3.632267735927588,
+      "tokens_seen": 1142423552
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006603273952816562,
+      "loss": 0.0747,
+      "theoretical_loss": 3.632190975014094,
+      "tokens_seen": 1142685696
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006602471513400738,
+      "loss": 0.0723,
+      "theoretical_loss": 3.6321142366377757,
+      "tokens_seen": 1142947840
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006601669073984914,
+      "loss": 0.072,
+      "theoretical_loss": 3.6320375207868483,
+      "tokens_seen": 1143209984
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000660086663456909,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6319608274495376,
+      "tokens_seen": 1143472128
+    },
+    {
+      "epoch": 0.35,
+      "objective/train/advantage_avg": 0.002307515824213624,
+      "objective/train/docs_used": 418760,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4072736501693726,
+      "objective/train/original_loss": 1.407273769378662,
+      "objective/train/theoretical_loss": 3.6319224892198108,
+      "objective/train/tokens_used": 1164063200,
+      "objective/train/value_avg": -0.00730133056640625,
+      "objective/train/value_loss": 0.00011779867054428905,
+      "objective/train/value_max": -0.00014889240264892578,
+      "objective/train/value_min": -0.4365234375,
+      "objective/train/value_reward_corr": 0.7994153780614749,
+      "objective/train/value_std": 0.01413726806640625,
+      "objective/train/weight_avg": 1.0023618936538696,
+      "objective/train/weighted_lm_loss": 1.4107744693756104,
+      "objective/train/weights_max": 1.1094887256622314,
+      "objective/train/weights_min": 0.37357577681541443,
+      "theoretical_loss": 3.6319224892198108,
+      "tokens_seen": 1143603200
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006600064195153265,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6318841566140767,
+      "tokens_seen": 1143734272
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006599261755737443,
+      "loss": 0.0724,
+      "theoretical_loss": 3.63180750826871,
+      "tokens_seen": 1143996416
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006598459316321618,
+      "loss": 0.0701,
+      "theoretical_loss": 3.6317308824016874,
+      "tokens_seen": 1144258560
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006597656876905794,
+      "loss": 0.0707,
+      "theoretical_loss": 3.6316542790012702,
+      "tokens_seen": 1144520704
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000659685443748997,
+      "loss": 0.0718,
+      "theoretical_loss": 3.631577698055727,
+      "tokens_seen": 1144782848
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006596051998074145,
+      "loss": 0.0708,
+      "theoretical_loss": 3.631501139553337,
+      "tokens_seen": 1145044992
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006595249558658322,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6314246034823867,
+      "tokens_seen": 1145307136
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006594447119242497,
+      "loss": 0.0751,
+      "theoretical_loss": 3.631348089831171,
+      "tokens_seen": 1145569280
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006593644679826673,
+      "loss": 0.0737,
+      "theoretical_loss": 3.631271598587995,
+      "tokens_seen": 1145831424
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006592842240410849,
+      "loss": 0.0741,
+      "theoretical_loss": 3.631195129741172,
+      "tokens_seen": 1146093568
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006592039800995026,
+      "loss": 0.0749,
+      "theoretical_loss": 3.631118683279024,
+      "tokens_seen": 1146355712
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006591237361579201,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6310422591898814,
+      "tokens_seen": 1146617856
+    },
+    {
+      "epoch": 0.35,
+      "objective/train/advantage_avg": 0.001571335014887154,
+      "objective/train/docs_used": 419954,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4320921897888184,
+      "objective/train/original_loss": 1.4320919513702393,
+      "objective/train/theoretical_loss": 3.630965857462084,
+      "objective/train/tokens_used": 1167340000,
+      "objective/train/value_avg": -0.00972747802734375,
+      "objective/train/value_loss": 0.00031021423637866974,
+      "objective/train/value_max": -0.00012934207916259766,
+      "objective/train/value_min": -0.625,
+      "objective/train/value_reward_corr": 0.6921582137917927,
+      "objective/train/value_std": 0.0173187255859375,
+      "objective/train/weight_avg": 1.0017169713974,
+      "objective/train/weighted_lm_loss": 1.4336016178131104,
+      "objective/train/weights_max": 1.7058966159820557,
+      "objective/train/weights_min": 0.5872719287872314,
+      "theoretical_loss": 3.630965857462084,
+      "tokens_seen": 1146880000
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006590434922163377,
+      "loss": 0.0716,
+      "theoretical_loss": 3.630965857462084,
+      "tokens_seen": 1146880000
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006589632482747553,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6308894780839798,
+      "tokens_seen": 1147142144
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006588830043331728,
+      "loss": 0.0716,
+      "theoretical_loss": 3.630813121043926,
+      "tokens_seen": 1147404288
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006588027603915905,
+      "loss": 0.0737,
+      "theoretical_loss": 3.630736786330288,
+      "tokens_seen": 1147666432
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000658722516450008,
+      "loss": 0.0746,
+      "theoretical_loss": 3.630660473931441,
+      "tokens_seen": 1147928576
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006586422725084256,
+      "loss": 0.0756,
+      "theoretical_loss": 3.6305841838357673,
+      "tokens_seen": 1148190720
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006585620285668432,
+      "loss": 0.0706,
+      "theoretical_loss": 3.630507916031659,
+      "tokens_seen": 1148452864
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006584817846252607,
+      "loss": 0.0745,
+      "theoretical_loss": 3.630431670507517,
+      "tokens_seen": 1148715008
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006584015406836785,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6303554472517496,
+      "tokens_seen": 1148977152
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000658321296742096,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6302792462527758,
+      "tokens_seen": 1149239296
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006582410528005136,
+      "loss": 0.071,
+      "theoretical_loss": 3.6302030674990213,
+      "tokens_seen": 1149501440
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006581608088589312,
+      "loss": 0.075,
+      "theoretical_loss": 3.6301269109789214,
+      "tokens_seen": 1149763584
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006580805649173488,
+      "loss": 0.0719,
+      "theoretical_loss": 3.63005077668092,
+      "tokens_seen": 1150025728
+    },
+    {
+      "epoch": 0.35,
+      "objective/train/advantage_avg": 0.0009393309592269361,
+      "objective/train/docs_used": 421268,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4447449445724487,
+      "objective/train/original_loss": 1.4447450637817383,
+      "objective/train/theoretical_loss": 3.630012717861597,
+      "objective/train/tokens_used": 1170616800,
+      "objective/train/value_avg": -0.00754547119140625,
+      "objective/train/value_loss": 0.0002241473994217813,
+      "objective/train/value_max": -0.0001767873764038086,
+      "objective/train/value_min": -0.6748046875,
+      "objective/train/value_reward_corr": 0.743822312179771,
+      "objective/train/value_std": 0.0171661376953125,
+      "objective/train/weight_avg": 1.0010449886322021,
+      "objective/train/weighted_lm_loss": 1.4458165168762207,
+      "objective/train/weights_max": 1.4855008125305176,
+      "objective/train/weights_min": 0.5288990139961243,
+      "theoretical_loss": 3.630012717861597,
+      "tokens_seen": 1150156800
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006580003209757663,
+      "loss": 0.0723,
+      "theoretical_loss": 3.62997466459347,
+      "tokens_seen": 1150287872
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006579200770341839,
+      "loss": 0.0705,
+      "theoretical_loss": 3.629898574705031,
+      "tokens_seen": 1150550016
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006578398330926015,
+      "loss": 0.0735,
+      "theoretical_loss": 3.629822507004075,
+      "tokens_seen": 1150812160
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000657759589151019,
+      "loss": 0.0723,
+      "theoretical_loss": 3.629746461479079,
+      "tokens_seen": 1151074304
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006576793452094368,
+      "loss": 0.0714,
+      "theoretical_loss": 3.62967043811853,
+      "tokens_seen": 1151336448
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006575991012678543,
+      "loss": 0.0736,
+      "theoretical_loss": 3.629594436910924,
+      "tokens_seen": 1151598592
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006575188573262719,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6295184578447643,
+      "tokens_seen": 1151860736
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006574386133846895,
+      "loss": 0.0699,
+      "theoretical_loss": 3.6294425009085645,
+      "tokens_seen": 1152122880
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000657358369443107,
+      "loss": 0.073,
+      "theoretical_loss": 3.6293665660908454,
+      "tokens_seen": 1152385024
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006572781255015246,
+      "loss": 0.0718,
+      "theoretical_loss": 3.6292906533801372,
+      "tokens_seen": 1152647168
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006571978815599422,
+      "loss": 0.0742,
+      "theoretical_loss": 3.6292147627649776,
+      "tokens_seen": 1152909312
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006571176376183598,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6291388942339147,
+      "tokens_seen": 1153171456
+    },
+    {
+      "epoch": 0.35,
+      "objective/train/advantage_avg": 0.00047070140135474503,
+      "objective/train/docs_used": 422447,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6008491516113281,
+      "objective/train/original_loss": 1.6008491516113281,
+      "objective/train/theoretical_loss": 3.6290630477755026,
+      "objective/train/tokens_used": 1173893600,
+      "objective/train/value_avg": -0.00966644287109375,
+      "objective/train/value_loss": 0.0003931581450160593,
+      "objective/train/value_max": -9.608268737792969e-05,
+      "objective/train/value_min": -0.392333984375,
+      "objective/train/value_reward_corr": 0.647847413214932,
+      "objective/train/value_std": 0.015899658203125,
+      "objective/train/weight_avg": 1.0006462335586548,
+      "objective/train/weighted_lm_loss": 1.601322054862976,
+      "objective/train/weights_max": 1.4302529096603394,
+      "objective/train/weights_min": 0.3789495527744293,
+      "theoretical_loss": 3.6290630477755026,
+      "tokens_seen": 1153433600
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006570373936767774,
+      "loss": 0.0714,
+      "theoretical_loss": 3.6290630477755026,
+      "tokens_seen": 1153433600
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006569571497351951,
+      "loss": 0.0717,
+      "theoretical_loss": 3.6289872233783065,
+      "tokens_seen": 1153695744
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006568769057936126,
+      "loss": 0.0705,
+      "theoretical_loss": 3.6289114210308977,
+      "tokens_seen": 1153957888
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006567966618520302,
+      "loss": 0.0728,
+      "theoretical_loss": 3.628835640721859,
+      "tokens_seen": 1154220032
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006567164179104478,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6287598824397787,
+      "tokens_seen": 1154482176
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006566361739688653,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6286841461732546,
+      "tokens_seen": 1154744320
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000656555930027283,
+      "loss": 0.0713,
+      "theoretical_loss": 3.6286084319108944,
+      "tokens_seen": 1155006464
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006564756860857005,
+      "loss": 0.0701,
+      "theoretical_loss": 3.628532739641312,
+      "tokens_seen": 1155268608
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006563954421441181,
+      "loss": 0.0699,
+      "theoretical_loss": 3.6284570693531317,
+      "tokens_seen": 1155530752
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006563151982025357,
+      "loss": 0.0717,
+      "theoretical_loss": 3.6283814210349847,
+      "tokens_seen": 1155792896
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006562349542609534,
+      "loss": 0.0709,
+      "theoretical_loss": 3.628305794675512,
+      "tokens_seen": 1156055040
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006561547103193709,
+      "loss": 0.0711,
+      "theoretical_loss": 3.6282301902633627,
+      "tokens_seen": 1156317184
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006560744663777885,
+      "loss": 0.0692,
+      "theoretical_loss": 3.628154607787194,
+      "tokens_seen": 1156579328
+    },
+    {
+      "epoch": 0.35,
+      "objective/train/advantage_avg": 0.000545949034858495,
+      "objective/train/docs_used": 423626,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2599838972091675,
+      "objective/train/original_loss": 1.2599838972091675,
+      "objective/train/theoretical_loss": 3.62811682477156,
+      "objective/train/tokens_used": 1177170400,
+      "objective/train/value_avg": -0.0085296630859375,
+      "objective/train/value_loss": 0.00013630140165332705,
+      "objective/train/value_max": -8.094310760498047e-05,
+      "objective/train/value_min": -0.234619140625,
+      "objective/train/value_reward_corr": 0.6883133819120735,
+      "objective/train/value_std": 0.01348876953125,
+      "objective/train/weight_avg": 1.000613808631897,
+      "objective/train/weighted_lm_loss": 1.2607276439666748,
+      "objective/train/weights_max": 1.1398074626922607,
+      "objective/train/weights_min": 0.8252780437469482,
+      "theoretical_loss": 3.62811682477156,
+      "tokens_seen": 1156710400
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006559942224362061,
+      "loss": 0.0703,
+      "theoretical_loss": 3.6280790472356705,
+      "tokens_seen": 1156841472
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006559139784946236,
+      "loss": 0.07,
+      "theoretical_loss": 3.628003508597468,
+      "tokens_seen": 1157103616
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006558337345530413,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6279279918612675,
+      "tokens_seen": 1157365760
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006557534906114588,
+      "loss": 0.0705,
+      "theoretical_loss": 3.6278524970157613,
+      "tokens_seen": 1157627904
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006556732466698765,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6277770240496476,
+      "tokens_seen": 1157890048
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000655593002728294,
+      "loss": 0.0695,
+      "theoretical_loss": 3.6277015729516355,
+      "tokens_seen": 1158152192
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006555127587867115,
+      "loss": 0.0709,
+      "theoretical_loss": 3.6276261437104402,
+      "tokens_seen": 1158414336
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006554325148451293,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6275507363147868,
+      "tokens_seen": 1158676480
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006553522709035468,
+      "loss": 0.0715,
+      "theoretical_loss": 3.6274753507534077,
+      "tokens_seen": 1158938624
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006552720269619644,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6273999870150444,
+      "tokens_seen": 1159200768
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000655191783020382,
+      "loss": 0.0684,
+      "theoretical_loss": 3.627324645088446,
+      "tokens_seen": 1159462912
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006551115390787996,
+      "loss": 0.0722,
+      "theoretical_loss": 3.627249324962371,
+      "tokens_seen": 1159725056
+    },
+    {
+      "epoch": 0.35,
+      "objective/train/advantage_avg": 0.0010166645515710115,
+      "objective/train/docs_used": 424758,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4893251657485962,
+      "objective/train/original_loss": 1.4893251657485962,
+      "objective/train/theoretical_loss": 3.6271740266255854,
+      "objective/train/tokens_used": 1180447200,
+      "objective/train/value_avg": -0.00568389892578125,
+      "objective/train/value_loss": 0.00012459652498364449,
+      "objective/train/value_max": -9.101629257202148e-05,
+      "objective/train/value_min": -0.2247314453125,
+      "objective/train/value_reward_corr": 0.6477667845280287,
+      "objective/train/value_std": 0.009185791015625,
+      "objective/train/weight_avg": 1.0010737180709839,
+      "objective/train/weighted_lm_loss": 1.4913969039916992,
+      "objective/train/weights_max": 1.1035935878753662,
+      "objective/train/weights_min": 0.36922982335090637,
+      "theoretical_loss": 3.6271740266255854,
+      "tokens_seen": 1159987200
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006550312951372171,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6271740266255854,
+      "tokens_seen": 1159987200
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006549510511956347,
+      "loss": 0.0727,
+      "theoretical_loss": 3.6270987500668648,
+      "tokens_seen": 1160249344
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006548708072540523,
+      "loss": 0.0727,
+      "theoretical_loss": 3.6270234952749902,
+      "tokens_seen": 1160511488
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006547905633124698,
+      "loss": 0.0712,
+      "theoretical_loss": 3.6269482622387548,
+      "tokens_seen": 1160773632
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006547103193708876,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6268730509469567,
+      "tokens_seen": 1161035776
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006546300754293051,
+      "loss": 0.0699,
+      "theoretical_loss": 3.626797861388404,
+      "tokens_seen": 1161297920
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006545498314877228,
+      "loss": 0.073,
+      "theoretical_loss": 3.6267226935519132,
+      "tokens_seen": 1161560064
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006544695875461403,
+      "loss": 0.0723,
+      "theoretical_loss": 3.626647547426309,
+      "tokens_seen": 1161822208
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006543893436045578,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6265724230004226,
+      "tokens_seen": 1162084352
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006543090996629755,
+      "loss": 0.0713,
+      "theoretical_loss": 3.6264973202630966,
+      "tokens_seen": 1162346496
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000654228855721393,
+      "loss": 0.0697,
+      "theoretical_loss": 3.6264222392031797,
+      "tokens_seen": 1162608640
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006541486117798106,
+      "loss": 0.0696,
+      "theoretical_loss": 3.6263471798095286,
+      "tokens_seen": 1162870784
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006540683678382282,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6262721420710093,
+      "tokens_seen": 1163132928
+    },
+    {
+      "epoch": 0.35,
+      "objective/train/advantage_avg": 0.0016251426422968507,
+      "objective/train/docs_used": 425912,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4275366067886353,
+      "objective/train/original_loss": 1.4275367259979248,
+      "objective/train/theoretical_loss": 3.6262346313189475,
+      "objective/train/tokens_used": 1183724000,
+      "objective/train/value_avg": -0.008636474609375,
+      "objective/train/value_loss": 0.00021610024850815535,
+      "objective/train/value_max": -0.0001366138458251953,
+      "objective/train/value_min": -0.300537109375,
+      "objective/train/value_reward_corr": 0.6914367282185478,
+      "objective/train/value_std": 0.01494598388671875,
+      "objective/train/weight_avg": 1.0017211437225342,
+      "objective/train/weighted_lm_loss": 1.4297378063201904,
+      "objective/train/weights_max": 1.1434952020645142,
+      "objective/train/weights_min": 0.37196043133735657,
+      "theoretical_loss": 3.6262346313189475,
+      "tokens_seen": 1163264000
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006539881238966459,
+      "loss": 0.0709,
+      "theoretical_loss": 3.6261971259764962,
+      "tokens_seen": 1163395072
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006539078799550634,
+      "loss": 0.0696,
+      "theoretical_loss": 3.626122131514871,
+      "tokens_seen": 1163657216
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.000653827636013481,
+      "loss": 0.0724,
+      "theoretical_loss": 3.626047158675024,
+      "tokens_seen": 1163919360
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006537473920718986,
+      "loss": 0.0702,
+      "theoretical_loss": 3.625972207445854,
+      "tokens_seen": 1164181504
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006536671481303161,
+      "loss": 0.0686,
+      "theoretical_loss": 3.625897277816267,
+      "tokens_seen": 1164443648
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006535869041887338,
+      "loss": 0.0709,
+      "theoretical_loss": 3.625822369775179,
+      "tokens_seen": 1164705792
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006535066602471513,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6257474833115113,
+      "tokens_seen": 1164967936
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006534264163055689,
+      "loss": 0.0726,
+      "theoretical_loss": 3.625672618414198,
+      "tokens_seen": 1165230080
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006533461723639865,
+      "loss": 0.0715,
+      "theoretical_loss": 3.6255977750721753,
+      "tokens_seen": 1165492224
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006532659284224042,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6255229532743933,
+      "tokens_seen": 1165754368
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006531856844808218,
+      "loss": 0.0713,
+      "theoretical_loss": 3.625448153009807,
+      "tokens_seen": 1166016512
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006531054405392393,
+      "loss": 0.0692,
+      "theoretical_loss": 3.6253733742673795,
+      "tokens_seen": 1166278656
+    },
+    {
+      "epoch": 0.35,
+      "objective/train/advantage_avg": 0.0018858092371374369,
+      "objective/train/docs_used": 427016,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.495458960533142,
+      "objective/train/original_loss": 1.4954588413238525,
+      "objective/train/theoretical_loss": 3.625298617036084,
+      "objective/train/tokens_used": 1187000800,
+      "objective/train/value_avg": -0.008880615234375,
+      "objective/train/value_loss": 0.00015289847215171903,
+      "objective/train/value_max": -0.00010973215103149414,
+      "objective/train/value_min": -0.302001953125,
+      "objective/train/value_reward_corr": 0.7871390335182029,
+      "objective/train/value_std": 0.01666259765625,
+      "objective/train/weight_avg": 1.001957893371582,
+      "objective/train/weighted_lm_loss": 1.4991862773895264,
+      "objective/train/weights_max": 1.1765505075454712,
+      "objective/train/weights_min": 0.373849481344223,
+      "theoretical_loss": 3.625298617036084,
+      "tokens_seen": 1166540800
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006530251965976569,
+      "loss": 0.0746,
+      "theoretical_loss": 3.625298617036084,
+      "tokens_seen": 1166540800
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006529449526560745,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6252238813049,
+      "tokens_seen": 1166802944
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006528647087144921,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6251491670628155,
+      "tokens_seen": 1167065088
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006527844647729096,
+      "loss": 0.071,
+      "theoretical_loss": 3.6250744742988275,
+      "tokens_seen": 1167327232
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006527042208313273,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6249998030019404,
+      "tokens_seen": 1167589376
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006526239768897448,
+      "loss": 0.0708,
+      "theoretical_loss": 3.6249251531611666,
+      "tokens_seen": 1167851520
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006525437329481623,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6248505247655265,
+      "tokens_seen": 1168113664
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006524634890065801,
+      "loss": 0.073,
+      "theoretical_loss": 3.6247759178040493,
+      "tokens_seen": 1168375808
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006523832450649976,
+      "loss": 0.0705,
+      "theoretical_loss": 3.624701332265772,
+      "tokens_seen": 1168637952
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006523030011234152,
+      "loss": 0.0703,
+      "theoretical_loss": 3.6246267681397386,
+      "tokens_seen": 1168900096
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006522227571818328,
+      "loss": 0.0713,
+      "theoretical_loss": 3.624552225415003,
+      "tokens_seen": 1169162240
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006521425132402504,
+      "loss": 0.0692,
+      "theoretical_loss": 3.6244777040806255,
+      "tokens_seen": 1169424384
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006520622692986679,
+      "loss": 0.0723,
+      "theoretical_loss": 3.624403204125676,
+      "tokens_seen": 1169686528
+    },
+    {
+      "epoch": 0.35,
+      "objective/train/advantage_avg": 0.00018043404270429164,
+      "objective/train/docs_used": 428177,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4640066623687744,
+      "objective/train/original_loss": 1.4640066623687744,
+      "objective/train/theoretical_loss": 3.6243659621620727,
+      "objective/train/tokens_used": 1190277600,
+      "objective/train/value_avg": -0.010498046875,
+      "objective/train/value_loss": 0.00026088516460731626,
+      "objective/train/value_max": -5.4776668548583984e-05,
+      "objective/train/value_min": -0.31982421875,
+      "objective/train/value_reward_corr": 0.7486174859686674,
+      "objective/train/value_std": 0.017181396484375,
+      "objective/train/weight_avg": 1.000304102897644,
+      "objective/train/weighted_lm_loss": 1.4634703397750854,
+      "objective/train/weights_max": 1.2624222040176392,
+      "objective/train/weights_min": 0.3704579770565033,
+      "theoretical_loss": 3.6243659621620727,
+      "tokens_seen": 1169817600
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006519820253570855,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6243287255392307,
+      "tokens_seen": 1169948672
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006519017814155031,
+      "loss": 0.0723,
+      "theoretical_loss": 3.624254268310375,
+      "tokens_seen": 1170210816
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006518215374739207,
+      "loss": 0.0697,
+      "theoretical_loss": 3.6241798324282017,
+      "tokens_seen": 1170472960
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006517412935323384,
+      "loss": 0.0731,
+      "theoretical_loss": 3.624105417881813,
+      "tokens_seen": 1170735104
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006516610495907559,
+      "loss": 0.0731,
+      "theoretical_loss": 3.624031024660317,
+      "tokens_seen": 1170997248
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 0.0006515808056491736,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6239566527528306,
+      "tokens_seen": 1171259392
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006515005617075911,
+      "loss": 0.0715,
+      "theoretical_loss": 3.6238823021484796,
+      "tokens_seen": 1171521536
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006514203177660086,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6238079728363974,
+      "tokens_seen": 1171783680
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006513400738244263,
+      "loss": 0.0711,
+      "theoretical_loss": 3.6237336648057243,
+      "tokens_seen": 1172045824
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006512598298828438,
+      "loss": 0.0692,
+      "theoretical_loss": 3.62365937804561,
+      "tokens_seen": 1172307968
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006511795859412614,
+      "loss": 0.0686,
+      "theoretical_loss": 3.6235851125452108,
+      "tokens_seen": 1172570112
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.000651099341999679,
+      "loss": 0.0722,
+      "theoretical_loss": 3.6235108682936916,
+      "tokens_seen": 1172832256
+    },
+    {
+      "epoch": 0.36,
+      "objective/train/advantage_avg": 0.0007615335052832961,
+      "objective/train/docs_used": 429267,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.439132809638977,
+      "objective/train/original_loss": 1.4391329288482666,
+      "objective/train/theoretical_loss": 3.623436645280226,
+      "objective/train/tokens_used": 1193554400,
+      "objective/train/value_avg": -0.005706787109375,
+      "objective/train/value_loss": 8.419386722380295e-05,
+      "objective/train/value_max": -5.692243576049805e-05,
+      "objective/train/value_min": -0.278564453125,
+      "objective/train/value_reward_corr": 0.6225248522847999,
+      "objective/train/value_std": 0.007747650146484375,
+      "objective/train/weight_avg": 1.0008031129837036,
+      "objective/train/weighted_lm_loss": 1.4398229122161865,
+      "objective/train/weights_max": 1.1341171264648438,
+      "objective/train/weights_min": 0.7353724837303162,
+      "theoretical_loss": 3.623436645280226,
+      "tokens_seen": 1173094400
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006510190980580967,
+      "loss": 0.0723,
+      "theoretical_loss": 3.623436645280226,
+      "tokens_seen": 1173094400
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006509388541165142,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6233624434939946,
+      "tokens_seen": 1173356544
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006508586101749318,
+      "loss": 0.07,
+      "theoretical_loss": 3.623288262924186,
+      "tokens_seen": 1173618688
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006507783662333494,
+      "loss": 0.0739,
+      "theoretical_loss": 3.623214103559997,
+      "tokens_seen": 1173880832
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.000650698122291767,
+      "loss": 0.072,
+      "theoretical_loss": 3.6231399653906315,
+      "tokens_seen": 1174142976
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006506178783501846,
+      "loss": 0.0686,
+      "theoretical_loss": 3.623065848405303,
+      "tokens_seen": 1174405120
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006505376344086021,
+      "loss": 0.0717,
+      "theoretical_loss": 3.622991752593231,
+      "tokens_seen": 1174667264
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006504573904670198,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6229176779436445,
+      "tokens_seen": 1174929408
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006503771465254373,
+      "loss": 0.0706,
+      "theoretical_loss": 3.6228436244457796,
+      "tokens_seen": 1175191552
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.000650296902583855,
+      "loss": 0.072,
+      "theoretical_loss": 3.6227695920888796,
+      "tokens_seen": 1175453696
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006502166586422726,
+      "loss": 0.0705,
+      "theoretical_loss": 3.622695580862197,
+      "tokens_seen": 1175715840
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006501364147006901,
+      "loss": 0.0702,
+      "theoretical_loss": 3.6226215907549912,
+      "tokens_seen": 1175977984
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006500561707591077,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6225476217565307,
+      "tokens_seen": 1176240128
+    },
+    {
+      "epoch": 0.36,
+      "objective/train/advantage_avg": 2.649899761308916e-05,
+      "objective/train/docs_used": 430645,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.540775179862976,
+      "objective/train/original_loss": 1.5407750606536865,
+      "objective/train/theoretical_loss": 3.6225106451697275,
+      "objective/train/tokens_used": 1196831200,
+      "objective/train/value_avg": -0.00846099853515625,
+      "objective/train/value_loss": 0.00018133767298422754,
+      "objective/train/value_max": -3.0219554901123047e-05,
+      "objective/train/value_min": -0.51318359375,
+      "objective/train/value_reward_corr": 0.7236088657344646,
+      "objective/train/value_std": 0.01409912109375,
+      "objective/train/weight_avg": 1.0001122951507568,
+      "objective/train/weighted_lm_loss": 1.5408101081848145,
+      "objective/train/weights_max": 1.1531023979187012,
+      "objective/train/weights_min": 0.3940271735191345,
+      "theoretical_loss": 3.6225106451697275,
+      "tokens_seen": 1176371200
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006499759268175253,
+      "loss": 0.0713,
+      "theoretical_loss": 3.6224736738560894,
+      "tokens_seen": 1176502272
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006498956828759429,
+      "loss": 0.0755,
+      "theoretical_loss": 3.6223997470429516,
+      "tokens_seen": 1176764416
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006498154389343604,
+      "loss": 0.0713,
+      "theoretical_loss": 3.6223258413064086,
+      "tokens_seen": 1177026560
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006497351949927781,
+      "loss": 0.0734,
+      "theoretical_loss": 3.6222519566357585,
+      "tokens_seen": 1177288704
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006496549510511956,
+      "loss": 0.0711,
+      "theoretical_loss": 3.6221780930203096,
+      "tokens_seen": 1177550848
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006495747071096131,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6221042504493743,
+      "tokens_seen": 1177812992
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006494944631680309,
+      "loss": 0.0738,
+      "theoretical_loss": 3.622030428912276,
+      "tokens_seen": 1178075136
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006494142192264484,
+      "loss": 0.0699,
+      "theoretical_loss": 3.6219566283983458,
+      "tokens_seen": 1178337280
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006493339752848661,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6218828488969197,
+      "tokens_seen": 1178599424
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006492537313432836,
+      "loss": 0.0706,
+      "theoretical_loss": 3.6218090903973446,
+      "tokens_seen": 1178861568
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006491734874017012,
+      "loss": 0.0764,
+      "theoretical_loss": 3.621735352888974,
+      "tokens_seen": 1179123712
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006490932434601188,
+      "loss": 0.0787,
+      "theoretical_loss": 3.621661636361169,
+      "tokens_seen": 1179385856
+    },
+    {
+      "epoch": 0.36,
+      "objective/train/advantage_avg": -0.00019712685025297105,
+      "objective/train/docs_used": 431781,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4777541160583496,
+      "objective/train/original_loss": 1.4777541160583496,
+      "objective/train/theoretical_loss": 3.6215879408032987,
+      "objective/train/tokens_used": 1200108000,
+      "objective/train/value_avg": -0.006439208984375,
+      "objective/train/value_loss": 0.000245880801230669,
+      "objective/train/value_max": -6.920099258422852e-05,
+      "objective/train/value_min": -0.25927734375,
+      "objective/train/value_reward_corr": 0.6481294948793725,
+      "objective/train/value_std": 0.011993408203125,
+      "objective/train/weight_avg": 0.999910295009613,
+      "objective/train/weighted_lm_loss": 1.476799488067627,
+      "objective/train/weights_max": 1.2023646831512451,
+      "objective/train/weights_min": 0.37629902362823486,
+      "theoretical_loss": 3.6215879408032987,
+      "tokens_seen": 1179648000
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006490129995185363,
+      "loss": 0.0742,
+      "theoretical_loss": 3.6215879408032987,
+      "tokens_seen": 1179648000
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006489327555769539,
+      "loss": 0.074,
+      "theoretical_loss": 3.6215142662047395,
+      "tokens_seen": 1179910144
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006488525116353715,
+      "loss": 0.0742,
+      "theoretical_loss": 3.621440612554876,
+      "tokens_seen": 1180172288
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006487722676937892,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6213669798431005,
+      "tokens_seen": 1180434432
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006486920237522067,
+      "loss": 0.0724,
+      "theoretical_loss": 3.621293368058813,
+      "tokens_seen": 1180696576
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006486117798106244,
+      "loss": 0.0739,
+      "theoretical_loss": 3.621219777191421,
+      "tokens_seen": 1180958720
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006485315358690419,
+      "loss": 0.0746,
+      "theoretical_loss": 3.62114620723034,
+      "tokens_seen": 1181220864
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006484512919274594,
+      "loss": 0.0738,
+      "theoretical_loss": 3.621072658164993,
+      "tokens_seen": 1181483008
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006483710479858771,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6209991299848108,
+      "tokens_seen": 1181745152
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006482908040442946,
+      "loss": 0.0751,
+      "theoretical_loss": 3.620925622679232,
+      "tokens_seen": 1182007296
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006482105601027123,
+      "loss": 0.0738,
+      "theoretical_loss": 3.620852136237702,
+      "tokens_seen": 1182269440
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006481303161611298,
+      "loss": 0.0741,
+      "theoretical_loss": 3.620778670649676,
+      "tokens_seen": 1182531584
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006480500722195475,
+      "loss": 0.0709,
+      "theoretical_loss": 3.620705225904614,
+      "tokens_seen": 1182793728
+    },
+    {
+      "epoch": 0.36,
+      "objective/train/advantage_avg": 0.0012156600132584572,
+      "objective/train/docs_used": 433029,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4418714046478271,
+      "objective/train/original_loss": 1.4418714046478271,
+      "objective/train/theoretical_loss": 3.6206685113449044,
+      "objective/train/tokens_used": 1203384800,
+      "objective/train/value_avg": -0.007598876953125,
+      "objective/train/value_loss": 0.0002602523018140346,
+      "objective/train/value_max": -0.00014889240264892578,
+      "objective/train/value_min": -0.91015625,
+      "objective/train/value_reward_corr": 0.7045290549833653,
+      "objective/train/value_std": 0.014434814453125,
+      "objective/train/weight_avg": 1.0013296604156494,
+      "objective/train/weighted_lm_loss": 1.4433629512786865,
+      "objective/train/weights_max": 1.1642762422561646,
+      "objective/train/weights_min": 0.37042829394340515,
+      "theoretical_loss": 3.6206685113449044,
+      "tokens_seen": 1182924800
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006479698282779651,
+      "loss": 0.0734,
+      "theoretical_loss": 3.620631801991987,
+      "tokens_seen": 1183055872
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006478895843363826,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6205583989012697,
+      "tokens_seen": 1183318016
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006478093403948002,
+      "loss": 0.075,
+      "theoretical_loss": 3.6204850166219478,
+      "tokens_seen": 1183580160
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006477290964532178,
+      "loss": 0.07,
+      "theoretical_loss": 3.6204116551435126,
+      "tokens_seen": 1183842304
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006476488525116354,
+      "loss": 0.0746,
+      "theoretical_loss": 3.620338314455465,
+      "tokens_seen": 1184104448
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006475686085700529,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6202649945473113,
+      "tokens_seen": 1184366592
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006474883646284706,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6201916954085664,
+      "tokens_seen": 1184628736
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006474081206868881,
+      "loss": 0.0724,
+      "theoretical_loss": 3.620118417028754,
+      "tokens_seen": 1184890880
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006473278767453056,
+      "loss": 0.0735,
+      "theoretical_loss": 3.620045159397403,
+      "tokens_seen": 1185153024
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006472476328037234,
+      "loss": 0.0741,
+      "theoretical_loss": 3.619971922504052,
+      "tokens_seen": 1185415168
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006471673888621409,
+      "loss": 0.0712,
+      "theoretical_loss": 3.6198987063382457,
+      "tokens_seen": 1185677312
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006470871449205585,
+      "loss": 0.0757,
+      "theoretical_loss": 3.6198255108895374,
+      "tokens_seen": 1185939456
+    },
+    {
+      "epoch": 0.36,
+      "objective/train/advantage_avg": 0.001317921094596386,
+      "objective/train/docs_used": 434243,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5565327405929565,
+      "objective/train/original_loss": 1.556532859802246,
+      "objective/train/theoretical_loss": 3.6197523361474877,
+      "objective/train/tokens_used": 1206661600,
+      "objective/train/value_avg": -0.0074920654296875,
+      "objective/train/value_loss": 0.00021194624423515052,
+      "objective/train/value_max": -0.00010389089584350586,
+      "objective/train/value_min": -0.22802734375,
+      "objective/train/value_reward_corr": 0.585920330003381,
+      "objective/train/value_std": 0.01111602783203125,
+      "objective/train/weight_avg": 1.0014089345932007,
+      "objective/train/weighted_lm_loss": 1.5591508150100708,
+      "objective/train/weights_max": 1.1952372789382935,
+      "objective/train/weights_min": 0.3731912076473236,
+      "theoretical_loss": 3.6197523361474877,
+      "tokens_seen": 1186201600
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006470069009789761,
+      "loss": 0.0754,
+      "theoretical_loss": 3.6197523361474877,
+      "tokens_seen": 1186201600
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006469266570373937,
+      "loss": 0.0723,
+      "theoretical_loss": 3.619679182101664,
+      "tokens_seen": 1186463744
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006468464130958113,
+      "loss": 0.0766,
+      "theoretical_loss": 3.619606048741643,
+      "tokens_seen": 1186725888
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006467661691542289,
+      "loss": 0.0751,
+      "theoretical_loss": 3.6195329360570065,
+      "tokens_seen": 1186988032
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006466859252126464,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6194598440373467,
+      "tokens_seen": 1187250176
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.000646605681271064,
+      "loss": 0.0721,
+      "theoretical_loss": 3.619386772672261,
+      "tokens_seen": 1187512320
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006465254373294817,
+      "loss": 0.0767,
+      "theoretical_loss": 3.6193137219513556,
+      "tokens_seen": 1187774464
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006464451933878992,
+      "loss": 0.0759,
+      "theoretical_loss": 3.619240691864243,
+      "tokens_seen": 1188036608
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006463649494463169,
+      "loss": 0.071,
+      "theoretical_loss": 3.619167682400545,
+      "tokens_seen": 1188298752
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006462847055047344,
+      "loss": 0.0769,
+      "theoretical_loss": 3.619094693549889,
+      "tokens_seen": 1188560896
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.000646204461563152,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6190217253019124,
+      "tokens_seen": 1188823040
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006461242176215696,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6189487776462568,
+      "tokens_seen": 1189085184
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006460439736799871,
+      "loss": 0.0707,
+      "theoretical_loss": 3.6188758505725738,
+      "tokens_seen": 1189347328
+    },
+    {
+      "epoch": 0.36,
+      "objective/train/advantage_avg": 0.0009901868179440498,
+      "objective/train/docs_used": 435462,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4739112854003906,
+      "objective/train/original_loss": 1.4739114046096802,
+      "objective/train/theoretical_loss": 3.61883939475074,
+      "objective/train/tokens_used": 1209938400,
+      "objective/train/value_avg": -0.00931549072265625,
+      "objective/train/value_loss": 0.0008738775504752994,
+      "objective/train/value_max": -0.00011771917343139648,
+      "objective/train/value_min": -0.97705078125,
+      "objective/train/value_reward_corr": 0.5927413814353182,
+      "objective/train/value_std": 0.02313232421875,
+      "objective/train/weight_avg": 1.0013524293899536,
+      "objective/train/weighted_lm_loss": 1.4759513139724731,
+      "objective/train/weights_max": 2.4594109058380127,
+      "objective/train/weights_min": 0.22472499310970306,
+      "theoretical_loss": 3.61883939475074,
+      "tokens_seen": 1189478400
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006459637297384047,
+      "loss": 0.0769,
+      "theoretical_loss": 3.618802944070522,
+      "tokens_seen": 1189609472
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006458834857968223,
+      "loss": 0.0741,
+      "theoretical_loss": 3.618730058129766,
+      "tokens_seen": 1189871616
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.00064580324185524,
+      "loss": 0.0746,
+      "theoretical_loss": 3.61865719273998,
+      "tokens_seen": 1190133760
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006457229979136575,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6185843478908453,
+      "tokens_seen": 1190395904
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006456427539720752,
+      "loss": 0.0736,
+      "theoretical_loss": 3.618511523572049,
+      "tokens_seen": 1190658048
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006455625100304927,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6184387197732875,
+      "tokens_seen": 1190920192
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006454822660889103,
+      "loss": 0.0773,
+      "theoretical_loss": 3.6183659364842624,
+      "tokens_seen": 1191182336
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006454020221473279,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6182931736946857,
+      "tokens_seen": 1191444480
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006453217782057454,
+      "loss": 0.0722,
+      "theoretical_loss": 3.618220431394274,
+      "tokens_seen": 1191706624
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006452415342641631,
+      "loss": 0.0704,
+      "theoretical_loss": 3.618147709572754,
+      "tokens_seen": 1191968768
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006451612903225806,
+      "loss": 0.0739,
+      "theoretical_loss": 3.618075008219858,
+      "tokens_seen": 1192230912
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006450810463809983,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6180023273253252,
+      "tokens_seen": 1192493056
+    },
+    {
+      "epoch": 0.36,
+      "objective/train/advantage_avg": 0.0011386601254343987,
+      "objective/train/docs_used": 436657,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5473337173461914,
+      "objective/train/original_loss": 1.5473337173461914,
+      "objective/train/theoretical_loss": 3.6179296668789043,
+      "objective/train/tokens_used": 1213215200,
+      "objective/train/value_avg": -0.01073455810546875,
+      "objective/train/value_loss": 0.00023775658337399364,
+      "objective/train/value_max": -0.00012934207916259766,
+      "objective/train/value_min": -0.73193359375,
+      "objective/train/value_reward_corr": 0.7040417032080557,
+      "objective/train/value_std": 0.0167999267578125,
+      "objective/train/weight_avg": 1.0012502670288086,
+      "objective/train/weighted_lm_loss": 1.5486918687820435,
+      "objective/train/weights_max": 1.341107964515686,
+      "objective/train/weights_min": 0.3696603775024414,
+      "theoretical_loss": 3.6179296668789043,
+      "tokens_seen": 1192755200
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006450008024394159,
+      "loss": 0.0754,
+      "theoretical_loss": 3.6179296668789043,
+      "tokens_seen": 1192755200
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006449205584978334,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6178570268703494,
+      "tokens_seen": 1193017344
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.000644840314556251,
+      "loss": 0.0734,
+      "theoretical_loss": 3.617784407289424,
+      "tokens_seen": 1193279488
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006447600706146686,
+      "loss": 0.0747,
+      "theoretical_loss": 3.617711808125896,
+      "tokens_seen": 1193541632
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006446798266730862,
+      "loss": 0.0753,
+      "theoretical_loss": 3.6176392293695434,
+      "tokens_seen": 1193803776
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006445995827315037,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6175666710101506,
+      "tokens_seen": 1194065920
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006445193387899214,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6174941330375097,
+      "tokens_seen": 1194328064
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006444390948483389,
+      "loss": 0.0752,
+      "theoretical_loss": 3.617421615441419,
+      "tokens_seen": 1194590208
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006443588509067564,
+      "loss": 0.0778,
+      "theoretical_loss": 3.617349118211685,
+      "tokens_seen": 1194852352
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006442786069651742,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6172766413381225,
+      "tokens_seen": 1195114496
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006441983630235917,
+      "loss": 0.0702,
+      "theoretical_loss": 3.617204184810552,
+      "tokens_seen": 1195376640
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006441181190820094,
+      "loss": 0.074,
+      "theoretical_loss": 3.6171317486188013,
+      "tokens_seen": 1195638784
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006440378751404269,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6170593327527074,
+      "tokens_seen": 1195900928
+    },
+    {
+      "epoch": 0.36,
+      "objective/train/advantage_avg": 0.0009150534751825035,
+      "objective/train/docs_used": 437781,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5300925970077515,
+      "objective/train/original_loss": 1.530092477798462,
+      "objective/train/theoretical_loss": 3.617023132438607,
+      "objective/train/tokens_used": 1216492000,
+      "objective/train/value_avg": -0.006450653076171875,
+      "objective/train/value_loss": 0.00013045086234342307,
+      "objective/train/value_max": -0.00013446807861328125,
+      "objective/train/value_min": -0.371826171875,
+      "objective/train/value_reward_corr": 0.6219127786332249,
+      "objective/train/value_std": 0.009521484375,
+      "objective/train/weight_avg": 1.000975251197815,
+      "objective/train/weighted_lm_loss": 1.5323002338409424,
+      "objective/train/weights_max": 1.1521174907684326,
+      "objective/train/weights_min": 0.3824407160282135,
+      "theoretical_loss": 3.617023132438607,
+      "tokens_seen": 1196032000
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006439576311988445,
+      "loss": 0.0727,
+      "theoretical_loss": 3.616986937202112,
+      "tokens_seen": 1196163072
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006438773872572621,
+      "loss": 0.0745,
+      "theoretical_loss": 3.616914561956867,
+      "tokens_seen": 1196425216
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006437971433156797,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6168422070068287,
+      "tokens_seen": 1196687360
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006437168993740972,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6167698723418624,
+      "tokens_seen": 1196949504
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006436366554325149,
+      "loss": 0.0761,
+      "theoretical_loss": 3.6166975579518406,
+      "tokens_seen": 1197211648
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006435564114909325,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6166252638266423,
+      "tokens_seen": 1197473792
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.00064347616754935,
+      "loss": 0.0746,
+      "theoretical_loss": 3.616552989956155,
+      "tokens_seen": 1197735936
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006433959236077677,
+      "loss": 0.0745,
+      "theoretical_loss": 3.616480736330272,
+      "tokens_seen": 1197998080
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006433156796661852,
+      "loss": 0.0711,
+      "theoretical_loss": 3.616408502938895,
+      "tokens_seen": 1198260224
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006432354357246028,
+      "loss": 0.0701,
+      "theoretical_loss": 3.616336289771932,
+      "tokens_seen": 1198522368
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006431551917830204,
+      "loss": 0.0711,
+      "theoretical_loss": 3.6162640968192994,
+      "tokens_seen": 1198784512
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006430749478414379,
+      "loss": 0.0728,
+      "theoretical_loss": 3.61619192407092,
+      "tokens_seen": 1199046656
+    },
+    {
+      "epoch": 0.36,
+      "objective/train/advantage_avg": 0.0014068173477426171,
+      "objective/train/docs_used": 438488,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5579599142074585,
+      "objective/train/original_loss": 1.5579596757888794,
+      "objective/train/theoretical_loss": 3.6161197715167237,
+      "objective/train/tokens_used": 1219768800,
+      "objective/train/value_avg": -0.00788116455078125,
+      "objective/train/value_loss": 0.0003655260952655226,
+      "objective/train/value_max": -9.840726852416992e-05,
+      "objective/train/value_min": -0.95556640625,
+      "objective/train/value_reward_corr": 0.7506662225236421,
+      "objective/train/value_std": 0.0210113525390625,
+      "objective/train/weight_avg": 1.0015637874603271,
+      "objective/train/weighted_lm_loss": 1.5613462924957275,
+      "objective/train/weights_max": 1.6007517576217651,
+      "objective/train/weights_min": 0.22900784015655518,
+      "theoretical_loss": 3.6161197715167237,
+      "tokens_seen": 1199308800
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006429947038998556,
+      "loss": 0.0722,
+      "theoretical_loss": 3.6161197715167237,
+      "tokens_seen": 1199308800
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006429144599582731,
+      "loss": 0.0719,
+      "theoretical_loss": 3.616047639146648,
+      "tokens_seen": 1199570944
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006428342160166908,
+      "loss": 0.0717,
+      "theoretical_loss": 3.6159755269506375,
+      "tokens_seen": 1199833088
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006427539720751084,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6159034349186445,
+      "tokens_seen": 1200095232
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.000642673728133526,
+      "loss": 0.0759,
+      "theoretical_loss": 3.615831363040628,
+      "tokens_seen": 1200357376
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006425934841919435,
+      "loss": 0.071,
+      "theoretical_loss": 3.615759311306553,
+      "tokens_seen": 1200619520
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006425132402503611,
+      "loss": 0.0715,
+      "theoretical_loss": 3.6156872797063944,
+      "tokens_seen": 1200881664
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006424329963087787,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6156152682301324,
+      "tokens_seen": 1201143808
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006423527523671962,
+      "loss": 0.0753,
+      "theoretical_loss": 3.6155432768677542,
+      "tokens_seen": 1201405952
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006422725084256139,
+      "loss": 0.0726,
+      "theoretical_loss": 3.615471305609255,
+      "tokens_seen": 1201668096
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006421922644840314,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6153993544446372,
+      "tokens_seen": 1201930240
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006421120205424491,
+      "loss": 0.0729,
+      "theoretical_loss": 3.61532742336391,
+      "tokens_seen": 1202192384
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006420317766008667,
+      "loss": 0.071,
+      "theoretical_loss": 3.6152555123570895,
+      "tokens_seen": 1202454528
+    },
+    {
+      "epoch": 0.36,
+      "objective/train/advantage_avg": 0.00142679491546005,
+      "objective/train/docs_used": 439824,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5132286548614502,
+      "objective/train/original_loss": 1.5132288932800293,
+      "objective/train/theoretical_loss": 3.6152195643782763,
+      "objective/train/tokens_used": 1223045600,
+      "objective/train/value_avg": -0.00974273681640625,
+      "objective/train/value_loss": 0.00020919894450344145,
+      "objective/train/value_max": -0.00011414289474487305,
+      "objective/train/value_min": -0.3330078125,
+      "objective/train/value_reward_corr": 0.7210474728015025,
+      "objective/train/value_std": 0.01556396484375,
+      "objective/train/weight_avg": 1.0015263557434082,
+      "objective/train/weighted_lm_loss": 1.5158843994140625,
+      "objective/train/weights_max": 1.3555892705917358,
+      "objective/train/weights_min": 0.3693855106830597,
+      "theoretical_loss": 3.6152195643782763,
+      "tokens_seen": 1202585600
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006419515326592842,
+      "loss": 0.0687,
+      "theoretical_loss": 3.6151836214141992,
+      "tokens_seen": 1202716672
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006418712887177018,
+      "loss": 0.0738,
+      "theoretical_loss": 3.61511175052527,
+      "tokens_seen": 1202978816
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006417910447761194,
+      "loss": 0.0712,
+      "theoretical_loss": 3.6150398996803395,
+      "tokens_seen": 1203240960
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.000641710800834537,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6149680688694525,
+      "tokens_seen": 1203503104
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006416305568929546,
+      "loss": 0.0736,
+      "theoretical_loss": 3.614896258082661,
+      "tokens_seen": 1203765248
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006415503129513722,
+      "loss": 0.071,
+      "theoretical_loss": 3.614824467310025,
+      "tokens_seen": 1204027392
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 0.0006414700690097897,
+      "loss": 0.0744,
+      "theoretical_loss": 3.614752696541609,
+      "tokens_seen": 1204289536
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006413898250682074,
+      "loss": 0.0733,
+      "theoretical_loss": 3.614680945767488,
+      "tokens_seen": 1204551680
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000641309581126625,
+      "loss": 0.0727,
+      "theoretical_loss": 3.614609214977741,
+      "tokens_seen": 1204813824
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006412293371850425,
+      "loss": 0.075,
+      "theoretical_loss": 3.614537504162457,
+      "tokens_seen": 1205075968
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006411490932434602,
+      "loss": 0.0737,
+      "theoretical_loss": 3.614465813311729,
+      "tokens_seen": 1205338112
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006410688493018777,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6143941424156596,
+      "tokens_seen": 1205600256
+    },
+    {
+      "epoch": 0.37,
+      "objective/train/advantage_avg": 0.00039125708281062543,
+      "objective/train/docs_used": 441116,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4485985040664673,
+      "objective/train/original_loss": 1.4485986232757568,
+      "objective/train/theoretical_loss": 3.614322491464357,
+      "objective/train/tokens_used": 1226322400,
+      "objective/train/value_avg": -0.01201629638671875,
+      "objective/train/value_loss": 0.0005621611489914358,
+      "objective/train/value_max": -0.00015115737915039062,
+      "objective/train/value_min": -0.95751953125,
+      "objective/train/value_reward_corr": 0.7828427996520164,
+      "objective/train/value_std": 0.02960205078125,
+      "objective/train/weight_avg": 1.0006436109542847,
+      "objective/train/weighted_lm_loss": 1.449430227279663,
+      "objective/train/weights_max": 1.638288974761963,
+      "objective/train/weights_min": 0.3702460527420044,
+      "theoretical_loss": 3.614322491464357,
+      "tokens_seen": 1205862400
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006409886053602953,
+      "loss": 0.0726,
+      "theoretical_loss": 3.614322491464357,
+      "tokens_seen": 1205862400
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006409083614187129,
+      "loss": 0.0761,
+      "theoretical_loss": 3.614250860447936,
+      "tokens_seen": 1206124544
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006408281174771304,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6141792493565212,
+      "tokens_seen": 1206386688
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000640747873535548,
+      "loss": 0.0711,
+      "theoretical_loss": 3.6141076581802416,
+      "tokens_seen": 1206648832
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006406676295939656,
+      "loss": 0.0718,
+      "theoretical_loss": 3.614036086909234,
+      "tokens_seen": 1206910976
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006405873856523833,
+      "loss": 0.075,
+      "theoretical_loss": 3.6139645355336425,
+      "tokens_seen": 1207173120
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006405071417108009,
+      "loss": 0.0761,
+      "theoretical_loss": 3.613893004043617,
+      "tokens_seen": 1207435264
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006404268977692185,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6138214924293166,
+      "tokens_seen": 1207697408
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000640346653827636,
+      "loss": 0.077,
+      "theoretical_loss": 3.6137500006809056,
+      "tokens_seen": 1207959552
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006402664098860537,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6136785287885553,
+      "tokens_seen": 1208221696
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006401861659444712,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6136070767424457,
+      "tokens_seen": 1208483840
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006401059220028887,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6135356445327624,
+      "tokens_seen": 1208745984
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006400256780613064,
+      "loss": 0.0691,
+      "theoretical_loss": 3.6134642321496977,
+      "tokens_seen": 1209008128
+    },
+    {
+      "epoch": 0.37,
+      "objective/train/advantage_avg": 0.0001867167593445629,
+      "objective/train/docs_used": 442331,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2952028512954712,
+      "objective/train/original_loss": 1.2952029705047607,
+      "objective/train/theoretical_loss": 3.613428533390085,
+      "objective/train/tokens_used": 1229599200,
+      "objective/train/value_avg": -0.00867462158203125,
+      "objective/train/value_loss": 0.00022595007612835616,
+      "objective/train/value_max": -4.756450653076172e-05,
+      "objective/train/value_min": -0.47412109375,
+      "objective/train/value_reward_corr": 0.820079703282189,
+      "objective/train/value_std": 0.019287109375,
+      "objective/train/weight_avg": 1.0002939701080322,
+      "objective/train/weighted_lm_loss": 1.2954621315002441,
+      "objective/train/weights_max": 1.459554672241211,
+      "objective/train/weights_min": 0.3717447817325592,
+      "theoretical_loss": 3.613428533390085,
+      "tokens_seen": 1209139200
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000639945434119724,
+      "loss": 0.0705,
+      "theoretical_loss": 3.613392839583452,
+      "tokens_seen": 1209270272
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006398651901781416,
+      "loss": 0.0742,
+      "theoretical_loss": 3.6133214668242317,
+      "tokens_seen": 1209532416
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006397849462365592,
+      "loss": 0.0707,
+      "theoretical_loss": 3.613250113862251,
+      "tokens_seen": 1209794560
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006397047022949768,
+      "loss": 0.0724,
+      "theoretical_loss": 3.61317878068773,
+      "tokens_seen": 1210056704
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006396244583533943,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6131074672908965,
+      "tokens_seen": 1210318848
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006395442144118119,
+      "loss": 0.0714,
+      "theoretical_loss": 3.6130361736619854,
+      "tokens_seen": 1210580992
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006394639704702295,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6129648997912382,
+      "tokens_seen": 1210843136
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000639383726528647,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6128936456689034,
+      "tokens_seen": 1211105280
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006393034825870647,
+      "loss": 0.073,
+      "theoretical_loss": 3.612822411285236,
+      "tokens_seen": 1211367424
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006392232386454822,
+      "loss": 0.0755,
+      "theoretical_loss": 3.612751196630499,
+      "tokens_seen": 1211629568
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006391429947039,
+      "loss": 0.0725,
+      "theoretical_loss": 3.61268000169496,
+      "tokens_seen": 1211891712
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006390627507623175,
+      "loss": 0.0744,
+      "theoretical_loss": 3.612608826468897,
+      "tokens_seen": 1212153856
+    },
+    {
+      "epoch": 0.37,
+      "objective/train/advantage_avg": -0.0002951657515950501,
+      "objective/train/docs_used": 443456,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.476730465888977,
+      "objective/train/original_loss": 1.476730465888977,
+      "objective/train/theoretical_loss": 3.6125376709425923,
+      "objective/train/tokens_used": 1232876000,
+      "objective/train/value_avg": -0.0062255859375,
+      "objective/train/value_loss": 0.00031703378772363067,
+      "objective/train/value_max": -0.00010073184967041016,
+      "objective/train/value_min": -0.1983642578125,
+      "objective/train/value_reward_corr": 0.5360043017511746,
+      "objective/train/value_std": 0.00960540771484375,
+      "objective/train/weight_avg": 0.999844491481781,
+      "objective/train/weighted_lm_loss": 1.4760133028030396,
+      "objective/train/weights_max": 1.2021102905273438,
+      "objective/train/weights_min": 0.3719916045665741,
+      "theoretical_loss": 3.6125376709425923,
+      "tokens_seen": 1212416000
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000638982506820735,
+      "loss": 0.0738,
+      "theoretical_loss": 3.6125376709425923,
+      "tokens_seen": 1212416000
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006389022628791527,
+      "loss": 0.0683,
+      "theoretical_loss": 3.6124665351063356,
+      "tokens_seen": 1212678144
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006388220189375702,
+      "loss": 0.0737,
+      "theoretical_loss": 3.612395418950424,
+      "tokens_seen": 1212940288
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006387417749959878,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6123243224651604,
+      "tokens_seen": 1213202432
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006386615310544054,
+      "loss": 0.0719,
+      "theoretical_loss": 3.612253245640856,
+      "tokens_seen": 1213464576
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000638581287112823,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6121821884678287,
+      "tokens_seen": 1213726720
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006385010431712405,
+      "loss": 0.0743,
+      "theoretical_loss": 3.6121111509364017,
+      "tokens_seen": 1213988864
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006384207992296582,
+      "loss": 0.075,
+      "theoretical_loss": 3.6120401330369067,
+      "tokens_seen": 1214251008
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006383405552880758,
+      "loss": 0.0722,
+      "theoretical_loss": 3.6119691347596814,
+      "tokens_seen": 1214513152
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006382603113464933,
+      "loss": 0.074,
+      "theoretical_loss": 3.611898156095071,
+      "tokens_seen": 1214775296
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000638180067404911,
+      "loss": 0.0736,
+      "theoretical_loss": 3.6118271970334264,
+      "tokens_seen": 1215037440
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006380998234633285,
+      "loss": 0.0722,
+      "theoretical_loss": 3.6117562575651068,
+      "tokens_seen": 1215299584
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006380195795217462,
+      "loss": 0.0728,
+      "theoretical_loss": 3.611685337680477,
+      "tokens_seen": 1215561728
+    },
+    {
+      "epoch": 0.37,
+      "objective/train/advantage_avg": 0.0022401802707463503,
+      "objective/train/docs_used": 444688,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3715991973876953,
+      "objective/train/original_loss": 1.3715991973876953,
+      "objective/train/theoretical_loss": 3.6116498850790366,
+      "objective/train/tokens_used": 1236152800,
+      "objective/train/value_avg": -0.005950927734375,
+      "objective/train/value_loss": 0.00014855283370707184,
+      "objective/train/value_max": -0.0001233816146850586,
+      "objective/train/value_min": -0.921875,
+      "objective/train/value_reward_corr": 0.452322266230042,
+      "objective/train/value_std": 0.01007080078125,
+      "objective/train/weight_avg": 1.0023151636123657,
+      "objective/train/weighted_lm_loss": 1.3756544589996338,
+      "objective/train/weights_max": 2.5139997005462646,
+      "objective/train/weights_min": 0.3757654130458832,
+      "theoretical_loss": 3.6116498850790366,
+      "tokens_seen": 1215692800
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006379393355801637,
+      "loss": 0.072,
+      "theoretical_loss": 3.6116144373699086,
+      "tokens_seen": 1215823872
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006378590916385812,
+      "loss": 0.0728,
+      "theoretical_loss": 3.611543556623782,
+      "tokens_seen": 1216086016
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006377788476969989,
+      "loss": 0.0728,
+      "theoretical_loss": 3.6114726954324814,
+      "tokens_seen": 1216348160
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006376986037554164,
+      "loss": 0.0724,
+      "theoretical_loss": 3.6114018537864,
+      "tokens_seen": 1216610304
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006376183598138341,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6113310316759373,
+      "tokens_seen": 1216872448
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006375381158722517,
+      "loss": 0.0692,
+      "theoretical_loss": 3.6112602290914984,
+      "tokens_seen": 1217134592
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006374578719306693,
+      "loss": 0.0715,
+      "theoretical_loss": 3.611189446023497,
+      "tokens_seen": 1217396736
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006373776279890868,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6111186824623527,
+      "tokens_seen": 1217658880
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006372973840475045,
+      "loss": 0.0736,
+      "theoretical_loss": 3.6110479383984915,
+      "tokens_seen": 1217921024
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000637217140105922,
+      "loss": 0.073,
+      "theoretical_loss": 3.6109772138223466,
+      "tokens_seen": 1218183168
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006371368961643395,
+      "loss": 0.0736,
+      "theoretical_loss": 3.6109065087243577,
+      "tokens_seen": 1218445312
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006370566522227572,
+      "loss": 0.0714,
+      "theoretical_loss": 3.610835823094972,
+      "tokens_seen": 1218707456
+    },
+    {
+      "epoch": 0.37,
+      "objective/train/advantage_avg": 0.0002287530805915594,
+      "objective/train/docs_used": 445814,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5659476518630981,
+      "objective/train/original_loss": 1.5659475326538086,
+      "objective/train/theoretical_loss": 3.6107651569246424,
+      "objective/train/tokens_used": 1239429600,
+      "objective/train/value_avg": -0.01335906982421875,
+      "objective/train/value_loss": 0.0003773129137698561,
+      "objective/train/value_max": -0.00010150671005249023,
+      "objective/train/value_min": -0.476318359375,
+      "objective/train/value_reward_corr": 0.8288847136796991,
+      "objective/train/value_std": 0.0251617431640625,
+      "objective/train/weight_avg": 1.0004096031188965,
+      "objective/train/weighted_lm_loss": 1.5659234523773193,
+      "objective/train/weights_max": 1.2777711153030396,
+      "objective/train/weights_min": 0.5512197613716125,
+      "theoretical_loss": 3.6107651569246424,
+      "tokens_seen": 1218969600
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006369764082811747,
+      "loss": 0.073,
+      "theoretical_loss": 3.6107651569246424,
+      "tokens_seen": 1218969600
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006368961643395924,
+      "loss": 0.0712,
+      "theoretical_loss": 3.610694510203829,
+      "tokens_seen": 1219231744
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.00063681592039801,
+      "loss": 0.0717,
+      "theoretical_loss": 3.610623882922999,
+      "tokens_seen": 1219493888
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006367356764564276,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6105532750726255,
+      "tokens_seen": 1219756032
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006366554325148452,
+      "loss": 0.0718,
+      "theoretical_loss": 3.6104826866431887,
+      "tokens_seen": 1220018176
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006365751885732627,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6104121176251764,
+      "tokens_seen": 1220280320
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006364949446316803,
+      "loss": 0.0723,
+      "theoretical_loss": 3.6103415680090816,
+      "tokens_seen": 1220542464
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006364147006900979,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6102710377854046,
+      "tokens_seen": 1220804608
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006363344567485155,
+      "loss": 0.0705,
+      "theoretical_loss": 3.610200526944652,
+      "tokens_seen": 1221066752
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000636254212806933,
+      "loss": 0.0723,
+      "theoretical_loss": 3.610130035477339,
+      "tokens_seen": 1221328896
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006361739688653508,
+      "loss": 0.0753,
+      "theoretical_loss": 3.6100595633739854,
+      "tokens_seen": 1221591040
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006360937249237683,
+      "loss": 0.073,
+      "theoretical_loss": 3.6099891106251176,
+      "tokens_seen": 1221853184
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006360134809821858,
+      "loss": 0.0748,
+      "theoretical_loss": 3.6099186772212697,
+      "tokens_seen": 1222115328
+    },
+    {
+      "epoch": 0.37,
+      "objective/train/advantage_avg": 0.0006381941493600607,
+      "objective/train/docs_used": 446976,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.419157862663269,
+      "objective/train/original_loss": 1.419157862663269,
+      "objective/train/theoretical_loss": 3.6098834677707723,
+      "objective/train/tokens_used": 1242706400,
+      "objective/train/value_avg": -0.0130615234375,
+      "objective/train/value_loss": 0.0005473028286360204,
+      "objective/train/value_max": -0.00012242794036865234,
+      "objective/train/value_min": -0.96533203125,
+      "objective/train/value_reward_corr": 0.8783592914327736,
+      "objective/train/value_std": 0.036163330078125,
+      "objective/train/weight_avg": 1.0008801221847534,
+      "objective/train/weighted_lm_loss": 1.4213128089904785,
+      "objective/train/weights_max": 1.5615370273590088,
+      "objective/train/weights_min": 0.368775874376297,
+      "theoretical_loss": 3.6098834677707723,
+      "tokens_seen": 1222246400
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006359332370406035,
+      "loss": 0.0709,
+      "theoretical_loss": 3.6098482631529825,
+      "tokens_seen": 1222377472
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000635852993099021,
+      "loss": 0.0727,
+      "theoretical_loss": 3.6097778684108026,
+      "tokens_seen": 1222639616
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006357727491574386,
+      "loss": 0.0718,
+      "theoretical_loss": 3.609707492985284,
+      "tokens_seen": 1222901760
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006356925052158562,
+      "loss": 0.0713,
+      "theoretical_loss": 3.609637136866987,
+      "tokens_seen": 1223163904
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006356122612742738,
+      "loss": 0.0696,
+      "theoretical_loss": 3.609566800046478,
+      "tokens_seen": 1223426048
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006355320173326913,
+      "loss": 0.0711,
+      "theoretical_loss": 3.609496482514332,
+      "tokens_seen": 1223688192
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000635451773391109,
+      "loss": 0.0733,
+      "theoretical_loss": 3.609426184261128,
+      "tokens_seen": 1223950336
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006353715294495266,
+      "loss": 0.0735,
+      "theoretical_loss": 3.6093559052774538,
+      "tokens_seen": 1224212480
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006352912855079442,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6092856455539017,
+      "tokens_seen": 1224474624
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006352110415663618,
+      "loss": 0.0746,
+      "theoretical_loss": 3.6092154050810725,
+      "tokens_seen": 1224736768
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006351307976247793,
+      "loss": 0.073,
+      "theoretical_loss": 3.6091451838495727,
+      "tokens_seen": 1224998912
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000635050553683197,
+      "loss": 0.0714,
+      "theoretical_loss": 3.609074981850016,
+      "tokens_seen": 1225261056
+    },
+    {
+      "epoch": 0.37,
+      "objective/train/advantage_avg": 0.0011464846320450306,
+      "objective/train/docs_used": 448161,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4441373348236084,
+      "objective/train/original_loss": 1.4441375732421875,
+      "objective/train/theoretical_loss": 3.6090047990730216,
+      "objective/train/tokens_used": 1245983200,
+      "objective/train/value_avg": -0.0064849853515625,
+      "objective/train/value_loss": 0.00032763619674369693,
+      "objective/train/value_max": -0.00011235475540161133,
+      "objective/train/value_min": -0.94091796875,
+      "objective/train/value_reward_corr": 0.7932949888128629,
+      "objective/train/value_std": 0.0202789306640625,
+      "objective/train/weight_avg": 1.0012918710708618,
+      "objective/train/weighted_lm_loss": 1.4470592737197876,
+      "objective/train/weights_max": 1.4689112901687622,
+      "objective/train/weights_min": 0.4429895877838135,
+      "theoretical_loss": 3.6090047990730216,
+      "tokens_seen": 1225523200
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006349703097416145,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6090047990730216,
+      "tokens_seen": 1225523200
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000634890065800032,
+      "loss": 0.0712,
+      "theoretical_loss": 3.6089346355092164,
+      "tokens_seen": 1225785344
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006348098218584497,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6088644911492334,
+      "tokens_seen": 1226047488
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006347295779168672,
+      "loss": 0.0731,
+      "theoretical_loss": 3.6087943659837114,
+      "tokens_seen": 1226309632
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006346493339752849,
+      "loss": 0.0708,
+      "theoretical_loss": 3.6087242600032976,
+      "tokens_seen": 1226571776
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006345690900337025,
+      "loss": 0.0721,
+      "theoretical_loss": 3.6086541731986443,
+      "tokens_seen": 1226833920
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006344888460921201,
+      "loss": 0.0688,
+      "theoretical_loss": 3.6085841055604106,
+      "tokens_seen": 1227096064
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006344086021505376,
+      "loss": 0.0729,
+      "theoretical_loss": 3.608514057079262,
+      "tokens_seen": 1227358208
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006343283582089553,
+      "loss": 0.0711,
+      "theoretical_loss": 3.6084440277458714,
+      "tokens_seen": 1227620352
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006342481142673728,
+      "loss": 0.0703,
+      "theoretical_loss": 3.6083740175509176,
+      "tokens_seen": 1227882496
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006341678703257904,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6083040264850856,
+      "tokens_seen": 1228144640
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000634087626384208,
+      "loss": 0.0713,
+      "theoretical_loss": 3.6082340545390674,
+      "tokens_seen": 1228406784
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006340073824426255,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6081641017035615,
+      "tokens_seen": 1228668928
+    },
+    {
+      "epoch": 0.37,
+      "objective/train/advantage_avg": -6.965985812712461e-05,
+      "objective/train/docs_used": 449484,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5495264530181885,
+      "objective/train/original_loss": 1.5495264530181885,
+      "objective/train/theoretical_loss": 3.6081291324493456,
+      "objective/train/tokens_used": 1249260000,
+      "objective/train/value_avg": -0.00957489013671875,
+      "objective/train/value_loss": 0.0002603928733151406,
+      "objective/train/value_max": -9.03010368347168e-05,
+      "objective/train/value_min": -0.61083984375,
+      "objective/train/value_reward_corr": 0.7232506307301012,
+      "objective/train/value_std": 0.01611328125,
+      "objective/train/weight_avg": 1.000057578086853,
+      "objective/train/weighted_lm_loss": 1.5498452186584473,
+      "objective/train/weights_max": 1.781590461730957,
+      "objective/train/weights_min": 0.5239829421043396,
+      "theoretical_loss": 3.6081291324493456,
+      "tokens_seen": 1228800000
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006339271385010433,
+      "loss": 0.0749,
+      "theoretical_loss": 3.608094167969273,
+      "tokens_seen": 1228931072
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006338468945594608,
+      "loss": 0.0717,
+      "theoretical_loss": 3.6080242533269136,
+      "tokens_seen": 1229193216
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006337666506178784,
+      "loss": 0.0702,
+      "theoretical_loss": 3.6079543577672,
+      "tokens_seen": 1229455360
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000633686406676296,
+      "loss": 0.074,
+      "theoretical_loss": 3.6078844812808577,
+      "tokens_seen": 1229717504
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006336061627347135,
+      "loss": 0.0757,
+      "theoretical_loss": 3.6078146238586175,
+      "tokens_seen": 1229979648
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006335259187931311,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6077447854912164,
+      "tokens_seen": 1230241792
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006334456748515487,
+      "loss": 0.0721,
+      "theoretical_loss": 3.6076749661693985,
+      "tokens_seen": 1230503936
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006333654309099663,
+      "loss": 0.0743,
+      "theoretical_loss": 3.607605165883914,
+      "tokens_seen": 1230766080
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006332851869683838,
+      "loss": 0.0729,
+      "theoretical_loss": 3.60753538462552,
+      "tokens_seen": 1231028224
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006332049430268016,
+      "loss": 0.0719,
+      "theoretical_loss": 3.607465622384979,
+      "tokens_seen": 1231290368
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006331246990852191,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6073958791530614,
+      "tokens_seen": 1231552512
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006330444551436366,
+      "loss": 0.0727,
+      "theoretical_loss": 3.607326154920543,
+      "tokens_seen": 1231814656
+    },
+    {
+      "epoch": 0.37,
+      "objective/train/advantage_avg": 0.001238655881024897,
+      "objective/train/docs_used": 450737,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.530470371246338,
+      "objective/train/original_loss": 1.530470371246338,
+      "objective/train/theoretical_loss": 3.6072564496782062,
+      "objective/train/tokens_used": 1252536800,
+      "objective/train/value_avg": -0.00855255126953125,
+      "objective/train/value_loss": 0.00031318055698648095,
+      "objective/train/value_max": -0.00013136863708496094,
+      "objective/train/value_min": -0.65771484375,
+      "objective/train/value_reward_corr": 0.6789382313438648,
+      "objective/train/value_std": 0.01499176025390625,
+      "objective/train/weight_avg": 1.0013794898986816,
+      "objective/train/weighted_lm_loss": 1.5326346158981323,
+      "objective/train/weights_max": 1.1837249994277954,
+      "objective/train/weights_min": 0.37208810448646545,
+      "theoretical_loss": 3.6072564496782062,
+      "tokens_seen": 1232076800
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006329642112020543,
+      "loss": 0.0717,
+      "theoretical_loss": 3.6072564496782062,
+      "tokens_seen": 1232076800
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006328839672604718,
+      "loss": 0.0703,
+      "theoretical_loss": 3.607186763416841,
+      "tokens_seen": 1232338944
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006328037233188895,
+      "loss": 0.073,
+      "theoretical_loss": 3.607117096127242,
+      "tokens_seen": 1232601088
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000632723479377307,
+      "loss": 0.0743,
+      "theoretical_loss": 3.60704744780021,
+      "tokens_seen": 1232863232
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006326432354357246,
+      "loss": 0.072,
+      "theoretical_loss": 3.606977818426555,
+      "tokens_seen": 1233125376
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006325629914941422,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6069082079970913,
+      "tokens_seen": 1233387520
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006324827475525598,
+      "loss": 0.0741,
+      "theoretical_loss": 3.60683861650264,
+      "tokens_seen": 1233649664
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006324025036109774,
+      "loss": 0.0708,
+      "theoretical_loss": 3.606769043934027,
+      "tokens_seen": 1233911808
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000632322259669395,
+      "loss": 0.0736,
+      "theoretical_loss": 3.606699490282088,
+      "tokens_seen": 1234173952
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006322420157278126,
+      "loss": 0.07,
+      "theoretical_loss": 3.606629955537663,
+      "tokens_seen": 1234436096
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006321617717862301,
+      "loss": 0.0712,
+      "theoretical_loss": 3.606560439691598,
+      "tokens_seen": 1234698240
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006320815278446478,
+      "loss": 0.0723,
+      "theoretical_loss": 3.6064909427347462,
+      "tokens_seen": 1234960384
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006320012839030653,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6064214646579673,
+      "tokens_seen": 1235222528
+    },
+    {
+      "epoch": 0.37,
+      "objective/train/advantage_avg": -0.0005033229826949537,
+      "objective/train/docs_used": 451987,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4716780185699463,
+      "objective/train/original_loss": 1.4716781377792358,
+      "objective/train/theoretical_loss": 3.60638673269675,
+      "objective/train/tokens_used": 1255813600,
+      "objective/train/value_avg": -0.0082550048828125,
+      "objective/train/value_loss": 0.00027764757396653295,
+      "objective/train/value_max": -5.9664249420166016e-05,
+      "objective/train/value_min": -0.393798828125,
+      "objective/train/value_reward_corr": 0.6338732515401595,
+      "objective/train/value_std": 0.01284027099609375,
+      "objective/train/weight_avg": 0.9996262192726135,
+      "objective/train/weighted_lm_loss": 1.4702317714691162,
+      "objective/train/weights_max": 1.3029727935791016,
+      "objective/train/weights_min": 0.39436399936676025,
+      "theoretical_loss": 3.60638673269675,
+      "tokens_seen": 1235353600
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006319210399614828,
+      "loss": 0.0742,
+      "theoretical_loss": 3.606352005452126,
+      "tokens_seen": 1235484672
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006318407960199005,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6062825651080956,
+      "tokens_seen": 1235746816
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000631760552078318,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6062131436167544,
+      "tokens_seen": 1236008960
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006316803081367358,
+      "loss": 0.0716,
+      "theoretical_loss": 3.606143740968986,
+      "tokens_seen": 1236271104
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006316000641951533,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6060743571556833,
+      "tokens_seen": 1236533248
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006315198202535709,
+      "loss": 0.0741,
+      "theoretical_loss": 3.606004992167742,
+      "tokens_seen": 1236795392
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.0006314395763119885,
+      "loss": 0.0716,
+      "theoretical_loss": 3.605935645996067,
+      "tokens_seen": 1237057536
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 0.000631359332370406,
+      "loss": 0.0697,
+      "theoretical_loss": 3.605866318631568,
+      "tokens_seen": 1237319680
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006312790884288236,
+      "loss": 0.0713,
+      "theoretical_loss": 3.605797010065161,
+      "tokens_seen": 1237581824
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006311988444872412,
+      "loss": 0.0715,
+      "theoretical_loss": 3.6057277202877698,
+      "tokens_seen": 1237843968
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006311186005456588,
+      "loss": 0.072,
+      "theoretical_loss": 3.6056584492903223,
+      "tokens_seen": 1238106112
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006310383566040763,
+      "loss": 0.0715,
+      "theoretical_loss": 3.6055891970637544,
+      "tokens_seen": 1238368256
+    },
+    {
+      "epoch": 0.38,
+      "objective/train/advantage_avg": 0.0013174168998375535,
+      "objective/train/docs_used": 453261,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5999850034713745,
+      "objective/train/original_loss": 1.5999850034713745,
+      "objective/train/theoretical_loss": 3.605519963599008,
+      "objective/train/tokens_used": 1259090400,
+      "objective/train/value_avg": -0.0081024169921875,
+      "objective/train/value_loss": 0.00033738784259185195,
+      "objective/train/value_max": -8.028745651245117e-05,
+      "objective/train/value_min": -0.69580078125,
+      "objective/train/value_reward_corr": 0.677158113803972,
+      "objective/train/value_std": 0.01568603515625,
+      "objective/train/weight_avg": 1.0014675855636597,
+      "objective/train/weighted_lm_loss": 1.6025694608688354,
+      "objective/train/weights_max": 1.330289602279663,
+      "objective/train/weights_min": 0.3702276945114136,
+      "theoretical_loss": 3.605519963599008,
+      "tokens_seen": 1238630400
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006309581126624941,
+      "loss": 0.0749,
+      "theoretical_loss": 3.605519963599008,
+      "tokens_seen": 1238630400
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006308778687209116,
+      "loss": 0.0724,
+      "theoretical_loss": 3.60545074888703,
+      "tokens_seen": 1238892544
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006307976247793292,
+      "loss": 0.0715,
+      "theoretical_loss": 3.6053815529187756,
+      "tokens_seen": 1239154688
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006307173808377468,
+      "loss": 0.0739,
+      "theoretical_loss": 3.6053123756852052,
+      "tokens_seen": 1239416832
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006306371368961643,
+      "loss": 0.0709,
+      "theoretical_loss": 3.605243217177285,
+      "tokens_seen": 1239678976
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006305568929545819,
+      "loss": 0.0711,
+      "theoretical_loss": 3.6051740773859877,
+      "tokens_seen": 1239941120
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006304766490129995,
+      "loss": 0.0729,
+      "theoretical_loss": 3.605104956302293,
+      "tokens_seen": 1240203264
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006303964050714171,
+      "loss": 0.0729,
+      "theoretical_loss": 3.605035853917187,
+      "tokens_seen": 1240465408
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006303161611298347,
+      "loss": 0.0747,
+      "theoretical_loss": 3.604966770221661,
+      "tokens_seen": 1240727552
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006302359171882524,
+      "loss": 0.0697,
+      "theoretical_loss": 3.604897705206713,
+      "tokens_seen": 1240989696
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006301556732466699,
+      "loss": 0.0722,
+      "theoretical_loss": 3.6048286588633465,
+      "tokens_seen": 1241251840
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006300754293050875,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6047596311825725,
+      "tokens_seen": 1241513984
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006299951853635051,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6046906221554087,
+      "tokens_seen": 1241776128
+    },
+    {
+      "epoch": 0.38,
+      "objective/train/advantage_avg": 0.0009522174368612468,
+      "objective/train/docs_used": 454384,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.473624348640442,
+      "objective/train/original_loss": 1.4736244678497314,
+      "objective/train/theoretical_loss": 3.604656124634124,
+      "objective/train/tokens_used": 1262367200,
+      "objective/train/value_avg": -0.00734710693359375,
+      "objective/train/value_loss": 0.00023657429846934974,
+      "objective/train/value_max": -9.995698928833008e-05,
+      "objective/train/value_min": -0.265869140625,
+      "objective/train/value_reward_corr": 0.6545632899431301,
+      "objective/train/value_std": 0.01197052001953125,
+      "objective/train/weight_avg": 1.0010582208633423,
+      "objective/train/weighted_lm_loss": 1.4750497341156006,
+      "objective/train/weights_max": 1.304564356803894,
+      "objective/train/weights_min": 0.39513498544692993,
+      "theoretical_loss": 3.604656124634124,
+      "tokens_seen": 1241907200
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006299149414219226,
+      "loss": 0.0737,
+      "theoretical_loss": 3.6046216317728765,
+      "tokens_seen": 1242038272
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006298346974803403,
+      "loss": 0.0723,
+      "theoretical_loss": 3.6045526600260054,
+      "tokens_seen": 1242300416
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006297544535387578,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6044837069058318,
+      "tokens_seen": 1242562560
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006296742095971754,
+      "loss": 0.0751,
+      "theoretical_loss": 3.6044147724033957,
+      "tokens_seen": 1242824704
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.000629593965655593,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6043458565097453,
+      "tokens_seen": 1243086848
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006295137217140106,
+      "loss": 0.0701,
+      "theoretical_loss": 3.6042769592159356,
+      "tokens_seen": 1243348992
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006294334777724282,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6042080805130254,
+      "tokens_seen": 1243611136
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006293532338308458,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6041392203920815,
+      "tokens_seen": 1243873280
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006292729898892634,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6040703788441757,
+      "tokens_seen": 1244135424
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006291927459476809,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6040015558603877,
+      "tokens_seen": 1244397568
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006291125020060986,
+      "loss": 0.0731,
+      "theoretical_loss": 3.603932751431802,
+      "tokens_seen": 1244659712
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006290322580645161,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6038639655495093,
+      "tokens_seen": 1244921856
+    },
+    {
+      "epoch": 0.38,
+      "objective/train/advantage_avg": -0.0012195850722491741,
+      "objective/train/docs_used": 455531,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5041210651397705,
+      "objective/train/original_loss": 1.5041210651397705,
+      "objective/train/theoretical_loss": 3.6037951982046064,
+      "objective/train/tokens_used": 1265644000,
+      "objective/train/value_avg": -0.01177215576171875,
+      "objective/train/value_loss": 0.0004948603454977274,
+      "objective/train/value_max": -0.0001366138458251953,
+      "objective/train/value_min": -0.5390625,
+      "objective/train/value_reward_corr": 0.73076830896824,
+      "objective/train/value_std": 0.0195465087890625,
+      "objective/train/weight_avg": 0.9990051984786987,
+      "objective/train/weighted_lm_loss": 1.5013059377670288,
+      "objective/train/weights_max": 1.3366363048553467,
+      "objective/train/weights_min": 0.3697971701622009,
+      "theoretical_loss": 3.6037951982046064,
+      "tokens_seen": 1245184000
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006289520141229337,
+      "loss": 0.0726,
+      "theoretical_loss": 3.6037951982046064,
+      "tokens_seen": 1245184000
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006288717701813513,
+      "loss": 0.0729,
+      "theoretical_loss": 3.6037264493881973,
+      "tokens_seen": 1245446144
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006287915262397688,
+      "loss": 0.075,
+      "theoretical_loss": 3.603657719091391,
+      "tokens_seen": 1245708288
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006287112822981866,
+      "loss": 0.0732,
+      "theoretical_loss": 3.603589007305303,
+      "tokens_seen": 1245970432
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006286310383566041,
+      "loss": 0.0749,
+      "theoretical_loss": 3.603520314021055,
+      "tokens_seen": 1246232576
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006285507944150217,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6034516392297746,
+      "tokens_seen": 1246494720
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006284705504734393,
+      "loss": 0.0759,
+      "theoretical_loss": 3.6033829829225965,
+      "tokens_seen": 1246756864
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006283903065318568,
+      "loss": 0.0749,
+      "theoretical_loss": 3.6033143450906593,
+      "tokens_seen": 1247019008
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006283100625902744,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6032457257251105,
+      "tokens_seen": 1247281152
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.000628229818648692,
+      "loss": 0.0701,
+      "theoretical_loss": 3.603177124817101,
+      "tokens_seen": 1247543296
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006281495747071096,
+      "loss": 0.0725,
+      "theoretical_loss": 3.603108542357791,
+      "tokens_seen": 1247805440
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006280693307655271,
+      "loss": 0.0743,
+      "theoretical_loss": 3.603039978338343,
+      "tokens_seen": 1248067584
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006279890868239449,
+      "loss": 0.0714,
+      "theoretical_loss": 3.6029714327499285,
+      "tokens_seen": 1248329728
+    },
+    {
+      "epoch": 0.38,
+      "objective/train/advantage_avg": 0.00015995489957276732,
+      "objective/train/docs_used": 456712,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4048641920089722,
+      "objective/train/original_loss": 1.4048643112182617,
+      "objective/train/theoretical_loss": 3.602937166864601,
+      "objective/train/tokens_used": 1268920800,
+      "objective/train/value_avg": -0.00669097900390625,
+      "objective/train/value_loss": 0.00019214232452213764,
+      "objective/train/value_max": -0.00010639429092407227,
+      "objective/train/value_min": -0.24560546875,
+      "objective/train/value_reward_corr": 0.6105372464341148,
+      "objective/train/value_std": 0.00984954833984375,
+      "objective/train/weight_avg": 1.0002514123916626,
+      "objective/train/weighted_lm_loss": 1.405979871749878,
+      "objective/train/weights_max": 1.278395175933838,
+      "objective/train/weights_min": 0.4344214200973511,
+      "theoretical_loss": 3.602937166864601,
+      "tokens_seen": 1248460800
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006279088428823624,
+      "loss": 0.0752,
+      "theoretical_loss": 3.602902905583724,
+      "tokens_seen": 1248591872
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006278285989407801,
+      "loss": 0.0717,
+      "theoretical_loss": 3.602834396830912,
+      "tokens_seen": 1248854016
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006277483549991976,
+      "loss": 0.0726,
+      "theoretical_loss": 3.6027659064826816,
+      "tokens_seen": 1249116160
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006276681110576151,
+      "loss": 0.0744,
+      "theoretical_loss": 3.602697434530227,
+      "tokens_seen": 1249378304
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006275878671160328,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6026289809647487,
+      "tokens_seen": 1249640448
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006275076231744503,
+      "loss": 0.0741,
+      "theoretical_loss": 3.6025605457774548,
+      "tokens_seen": 1249902592
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006274273792328679,
+      "loss": 0.0721,
+      "theoretical_loss": 3.602492128959558,
+      "tokens_seen": 1250164736
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006273471352912855,
+      "loss": 0.0732,
+      "theoretical_loss": 3.6024237305022764,
+      "tokens_seen": 1250426880
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006272668913497032,
+      "loss": 0.0741,
+      "theoretical_loss": 3.602355350396836,
+      "tokens_seen": 1250689024
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006271866474081207,
+      "loss": 0.072,
+      "theoretical_loss": 3.602286988634467,
+      "tokens_seen": 1250951168
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006271064034665383,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6022186452064076,
+      "tokens_seen": 1251213312
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006270261595249559,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6021503201039,
+      "tokens_seen": 1251475456
+    },
+    {
+      "epoch": 0.38,
+      "objective/train/advantage_avg": 0.0013131473679095507,
+      "objective/train/docs_used": 458046,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4418436288833618,
+      "objective/train/original_loss": 1.4418433904647827,
+      "objective/train/theoretical_loss": 3.6020820133181934,
+      "objective/train/tokens_used": 1272197600,
+      "objective/train/value_avg": -0.0069580078125,
+      "objective/train/value_loss": 0.00013802826288156211,
+      "objective/train/value_max": -7.486343383789062e-05,
+      "objective/train/value_min": -0.235595703125,
+      "objective/train/value_reward_corr": 0.722366365333374,
+      "objective/train/value_std": 0.0119781494140625,
+      "objective/train/weight_avg": 1.0013810396194458,
+      "objective/train/weighted_lm_loss": 1.4443731307983398,
+      "objective/train/weights_max": 1.2093919515609741,
+      "objective/train/weights_min": 0.6394911408424377,
+      "theoretical_loss": 3.6020820133181934,
+      "tokens_seen": 1251737600
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006269459155833734,
+      "loss": 0.071,
+      "theoretical_loss": 3.6020820133181934,
+      "tokens_seen": 1251737600
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006268656716417911,
+      "loss": 0.0744,
+      "theoretical_loss": 3.6020137248405435,
+      "tokens_seen": 1251999744
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006267854277002086,
+      "loss": 0.0733,
+      "theoretical_loss": 3.6019454546622107,
+      "tokens_seen": 1252261888
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006267051837586262,
+      "loss": 0.0711,
+      "theoretical_loss": 3.601877202774463,
+      "tokens_seen": 1252524032
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006266249398170438,
+      "loss": 0.0729,
+      "theoretical_loss": 3.601808969168573,
+      "tokens_seen": 1252786176
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006265446958754613,
+      "loss": 0.0719,
+      "theoretical_loss": 3.60174075383582,
+      "tokens_seen": 1253048320
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006264644519338791,
+      "loss": 0.0752,
+      "theoretical_loss": 3.6016725567674883,
+      "tokens_seen": 1253310464
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006263842079922966,
+      "loss": 0.0728,
+      "theoretical_loss": 3.60160437795487,
+      "tokens_seen": 1253572608
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006263039640507142,
+      "loss": 0.073,
+      "theoretical_loss": 3.6015362173892624,
+      "tokens_seen": 1253834752
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006262237201091318,
+      "loss": 0.0745,
+      "theoretical_loss": 3.6014680750619674,
+      "tokens_seen": 1254096896
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006261434761675494,
+      "loss": 0.0725,
+      "theoretical_loss": 3.6013999509642947,
+      "tokens_seen": 1254359040
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006260632322259669,
+      "loss": 0.0724,
+      "theoretical_loss": 3.601331845087559,
+      "tokens_seen": 1254621184
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006259829882843845,
+      "loss": 0.073,
+      "theoretical_loss": 3.601263757423082,
+      "tokens_seen": 1254883328
+    },
+    {
+      "epoch": 0.38,
+      "objective/train/advantage_avg": 0.0014171048533171415,
+      "objective/train/docs_used": 459153,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5080225467681885,
+      "objective/train/original_loss": 1.508022427558899,
+      "objective/train/theoretical_loss": 3.601229720417729,
+      "objective/train/tokens_used": 1275474400,
+      "objective/train/value_avg": -0.00691986083984375,
+      "objective/train/value_loss": 0.00018681980145629495,
+      "objective/train/value_max": -0.00014090538024902344,
+      "objective/train/value_min": -0.630859375,
+      "objective/train/value_reward_corr": 0.6682120767318891,
+      "objective/train/value_std": 0.0105743408203125,
+      "objective/train/weight_avg": 1.001495361328125,
+      "objective/train/weighted_lm_loss": 1.5094605684280396,
+      "objective/train/weights_max": 1.1394519805908203,
+      "objective/train/weights_min": 0.22579480707645416,
+      "theoretical_loss": 3.601229720417729,
+      "tokens_seen": 1255014400
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006259027443428021,
+      "loss": 0.0728,
+      "theoretical_loss": 3.601195687962189,
+      "tokens_seen": 1255145472
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006258225004012196,
+      "loss": 0.0718,
+      "theoretical_loss": 3.601127636696214,
+      "tokens_seen": 1255407616
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006257422564596374,
+      "loss": 0.0727,
+      "theoretical_loss": 3.6010596036164957,
+      "tokens_seen": 1255669760
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006256620125180549,
+      "loss": 0.0725,
+      "theoretical_loss": 3.600991588714378,
+      "tokens_seen": 1255931904
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006255817685764725,
+      "loss": 0.0705,
+      "theoretical_loss": 3.6009235919812124,
+      "tokens_seen": 1256194048
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006255015246348901,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6008556134083545,
+      "tokens_seen": 1256456192
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006254212806933076,
+      "loss": 0.0732,
+      "theoretical_loss": 3.600787652987167,
+      "tokens_seen": 1256718336
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006253410367517252,
+      "loss": 0.0716,
+      "theoretical_loss": 3.6007197107090185,
+      "tokens_seen": 1256980480
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006252607928101428,
+      "loss": 0.0735,
+      "theoretical_loss": 3.600651786565283,
+      "tokens_seen": 1257242624
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006251805488685604,
+      "loss": 0.0719,
+      "theoretical_loss": 3.6005838805473402,
+      "tokens_seen": 1257504768
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.000625100304926978,
+      "loss": 0.0717,
+      "theoretical_loss": 3.600515992646577,
+      "tokens_seen": 1257766912
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006250200609853957,
+      "loss": 0.072,
+      "theoretical_loss": 3.6004481228543854,
+      "tokens_seen": 1258029056
+    },
+    {
+      "epoch": 0.38,
+      "objective/train/advantage_avg": 0.0018183041829615831,
+      "objective/train/docs_used": 460354,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.566368579864502,
+      "objective/train/original_loss": 1.5663683414459229,
+      "objective/train/theoretical_loss": 3.600380271162162,
+      "objective/train/tokens_used": 1278751200,
+      "objective/train/value_avg": -0.01047515869140625,
+      "objective/train/value_loss": 0.0005354974418878555,
+      "objective/train/value_max": -0.00010639429092407227,
+      "objective/train/value_min": -0.8798828125,
+      "objective/train/value_reward_corr": 0.6700587084228133,
+      "objective/train/value_std": 0.019775390625,
+      "objective/train/weight_avg": 1.0020546913146973,
+      "objective/train/weighted_lm_loss": 1.5695098638534546,
+      "objective/train/weights_max": 1.7010573148727417,
+      "objective/train/weights_min": 0.3695588707923889,
+      "theoretical_loss": 3.600380271162162,
+      "tokens_seen": 1258291200
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006249398170438132,
+      "loss": 0.071,
+      "theoretical_loss": 3.600380271162162,
+      "tokens_seen": 1258291200
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006248595731022309,
+      "loss": 0.0693,
+      "theoretical_loss": 3.6003124375613114,
+      "tokens_seen": 1258553344
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006247793291606484,
+      "loss": 0.0747,
+      "theoretical_loss": 3.6002446220432427,
+      "tokens_seen": 1258815488
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006246990852190659,
+      "loss": 0.074,
+      "theoretical_loss": 3.6001768245993713,
+      "tokens_seen": 1259077632
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006246188412774836,
+      "loss": 0.0731,
+      "theoretical_loss": 3.600109045221119,
+      "tokens_seen": 1259339776
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006245385973359011,
+      "loss": 0.0748,
+      "theoretical_loss": 3.600041283899912,
+      "tokens_seen": 1259601920
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006244583533943187,
+      "loss": 0.0714,
+      "theoretical_loss": 3.599973540627184,
+      "tokens_seen": 1259864064
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006243781094527363,
+      "loss": 0.0741,
+      "theoretical_loss": 3.599905815394374,
+      "tokens_seen": 1260126208
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.000624297865511154,
+      "loss": 0.0736,
+      "theoretical_loss": 3.599838108192926,
+      "tokens_seen": 1260388352
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006242176215695715,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5997704190142907,
+      "tokens_seen": 1260650496
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006241373776279891,
+      "loss": 0.0741,
+      "theoretical_loss": 3.599702747849925,
+      "tokens_seen": 1260912640
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006240571336864067,
+      "loss": 0.073,
+      "theoretical_loss": 3.5996350946912896,
+      "tokens_seen": 1261174784
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006239768897448243,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5995674595298537,
+      "tokens_seen": 1261436928
+    },
+    {
+      "epoch": 0.38,
+      "objective/train/advantage_avg": 0.0015806846786290407,
+      "objective/train/docs_used": 461530,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3945951461791992,
+      "objective/train/original_loss": 1.3945950269699097,
+      "objective/train/theoretical_loss": 3.599533648695421,
+      "objective/train/tokens_used": 1282028000,
+      "objective/train/value_avg": -0.006893157958984375,
+      "objective/train/value_loss": 0.00015965731290634722,
+      "objective/train/value_max": -7.31348991394043e-05,
+      "objective/train/value_min": -0.7705078125,
+      "objective/train/value_reward_corr": 0.6894917108257843,
+      "objective/train/value_std": 0.0124969482421875,
+      "objective/train/weight_avg": 1.0016539096832275,
+      "objective/train/weighted_lm_loss": 1.39777410030365,
+      "objective/train/weights_max": 1.2735666036605835,
+      "objective/train/weights_min": 0.36959269642829895,
+      "theoretical_loss": 3.599533648695421,
+      "tokens_seen": 1261568000
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006238966458032419,
+      "loss": 0.072,
+      "theoretical_loss": 3.5994998423570914,
+      "tokens_seen": 1261699072
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006238164018616594,
+      "loss": 0.0713,
+      "theoretical_loss": 3.599432243164481,
+      "tokens_seen": 1261961216
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006237361579200771,
+      "loss": 0.0714,
+      "theoretical_loss": 3.599364661943509,
+      "tokens_seen": 1262223360
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006236559139784946,
+      "loss": 0.0726,
+      "theoretical_loss": 3.599297098685666,
+      "tokens_seen": 1262485504
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006235756700369121,
+      "loss": 0.071,
+      "theoretical_loss": 3.599229553382449,
+      "tokens_seen": 1262747648
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006234954260953299,
+      "loss": 0.0725,
+      "theoretical_loss": 3.599162026025361,
+      "tokens_seen": 1263009792
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006234151821537474,
+      "loss": 0.0706,
+      "theoretical_loss": 3.599094516605911,
+      "tokens_seen": 1263271936
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.000623334938212165,
+      "loss": 0.0737,
+      "theoretical_loss": 3.5990270251156122,
+      "tokens_seen": 1263534080
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006232546942705826,
+      "loss": 0.0748,
+      "theoretical_loss": 3.5989595515459856,
+      "tokens_seen": 1263796224
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006231744503290002,
+      "loss": 0.0743,
+      "theoretical_loss": 3.598892095888557,
+      "tokens_seen": 1264058368
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006230942063874177,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5988246581348573,
+      "tokens_seen": 1264320512
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006230139624458353,
+      "loss": 0.073,
+      "theoretical_loss": 3.5987572382764252,
+      "tokens_seen": 1264582656
+    },
+    {
+      "epoch": 0.38,
+      "objective/train/advantage_avg": 7.375935092568398e-05,
+      "objective/train/docs_used": 462829,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4854201078414917,
+      "objective/train/original_loss": 1.4854199886322021,
+      "objective/train/theoretical_loss": 3.5986898363048025,
+      "objective/train/tokens_used": 1285304800,
+      "objective/train/value_avg": -0.0058441162109375,
+      "objective/train/value_loss": 0.00019366215565241873,
+      "objective/train/value_max": -0.00012242794036865234,
+      "objective/train/value_min": -0.383544921875,
+      "objective/train/value_reward_corr": 0.5646258964894881,
+      "objective/train/value_std": 0.00856781005859375,
+      "objective/train/weight_avg": 1.0001623630523682,
+      "objective/train/weighted_lm_loss": 1.4864187240600586,
+      "objective/train/weights_max": 1.2285127639770508,
+      "objective/train/weights_min": 0.39554521441459656,
+      "theoretical_loss": 3.5986898363048025,
+      "tokens_seen": 1264844800
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006229337185042529,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5986898363048025,
+      "tokens_seen": 1264844800
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006228534745626704,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5986224522115395,
+      "tokens_seen": 1265106944
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006227732306210882,
+      "loss": 0.0704,
+      "theoretical_loss": 3.59855508598819,
+      "tokens_seen": 1265369088
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006226929866795057,
+      "loss": 0.0737,
+      "theoretical_loss": 3.5984877376263142,
+      "tokens_seen": 1265631232
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006226127427379234,
+      "loss": 0.07,
+      "theoretical_loss": 3.5984204071174783,
+      "tokens_seen": 1265893376
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006225324987963409,
+      "loss": 0.0763,
+      "theoretical_loss": 3.598353094453255,
+      "tokens_seen": 1266155520
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006224522548547584,
+      "loss": 0.0724,
+      "theoretical_loss": 3.598285799625221,
+      "tokens_seen": 1266417664
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006223720109131761,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5982185226249594,
+      "tokens_seen": 1266679808
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006222917669715936,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5981512634440604,
+      "tokens_seen": 1266941952
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006222115230300112,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5980840220741177,
+      "tokens_seen": 1267204096
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006221312790884289,
+      "loss": 0.0748,
+      "theoretical_loss": 3.5980167985067317,
+      "tokens_seen": 1267466240
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006220510351468465,
+      "loss": 0.0739,
+      "theoretical_loss": 3.5979495927335092,
+      "tokens_seen": 1267728384
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.000621970791205264,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5978824047460614,
+      "tokens_seen": 1267990528
+    },
+    {
+      "epoch": 0.38,
+      "objective/train/advantage_avg": 0.0012868504272773862,
+      "objective/train/docs_used": 464032,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5202511548995972,
+      "objective/train/original_loss": 1.5202511548995972,
+      "objective/train/theoretical_loss": 3.5978488174193832,
+      "objective/train/tokens_used": 1288581600,
+      "objective/train/value_avg": -0.00853729248046875,
+      "objective/train/value_loss": 0.0002907871676143259,
+      "objective/train/value_max": -0.00011771917343139648,
+      "objective/train/value_min": -0.3349609375,
+      "objective/train/value_reward_corr": 0.5771171767033785,
+      "objective/train/value_std": 0.013885498046875,
+      "objective/train/weight_avg": 1.0014193058013916,
+      "objective/train/weighted_lm_loss": 1.5222471952438354,
+      "objective/train/weights_max": 1.397885799407959,
+      "objective/train/weights_min": 0.37770265340805054,
+      "theoretical_loss": 3.5978488174193832,
+      "tokens_seen": 1268121600
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006218905472636816,
+      "loss": 0.0733,
+      "theoretical_loss": 3.5978152345360055,
+      "tokens_seen": 1268252672
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006218103033220992,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5977480820949657,
+      "tokens_seen": 1268514816
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006217300593805167,
+      "loss": 0.0741,
+      "theoretical_loss": 3.5976809474145703,
+      "tokens_seen": 1268776960
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006216498154389344,
+      "loss": 0.0731,
+      "theoretical_loss": 3.597613830486454,
+      "tokens_seen": 1269039104
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006215695714973519,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5975467313022564,
+      "tokens_seen": 1269301248
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006214893275557696,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5974796498536237,
+      "tokens_seen": 1269563392
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006214090836141871,
+      "loss": 0.0708,
+      "theoretical_loss": 3.597412586132208,
+      "tokens_seen": 1269825536
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006213288396726048,
+      "loss": 0.0728,
+      "theoretical_loss": 3.5973455401296652,
+      "tokens_seen": 1270087680
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 0.0006212485957310224,
+      "loss": 0.07,
+      "theoretical_loss": 3.597278511837659,
+      "tokens_seen": 1270349824
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006211683517894399,
+      "loss": 0.0716,
+      "theoretical_loss": 3.597211501247858,
+      "tokens_seen": 1270611968
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006210881078478575,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5971445083519358,
+      "tokens_seen": 1270874112
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006210078639062751,
+      "loss": 0.0733,
+      "theoretical_loss": 3.5970775331415723,
+      "tokens_seen": 1271136256
+    },
+    {
+      "epoch": 0.39,
+      "objective/train/advantage_avg": 0.0005056922673247755,
+      "objective/train/docs_used": 465264,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4634767770767212,
+      "objective/train/original_loss": 1.4634767770767212,
+      "objective/train/theoretical_loss": 3.5970105756084525,
+      "objective/train/tokens_used": 1291858400,
+      "objective/train/value_avg": -0.007724761962890625,
+      "objective/train/value_loss": 0.00021731469314545393,
+      "objective/train/value_max": -0.00010639429092407227,
+      "objective/train/value_min": -0.34912109375,
+      "objective/train/value_reward_corr": 0.6817886298040029,
+      "objective/train/value_std": 0.016876220703125,
+      "objective/train/weight_avg": 1.0006133317947388,
+      "objective/train/weighted_lm_loss": 1.4629930257797241,
+      "objective/train/weights_max": 1.2334792613983154,
+      "objective/train/weights_min": 0.729193389415741,
+      "theoretical_loss": 3.5970105756084525,
+      "tokens_seen": 1271398400
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006209276199646927,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5970105756084525,
+      "tokens_seen": 1271398400
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006208473760231102,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5969436357442683,
+      "tokens_seen": 1271660544
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006207671320815279,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5968767135407154,
+      "tokens_seen": 1271922688
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006206868881399454,
+      "loss": 0.0712,
+      "theoretical_loss": 3.596809808989496,
+      "tokens_seen": 1272184832
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000620606644198363,
+      "loss": 0.0728,
+      "theoretical_loss": 3.596742922082319,
+      "tokens_seen": 1272446976
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006205264002567807,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5966760528108965,
+      "tokens_seen": 1272709120
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006204461563151982,
+      "loss": 0.0726,
+      "theoretical_loss": 3.596609201166948,
+      "tokens_seen": 1272971264
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006203659123736158,
+      "loss": 0.0726,
+      "theoretical_loss": 3.596542367142198,
+      "tokens_seen": 1273233408
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006202856684320334,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5964755507283774,
+      "tokens_seen": 1273495552
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000620205424490451,
+      "loss": 0.0736,
+      "theoretical_loss": 3.5964087519172203,
+      "tokens_seen": 1273757696
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006201251805488686,
+      "loss": 0.0745,
+      "theoretical_loss": 3.59634197070047,
+      "tokens_seen": 1274019840
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006200449366072861,
+      "loss": 0.0751,
+      "theoretical_loss": 3.5962752070698727,
+      "tokens_seen": 1274281984
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006199646926657037,
+      "loss": 0.0694,
+      "theoretical_loss": 3.59620846101718,
+      "tokens_seen": 1274544128
+    },
+    {
+      "epoch": 0.39,
+      "objective/train/advantage_avg": 0.0023682687897235155,
+      "objective/train/docs_used": 466434,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4091410636901855,
+      "objective/train/original_loss": 1.4091408252716064,
+      "objective/train/theoretical_loss": 3.5961750945799724,
+      "objective/train/tokens_used": 1295135200,
+      "objective/train/value_avg": -0.012420654296875,
+      "objective/train/value_loss": 0.00021746121637988836,
+      "objective/train/value_max": -0.00011414289474487305,
+      "objective/train/value_min": -0.407958984375,
+      "objective/train/value_reward_corr": 0.8011572186934655,
+      "objective/train/value_std": 0.0207366943359375,
+      "objective/train/weight_avg": 1.002469778060913,
+      "objective/train/weighted_lm_loss": 1.4116214513778687,
+      "objective/train/weights_max": 1.1912201642990112,
+      "objective/train/weights_min": 0.3865475058555603,
+      "theoretical_loss": 3.5961750945799724,
+      "tokens_seen": 1274675200
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006198844487241214,
+      "loss": 0.0727,
+      "theoretical_loss": 3.596141732534151,
+      "tokens_seen": 1274806272
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000619804204782539,
+      "loss": 0.0749,
+      "theoretical_loss": 3.596075021612549,
+      "tokens_seen": 1275068416
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006197239608409565,
+      "loss": 0.0739,
+      "theoretical_loss": 3.5960083282441433,
+      "tokens_seen": 1275330560
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006196437168993742,
+      "loss": 0.0741,
+      "theoretical_loss": 3.5959416524207084,
+      "tokens_seen": 1275592704
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006195634729577917,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5958749941340242,
+      "tokens_seen": 1275854848
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006194832290162092,
+      "loss": 0.0735,
+      "theoretical_loss": 3.595808353375877,
+      "tokens_seen": 1276116992
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006194029850746269,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5957417301380588,
+      "tokens_seen": 1276379136
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006193227411330444,
+      "loss": 0.0735,
+      "theoretical_loss": 3.595675124412365,
+      "tokens_seen": 1276641280
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000619242497191462,
+      "loss": 0.0762,
+      "theoretical_loss": 3.595608536190599,
+      "tokens_seen": 1276903424
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006191622532498797,
+      "loss": 0.0723,
+      "theoretical_loss": 3.595541965464568,
+      "tokens_seen": 1277165568
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006190820093082973,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5954754122260866,
+      "tokens_seen": 1277427712
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006190017653667148,
+      "loss": 0.074,
+      "theoretical_loss": 3.595408876466972,
+      "tokens_seen": 1277689856
+    },
+    {
+      "epoch": 0.39,
+      "objective/train/advantage_avg": -0.0013968882849439979,
+      "objective/train/docs_used": 467639,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.402967095375061,
+      "objective/train/original_loss": 1.4029669761657715,
+      "objective/train/theoretical_loss": 3.59534235817905,
+      "objective/train/tokens_used": 1298412000,
+      "objective/train/value_avg": -0.01238250732421875,
+      "objective/train/value_loss": 0.0012696925550699234,
+      "objective/train/value_max": -0.00010150671005249023,
+      "objective/train/value_min": -0.794921875,
+      "objective/train/value_reward_corr": 0.703285806535258,
+      "objective/train/value_std": 0.029815673828125,
+      "objective/train/weight_avg": 0.9991534948348999,
+      "objective/train/weighted_lm_loss": 1.3983840942382812,
+      "objective/train/weights_max": 2.1239025592803955,
+      "objective/train/weights_min": 0.3922275900840759,
+      "theoretical_loss": 3.59534235817905,
+      "tokens_seen": 1277952000
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006189215214251324,
+      "loss": 0.0727,
+      "theoretical_loss": 3.59534235817905,
+      "tokens_seen": 1277952000
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.00061884127748355,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5952758573541503,
+      "tokens_seen": 1278214144
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006187610335419676,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5952093739841073,
+      "tokens_seen": 1278476288
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006186807896003852,
+      "loss": 0.0743,
+      "theoretical_loss": 3.5951429080607635,
+      "tokens_seen": 1278738432
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006186005456588027,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5950764595759646,
+      "tokens_seen": 1279000576
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006185203017172204,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5950100285215623,
+      "tokens_seen": 1279262720
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000618440057775638,
+      "loss": 0.0705,
+      "theoretical_loss": 3.594943614889414,
+      "tokens_seen": 1279524864
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006183598138340556,
+      "loss": 0.0746,
+      "theoretical_loss": 3.5948772186713827,
+      "tokens_seen": 1279787008
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006182795698924732,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5948108398593357,
+      "tokens_seen": 1280049152
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006181993259508907,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5947444784451488,
+      "tokens_seen": 1280311296
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006181190820093083,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5946781344206995,
+      "tokens_seen": 1280573440
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006180388380677259,
+      "loss": 0.0745,
+      "theoretical_loss": 3.5946118077778735,
+      "tokens_seen": 1280835584
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006179585941261435,
+      "loss": 0.0758,
+      "theoretical_loss": 3.59454549850856,
+      "tokens_seen": 1281097728
+    },
+    {
+      "epoch": 0.39,
+      "objective/train/advantage_avg": 0.0009923300240188837,
+      "objective/train/docs_used": 468763,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6140085458755493,
+      "objective/train/original_loss": 1.6140084266662598,
+      "objective/train/theoretical_loss": 3.594512350386438,
+      "objective/train/tokens_used": 1301688800,
+      "objective/train/value_avg": -0.007724761962890625,
+      "objective/train/value_loss": 0.00035045703407377005,
+      "objective/train/value_max": -8.684396743774414e-05,
+      "objective/train/value_min": -0.94189453125,
+      "objective/train/value_reward_corr": 0.6209565433413602,
+      "objective/train/value_std": 0.01407623291015625,
+      "objective/train/weight_avg": 1.0011414289474487,
+      "objective/train/weighted_lm_loss": 1.615778923034668,
+      "objective/train/weights_max": 1.344037413597107,
+      "objective/train/weights_min": 0.36872875690460205,
+      "theoretical_loss": 3.594512350386438,
+      "tokens_seen": 1281228800
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000617878350184561,
+      "loss": 0.0737,
+      "theoretical_loss": 3.594479206604655,
+      "tokens_seen": 1281359872
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006177981062429787,
+      "loss": 0.0747,
+      "theoretical_loss": 3.59441293205806,
+      "tokens_seen": 1281622016
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006177178623013962,
+      "loss": 0.0724,
+      "theoretical_loss": 3.59434667486068,
+      "tokens_seen": 1281884160
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006176376183598139,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5942804350044284,
+      "tokens_seen": 1282146304
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006175573744182315,
+      "loss": 0.0724,
+      "theoretical_loss": 3.594214212481222,
+      "tokens_seen": 1282408448
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000617477130476649,
+      "loss": 0.0702,
+      "theoretical_loss": 3.594148007282983,
+      "tokens_seen": 1282670592
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006173968865350667,
+      "loss": 0.0671,
+      "theoretical_loss": 3.59408181940164,
+      "tokens_seen": 1282932736
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006173166425934842,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5940156488291266,
+      "tokens_seen": 1283194880
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006172363986519018,
+      "loss": 0.0703,
+      "theoretical_loss": 3.593949495557381,
+      "tokens_seen": 1283457024
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006171561547103194,
+      "loss": 0.072,
+      "theoretical_loss": 3.5938833595783484,
+      "tokens_seen": 1283719168
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006170759107687369,
+      "loss": 0.0722,
+      "theoretical_loss": 3.593817240883978,
+      "tokens_seen": 1283981312
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006169956668271545,
+      "loss": 0.072,
+      "theoretical_loss": 3.5937511394662254,
+      "tokens_seen": 1284243456
+    },
+    {
+      "epoch": 0.39,
+      "objective/train/advantage_avg": 0.0013025725493207574,
+      "objective/train/docs_used": 470107,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4686199426651,
+      "objective/train/original_loss": 1.4686200618743896,
+      "objective/train/theoretical_loss": 3.5936850553170503,
+      "objective/train/tokens_used": 1304965600,
+      "objective/train/value_avg": -0.005847930908203125,
+      "objective/train/value_loss": 9.238809434464201e-05,
+      "objective/train/value_max": -8.958578109741211e-05,
+      "objective/train/value_min": -0.5205078125,
+      "objective/train/value_reward_corr": 0.705613709892375,
+      "objective/train/value_std": 0.01006317138671875,
+      "objective/train/weight_avg": 1.0013480186462402,
+      "objective/train/weighted_lm_loss": 1.4713242053985596,
+      "objective/train/weights_max": 1.1862255334854126,
+      "objective/train/weights_min": 0.6386135220527649,
+      "theoretical_loss": 3.5936850553170503,
+      "tokens_seen": 1284505600
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006169154228855722,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5936850553170503,
+      "tokens_seen": 1284505600
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006168351789439898,
+      "loss": 0.0736,
+      "theoretical_loss": 3.593618988428419,
+      "tokens_seen": 1284767744
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006167549350024073,
+      "loss": 0.0701,
+      "theoretical_loss": 3.593552938792303,
+      "tokens_seen": 1285029888
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000616674691060825,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5934869064006785,
+      "tokens_seen": 1285292032
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006165944471192425,
+      "loss": 0.0734,
+      "theoretical_loss": 3.593420891245527,
+      "tokens_seen": 1285554176
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.00061651420317766,
+      "loss": 0.0748,
+      "theoretical_loss": 3.5933548933188377,
+      "tokens_seen": 1285816320
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006164339592360777,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5932889126126017,
+      "tokens_seen": 1286078464
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006163537152944952,
+      "loss": 0.0739,
+      "theoretical_loss": 3.5932229491188172,
+      "tokens_seen": 1286340608
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006162734713529129,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5931570028294884,
+      "tokens_seen": 1286602752
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006161932274113304,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5930910737366233,
+      "tokens_seen": 1286864896
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006161129834697481,
+      "loss": 0.0729,
+      "theoretical_loss": 3.5930251618322364,
+      "tokens_seen": 1287127040
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006160327395281657,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5929592671083466,
+      "tokens_seen": 1287389184
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006159524955865832,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5928933895569792,
+      "tokens_seen": 1287651328
+    },
+    {
+      "epoch": 0.39,
+      "objective/train/advantage_avg": -0.0021324241533875465,
+      "objective/train/docs_used": 471101,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3588120937347412,
+      "objective/train/original_loss": 1.358811855316162,
+      "objective/train/theoretical_loss": 3.592860457218501,
+      "objective/train/tokens_used": 1308242400,
+      "objective/train/value_avg": -0.01183319091796875,
+      "objective/train/value_loss": 0.0003357592795509845,
+      "objective/train/value_max": -0.0001596212387084961,
+      "objective/train/value_min": -0.321044921875,
+      "objective/train/value_reward_corr": 0.8440921060828028,
+      "objective/train/value_std": 0.02423095703125,
+      "objective/train/weight_avg": 0.9980318546295166,
+      "objective/train/weighted_lm_loss": 1.3537436723709106,
+      "objective/train/weights_max": 1.2598665952682495,
+      "objective/train/weights_min": 0.7315114736557007,
+      "theoretical_loss": 3.592860457218501,
+      "tokens_seen": 1287782400
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006158722516450008,
+      "loss": 0.07,
+      "theoretical_loss": 3.5928275291701643,
+      "tokens_seen": 1287913472
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006157920077034184,
+      "loss": 0.0711,
+      "theoretical_loss": 3.592761685939937,
+      "tokens_seen": 1288175616
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000615711763761836,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5926958598583383,
+      "tokens_seen": 1288437760
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006156315198202535,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5926300509174136,
+      "tokens_seen": 1288699904
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006155512758786712,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5925642591092153,
+      "tokens_seen": 1288962048
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006154710319370887,
+      "loss": 0.0696,
+      "theoretical_loss": 3.592498484425799,
+      "tokens_seen": 1289224192
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006153907879955064,
+      "loss": 0.0737,
+      "theoretical_loss": 3.5924327268592267,
+      "tokens_seen": 1289486336
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000615310544053924,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5923669864015664,
+      "tokens_seen": 1289748480
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006152303001123415,
+      "loss": 0.0699,
+      "theoretical_loss": 3.59230126304489,
+      "tokens_seen": 1290010624
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006151500561707592,
+      "loss": 0.0733,
+      "theoretical_loss": 3.592235556781276,
+      "tokens_seen": 1290272768
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006150698122291767,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5921698676028058,
+      "tokens_seen": 1290534912
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006149895682875943,
+      "loss": 0.0683,
+      "theoretical_loss": 3.59210419550157,
+      "tokens_seen": 1290797056
+    },
+    {
+      "epoch": 0.39,
+      "objective/train/advantage_avg": 0.0010878384346142411,
+      "objective/train/docs_used": 472355,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5228557586669922,
+      "objective/train/original_loss": 1.5228557586669922,
+      "objective/train/theoretical_loss": 3.5920385404696606,
+      "objective/train/tokens_used": 1311519200,
+      "objective/train/value_avg": -0.0082855224609375,
+      "objective/train/value_loss": 0.00018637104949448258,
+      "objective/train/value_max": -0.00010973215103149414,
+      "objective/train/value_min": -0.97314453125,
+      "objective/train/value_reward_corr": 0.8474394273861514,
+      "objective/train/value_std": 0.019683837890625,
+      "objective/train/weight_avg": 1.0011759996414185,
+      "objective/train/weighted_lm_loss": 1.5245059728622437,
+      "objective/train/weights_max": 1.336806058883667,
+      "objective/train/weights_min": 0.5558143258094788,
+      "theoretical_loss": 3.5920385404696606,
+      "tokens_seen": 1291059200
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006149093243460119,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5920385404696606,
+      "tokens_seen": 1291059200
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006148290804044295,
+      "loss": 0.0714,
+      "theoretical_loss": 3.591972902499177,
+      "tokens_seen": 1291321344
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000614748836462847,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5919072815822233,
+      "tokens_seen": 1291583488
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006146685925212647,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5918416777109092,
+      "tokens_seen": 1291845632
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006145883485796823,
+      "loss": 0.0697,
+      "theoretical_loss": 3.591776090877349,
+      "tokens_seen": 1292107776
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006145081046380998,
+      "loss": 0.0752,
+      "theoretical_loss": 3.5917105210736624,
+      "tokens_seen": 1292369920
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006144278606965175,
+      "loss": 0.072,
+      "theoretical_loss": 3.591644968291975,
+      "tokens_seen": 1292632064
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000614347616754935,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5915794325244175,
+      "tokens_seen": 1292894208
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006142673728133526,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5915139137631242,
+      "tokens_seen": 1293156352
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006141871288717702,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5914484120002372,
+      "tokens_seen": 1293418496
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006141068849301877,
+      "loss": 0.0718,
+      "theoretical_loss": 3.591382927227902,
+      "tokens_seen": 1293680640
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006140266409886053,
+      "loss": 0.0746,
+      "theoretical_loss": 3.5913174594382693,
+      "tokens_seen": 1293942784
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000613946397047023,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5912520086234974,
+      "tokens_seen": 1294204928
+    },
+    {
+      "epoch": 0.39,
+      "objective/train/advantage_avg": 0.0006294844206422567,
+      "objective/train/docs_used": 473604,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4397389888763428,
+      "objective/train/original_loss": 1.4397391080856323,
+      "objective/train/theoretical_loss": 3.591219289579233,
+      "objective/train/tokens_used": 1314796000,
+      "objective/train/value_avg": -0.0067596435546875,
+      "objective/train/value_loss": 0.00012036753469146788,
+      "objective/train/value_max": -8.219480514526367e-05,
+      "objective/train/value_min": -0.262939453125,
+      "objective/train/value_reward_corr": 0.6750268958365275,
+      "objective/train/value_std": 0.00977325439453125,
+      "objective/train/weight_avg": 1.0006887912750244,
+      "objective/train/weighted_lm_loss": 1.4400802850723267,
+      "objective/train/weights_max": 1.1841590404510498,
+      "objective/train/weights_min": 0.8148074746131897,
+      "theoretical_loss": 3.591219289579233,
+      "tokens_seen": 1294336000
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006138661531054406,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5911865747757457,
+      "tokens_seen": 1294467072
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006137859091638582,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5911211578871827,
+      "tokens_seen": 1294729216
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006137056652222758,
+      "loss": 0.0701,
+      "theoretical_loss": 3.591055757949981,
+      "tokens_seen": 1294991360
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006136254212806933,
+      "loss": 0.074,
+      "theoretical_loss": 3.590990374956316,
+      "tokens_seen": 1295253504
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006135451773391109,
+      "loss": 0.074,
+      "theoretical_loss": 3.5909250088983713,
+      "tokens_seen": 1295515648
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006134649333975285,
+      "loss": 0.0741,
+      "theoretical_loss": 3.5908596597683347,
+      "tokens_seen": 1295777792
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000613384689455946,
+      "loss": 0.0714,
+      "theoretical_loss": 3.590794327558399,
+      "tokens_seen": 1296039936
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006133044455143637,
+      "loss": 0.0697,
+      "theoretical_loss": 3.590729012260762,
+      "tokens_seen": 1296302080
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006132242015727812,
+      "loss": 0.0739,
+      "theoretical_loss": 3.5906637138676265,
+      "tokens_seen": 1296564224
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006131439576311989,
+      "loss": 0.0708,
+      "theoretical_loss": 3.590598432371202,
+      "tokens_seen": 1296826368
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006130637136896165,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5905331677637013,
+      "tokens_seen": 1297088512
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000612983469748034,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5904679200373435,
+      "tokens_seen": 1297350656
+    },
+    {
+      "epoch": 0.39,
+      "objective/train/advantage_avg": 0.0006142263882793486,
+      "objective/train/docs_used": 474727,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.378791332244873,
+      "objective/train/original_loss": 1.378791332244873,
+      "objective/train/theoretical_loss": 3.5904026891843523,
+      "objective/train/tokens_used": 1318072800,
+      "objective/train/value_avg": -0.005767822265625,
+      "objective/train/value_loss": 0.000269250973360613,
+      "objective/train/value_max": -6.0617923736572266e-05,
+      "objective/train/value_min": -0.232666015625,
+      "objective/train/value_reward_corr": 0.4479749164985532,
+      "objective/train/value_std": 0.00815582275390625,
+      "objective/train/weight_avg": 1.0007208585739136,
+      "objective/train/weighted_lm_loss": 1.379752278327942,
+      "objective/train/weights_max": 1.1730239391326904,
+      "objective/train/weights_min": 0.22749631106853485,
+      "theoretical_loss": 3.5904026891843523,
+      "tokens_seen": 1297612800
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006129032258064516,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5904026891843523,
+      "tokens_seen": 1297612800
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006128229818648692,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5903374751969563,
+      "tokens_seen": 1297874944
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006127427379232868,
+      "loss": 0.0687,
+      "theoretical_loss": 3.59027227806739,
+      "tokens_seen": 1298137088
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006126624939817043,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5902070977878937,
+      "tokens_seen": 1298399232
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000612582250040122,
+      "loss": 0.07,
+      "theoretical_loss": 3.5901419343507106,
+      "tokens_seen": 1298661376
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006125020060985395,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5900767877480906,
+      "tokens_seen": 1298923520
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006124217621569572,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5900116579722883,
+      "tokens_seen": 1299185664
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006123415182153748,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5899465450155637,
+      "tokens_seen": 1299447808
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006122612742737923,
+      "loss": 0.0727,
+      "theoretical_loss": 3.589881448870182,
+      "tokens_seen": 1299709952
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.00061218103033221,
+      "loss": 0.0723,
+      "theoretical_loss": 3.589816369528413,
+      "tokens_seen": 1299972096
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006121007863906275,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5897513069825324,
+      "tokens_seen": 1300234240
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006120205424490451,
+      "loss": 0.0721,
+      "theoretical_loss": 3.589686261224819,
+      "tokens_seen": 1300496384
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006119402985074627,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5896212322475605,
+      "tokens_seen": 1300758528
+    },
+    {
+      "epoch": 0.39,
+      "objective/train/advantage_avg": 0.001577448914758861,
+      "objective/train/docs_used": 475730,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3788347244262695,
+      "objective/train/original_loss": 1.3788347244262695,
+      "objective/train/theoretical_loss": 3.5895887240491917,
+      "objective/train/tokens_used": 1321349600,
+      "objective/train/value_avg": -0.00556182861328125,
+      "objective/train/value_loss": 0.0001541156234452501,
+      "objective/train/value_max": -0.00013768672943115234,
+      "objective/train/value_min": -0.29736328125,
+      "objective/train/value_reward_corr": 0.6665854299912869,
+      "objective/train/value_std": 0.009857177734375,
+      "objective/train/weight_avg": 1.0016462802886963,
+      "objective/train/weighted_lm_loss": 1.3818347454071045,
+      "objective/train/weights_max": 1.1952372789382935,
+      "objective/train/weights_min": 0.3779159486293793,
+      "theoretical_loss": 3.5895887240491917,
+      "tokens_seen": 1300889600
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006118600545658803,
+      "loss": 0.0712,
+      "theoretical_loss": 3.589556220043046,
+      "tokens_seen": 1301020672
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006117798106242978,
+      "loss": 0.0713,
+      "theoretical_loss": 3.589491224603571,
+      "tokens_seen": 1301282816
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006116995666827155,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5894262459214366,
+      "tokens_seen": 1301544960
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006116193227411331,
+      "loss": 0.0698,
+      "theoretical_loss": 3.589361283988948,
+      "tokens_seen": 1301807104
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006115390787995506,
+      "loss": 0.0711,
+      "theoretical_loss": 3.589296338798418,
+      "tokens_seen": 1302069248
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006114588348579683,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5892314103421596,
+      "tokens_seen": 1302331392
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006113785909163858,
+      "loss": 0.0712,
+      "theoretical_loss": 3.589166498612496,
+      "tokens_seen": 1302593536
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006112983469748035,
+      "loss": 0.0719,
+      "theoretical_loss": 3.589101603601752,
+      "tokens_seen": 1302855680
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.000611218103033221,
+      "loss": 0.0733,
+      "theoretical_loss": 3.58903672530226,
+      "tokens_seen": 1303117824
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 0.0006111378590916385,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5889718637063552,
+      "tokens_seen": 1303379968
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006110576151500562,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5889070188063794,
+      "tokens_seen": 1303642112
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006109773712084738,
+      "loss": 0.0733,
+      "theoretical_loss": 3.5888421905946783,
+      "tokens_seen": 1303904256
+    },
+    {
+      "epoch": 0.4,
+      "objective/train/advantage_avg": 0.0006969193345867097,
+      "objective/train/docs_used": 476908,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3389830589294434,
+      "objective/train/original_loss": 1.3389828205108643,
+      "objective/train/theoretical_loss": 3.588777379063604,
+      "objective/train/tokens_used": 1324626400,
+      "objective/train/value_avg": -0.00836181640625,
+      "objective/train/value_loss": 0.0003036280977539718,
+      "objective/train/value_max": -3.24249267578125e-05,
+      "objective/train/value_min": -0.76318359375,
+      "objective/train/value_reward_corr": 0.5790845446495612,
+      "objective/train/value_std": 0.01256561279296875,
+      "objective/train/weight_avg": 1.0008283853530884,
+      "objective/train/weighted_lm_loss": 1.3403379917144775,
+      "objective/train/weights_max": 1.2200902700424194,
+      "objective/train/weights_min": 0.36939677596092224,
+      "theoretical_loss": 3.588777379063604,
+      "tokens_seen": 1304166400
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006108971272668914,
+      "loss": 0.0701,
+      "theoretical_loss": 3.588777379063604,
+      "tokens_seen": 1304166400
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.000610816883325309,
+      "loss": 0.074,
+      "theoretical_loss": 3.5887125842055116,
+      "tokens_seen": 1304428544
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006107366393837266,
+      "loss": 0.07,
+      "theoretical_loss": 3.588647806012765,
+      "tokens_seen": 1304690688
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006106563954421441,
+      "loss": 0.0712,
+      "theoretical_loss": 3.588583044477728,
+      "tokens_seen": 1304952832
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006105761515005617,
+      "loss": 0.0742,
+      "theoretical_loss": 3.5885182995927734,
+      "tokens_seen": 1305214976
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006104959075589793,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5884535713502776,
+      "tokens_seen": 1305477120
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006104156636173968,
+      "loss": 0.0702,
+      "theoretical_loss": 3.588388859742622,
+      "tokens_seen": 1305739264
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006103354196758145,
+      "loss": 0.0678,
+      "theoretical_loss": 3.5883241647621933,
+      "tokens_seen": 1306001408
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.000610255175734232,
+      "loss": 0.0701,
+      "theoretical_loss": 3.588259486401383,
+      "tokens_seen": 1306263552
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006101749317926497,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5881948246525877,
+      "tokens_seen": 1306525696
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006100946878510673,
+      "loss": 0.071,
+      "theoretical_loss": 3.588130179508209,
+      "tokens_seen": 1306787840
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006100144439094848,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5880655509606534,
+      "tokens_seen": 1307049984
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006099341999679025,
+      "loss": 0.0734,
+      "theoretical_loss": 3.5880009390023324,
+      "tokens_seen": 1307312128
+    },
+    {
+      "epoch": 0.4,
+      "objective/train/advantage_avg": -7.018262112978846e-05,
+      "objective/train/docs_used": 478148,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4914653301239014,
+      "objective/train/original_loss": 1.4914653301239014,
+      "objective/train/theoretical_loss": 3.587968639241765,
+      "objective/train/tokens_used": 1327903200,
+      "objective/train/value_avg": -0.006618499755859375,
+      "objective/train/value_loss": 0.00026124040596187115,
+      "objective/train/value_max": -6.920099258422852e-05,
+      "objective/train/value_min": -0.21337890625,
+      "objective/train/value_reward_corr": 0.6777787710933085,
+      "objective/train/value_std": 0.01326751708984375,
+      "objective/train/weight_avg": 1.0000460147857666,
+      "objective/train/weighted_lm_loss": 1.4916276931762695,
+      "objective/train/weights_max": 1.1570359468460083,
+      "objective/train/weights_min": 0.36851781606674194,
+      "theoretical_loss": 3.587968639241765,
+      "tokens_seen": 1307443200
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.00060985395602632,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5879363436256626,
+      "tokens_seen": 1307574272
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006097737120847376,
+      "loss": 0.0723,
+      "theoretical_loss": 3.587871764823066,
+      "tokens_seen": 1307836416
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006096934681431552,
+      "loss": 0.0729,
+      "theoretical_loss": 3.5878072025869683,
+      "tokens_seen": 1308098560
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006096132242015728,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5877426569098017,
+      "tokens_seen": 1308360704
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006095329802599903,
+      "loss": 0.069,
+      "theoretical_loss": 3.5876781277840024,
+      "tokens_seen": 1308622848
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.000609452736318408,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5876136152020117,
+      "tokens_seen": 1308884992
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006093724923768256,
+      "loss": 0.0714,
+      "theoretical_loss": 3.587549119156276,
+      "tokens_seen": 1309147136
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006092922484352431,
+      "loss": 0.0748,
+      "theoretical_loss": 3.5874846396392472,
+      "tokens_seen": 1309409280
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006092120044936608,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5874201766433815,
+      "tokens_seen": 1309671424
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006091317605520783,
+      "loss": 0.0728,
+      "theoretical_loss": 3.58735573016114,
+      "tokens_seen": 1309933568
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006090515166104959,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5872913001849884,
+      "tokens_seen": 1310195712
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006089712726689135,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5872268867073993,
+      "tokens_seen": 1310457856
+    },
+    {
+      "debugging/Compilability": 0.9,
+      "debugging/distinct-1-grams": 0.7732995475821178,
+      "debugging/entropy-1-grams": 5.431960387820077,
+      "debugging/length": 501.4,
+      "debugging/num_segments": 10,
+      "debugging/raw_token_scores_avg": 0.009144442155957222,
+      "debugging/raw_token_scores_std": 0.03945710137486458,
+      "debugging/score": 0.007191963134632398,
+      "debugging/score_std": 0.006508080924945895,
+      "epoch": 0.4,
+      "objective/train/advantage_avg": 0.0012943374458700418,
+      "objective/train/docs_used": 479375,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4710427522659302,
+      "objective/train/original_loss": 1.4710428714752197,
+      "objective/train/theoretical_loss": 3.587162489720847,
+      "objective/train/tokens_used": 1331180000,
+      "objective/train/value_avg": -0.01043701171875,
+      "objective/train/value_loss": 0.0005823342362418771,
+      "objective/train/value_max": -0.00010973215103149414,
+      "objective/train/value_min": -0.89501953125,
+      "objective/train/value_reward_corr": 0.7931646191164867,
+      "objective/train/value_std": 0.029541015625,
+      "objective/train/weight_avg": 1.0015616416931152,
+      "objective/train/weighted_lm_loss": 1.4743027687072754,
+      "objective/train/weights_max": 2.00335693359375,
+      "objective/train/weights_min": 0.40475231409072876,
+      "theoretical_loss": 3.587162489720847,
+      "tokens_seen": 1310720000
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006088910287273311,
+      "loss": 0.0695,
+      "theoretical_loss": 3.587162489720847,
+      "tokens_seen": 1310720000
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006088107847857486,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5870981092178136,
+      "tokens_seen": 1310982144
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006087305408441663,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5870337451907854,
+      "tokens_seen": 1311244288
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006086502969025839,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5869693976322523,
+      "tokens_seen": 1311506432
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006085700529610015,
+      "loss": 0.0677,
+      "theoretical_loss": 3.586905066534711,
+      "tokens_seen": 1311768576
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006084898090194191,
+      "loss": 0.0663,
+      "theoretical_loss": 3.5868407518906618,
+      "tokens_seen": 1312030720
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006084095650778366,
+      "loss": 0.0732,
+      "theoretical_loss": 3.586776453692611,
+      "tokens_seen": 1312292864
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006083293211362543,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5867121719330677,
+      "tokens_seen": 1312555008
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006082490771946718,
+      "loss": 0.0713,
+      "theoretical_loss": 3.586647906604549,
+      "tokens_seen": 1312817152
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006081688332530893,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5865836576995744,
+      "tokens_seen": 1313079296
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.000608088589311507,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5865194252106694,
+      "tokens_seen": 1313341440
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006080083453699246,
+      "loss": 0.0727,
+      "theoretical_loss": 3.586455209130364,
+      "tokens_seen": 1313603584
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006079281014283422,
+      "loss": 0.0738,
+      "theoretical_loss": 3.5863910094511935,
+      "tokens_seen": 1313865728
+    },
+    {
+      "epoch": 0.4,
+      "objective/train/advantage_avg": 0.0004384119529277086,
+      "objective/train/docs_used": 480644,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4312394857406616,
+      "objective/train/original_loss": 1.431239128112793,
+      "objective/train/theoretical_loss": 3.5863589157597024,
+      "objective/train/tokens_used": 1334456800,
+      "objective/train/value_avg": -0.00553131103515625,
+      "objective/train/value_loss": 0.00011002233804902062,
+      "objective/train/value_max": -0.00011146068572998047,
+      "objective/train/value_min": -0.28369140625,
+      "objective/train/value_reward_corr": 0.5407403599164424,
+      "objective/train/value_std": 0.00691986083984375,
+      "objective/train/weight_avg": 1.0004894733428955,
+      "objective/train/weighted_lm_loss": 1.4322750568389893,
+      "objective/train/weights_max": 1.328023076057434,
+      "objective/train/weights_min": 0.384535551071167,
+      "theoretical_loss": 3.5863589157597024,
+      "tokens_seen": 1313996800
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006078478574867598,
+      "loss": 0.0713,
+      "theoretical_loss": 3.586326826165698,
+      "tokens_seen": 1314127872
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006077676135451774,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5862626592664215,
+      "tokens_seen": 1314390016
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006076873696035949,
+      "loss": 0.0702,
+      "theoretical_loss": 3.586198508745915,
+      "tokens_seen": 1314652160
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006076071256620125,
+      "loss": 0.0706,
+      "theoretical_loss": 3.586134374596732,
+      "tokens_seen": 1314914304
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006075268817204301,
+      "loss": 0.0721,
+      "theoretical_loss": 3.586070256811432,
+      "tokens_seen": 1315176448
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006074466377788477,
+      "loss": 0.0689,
+      "theoretical_loss": 3.58600615538258,
+      "tokens_seen": 1315438592
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006073663938372653,
+      "loss": 0.0737,
+      "theoretical_loss": 3.5859420703027447,
+      "tokens_seen": 1315700736
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006072861498956828,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5858780015644998,
+      "tokens_seen": 1315962880
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006072059059541006,
+      "loss": 0.075,
+      "theoretical_loss": 3.585813949160425,
+      "tokens_seen": 1316225024
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006071256620125181,
+      "loss": 0.0699,
+      "theoretical_loss": 3.585749913083103,
+      "tokens_seen": 1316487168
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006070454180709356,
+      "loss": 0.0753,
+      "theoretical_loss": 3.5856858933251234,
+      "tokens_seen": 1316749312
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006069651741293533,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5856218898790786,
+      "tokens_seen": 1317011456
+    },
+    {
+      "epoch": 0.4,
+      "objective/train/advantage_avg": 0.0007247466128319502,
+      "objective/train/docs_used": 481887,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5275864601135254,
+      "objective/train/original_loss": 1.5275864601135254,
+      "objective/train/theoretical_loss": 3.585557902737568,
+      "objective/train/tokens_used": 1337733600,
+      "objective/train/value_avg": -0.01215362548828125,
+      "objective/train/value_loss": 0.00025770600768737495,
+      "objective/train/value_max": -7.599592208862305e-05,
+      "objective/train/value_min": -0.9384765625,
+      "objective/train/value_reward_corr": 0.9506689362153705,
+      "objective/train/value_std": 0.051605224609375,
+      "objective/train/weight_avg": 1.0008461475372314,
+      "objective/train/weighted_lm_loss": 1.5283203125,
+      "objective/train/weights_max": 1.57668137550354,
+      "objective/train/weights_min": 0.37430042028427124,
+      "theoretical_loss": 3.585557902737568,
+      "tokens_seen": 1317273600
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006068849301877708,
+      "loss": 0.0741,
+      "theoretical_loss": 3.585557902737568,
+      "tokens_seen": 1317273600
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006068046862461884,
+      "loss": 0.0768,
+      "theoretical_loss": 3.5854939318931933,
+      "tokens_seen": 1317535744
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.000606724442304606,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5854299773385634,
+      "tokens_seen": 1317797888
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006066441983630236,
+      "loss": 0.0725,
+      "theoretical_loss": 3.585366039066291,
+      "tokens_seen": 1318060032
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006065639544214411,
+      "loss": 0.0716,
+      "theoretical_loss": 3.585302117068993,
+      "tokens_seen": 1318322176
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006064837104798588,
+      "loss": 0.0721,
+      "theoretical_loss": 3.585238211339292,
+      "tokens_seen": 1318584320
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006064034665382764,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5851743218698156,
+      "tokens_seen": 1318846464
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006063232225966939,
+      "loss": 0.0715,
+      "theoretical_loss": 3.585110448653195,
+      "tokens_seen": 1319108608
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006062429786551116,
+      "loss": 0.0698,
+      "theoretical_loss": 3.585046591682068,
+      "tokens_seen": 1319370752
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006061627347135291,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5849827509490746,
+      "tokens_seen": 1319632896
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006060824907719468,
+      "loss": 0.0719,
+      "theoretical_loss": 3.584918926446863,
+      "tokens_seen": 1319895040
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006060022468303643,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5848551181680826,
+      "tokens_seen": 1320157184
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006059220028887819,
+      "loss": 0.0677,
+      "theoretical_loss": 3.5847913261053908,
+      "tokens_seen": 1320419328
+    },
+    {
+      "epoch": 0.4,
+      "objective/train/advantage_avg": 0.0010326327756047249,
+      "objective/train/docs_used": 483102,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3745527267456055,
+      "objective/train/original_loss": 1.374552845954895,
+      "objective/train/theoretical_loss": 3.584759436152784,
+      "objective/train/tokens_used": 1341010400,
+      "objective/train/value_avg": -0.008331298828125,
+      "objective/train/value_loss": 0.00027299131033942103,
+      "objective/train/value_max": -0.00014770030975341797,
+      "objective/train/value_min": -0.57470703125,
+      "objective/train/value_reward_corr": 0.6821128634573694,
+      "objective/train/value_std": 0.015960693359375,
+      "objective/train/weight_avg": 1.0011539459228516,
+      "objective/train/weighted_lm_loss": 1.3762563467025757,
+      "objective/train/weights_max": 1.522135853767395,
+      "objective/train/weights_min": 0.3687252402305603,
+      "theoretical_loss": 3.584759436152784,
+      "tokens_seen": 1320550400
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006058417589471995,
+      "loss": 0.072,
+      "theoretical_loss": 3.584727550251447,
+      "tokens_seen": 1320681472
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.000605761515005617,
+      "loss": 0.0734,
+      "theoretical_loss": 3.5846637905989183,
+      "tokens_seen": 1320943616
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006056812710640347,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5846000471404738,
+      "tokens_seen": 1321205760
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006056010271224523,
+      "loss": 0.0762,
+      "theoretical_loss": 3.5845363198687883,
+      "tokens_seen": 1321467904
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006055207831808699,
+      "loss": 0.0727,
+      "theoretical_loss": 3.584472608776542,
+      "tokens_seen": 1321730048
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006054405392392874,
+      "loss": 0.0733,
+      "theoretical_loss": 3.5844089138564197,
+      "tokens_seen": 1321992192
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006053602952977051,
+      "loss": 0.0694,
+      "theoretical_loss": 3.584345235101111,
+      "tokens_seen": 1322254336
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006052800513561226,
+      "loss": 0.071,
+      "theoretical_loss": 3.584281572503309,
+      "tokens_seen": 1322516480
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006051998074145401,
+      "loss": 0.0728,
+      "theoretical_loss": 3.584217926055713,
+      "tokens_seen": 1322778624
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006051195634729578,
+      "loss": 0.0733,
+      "theoretical_loss": 3.584154295751027,
+      "tokens_seen": 1323040768
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006050393195313753,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5840906815819586,
+      "tokens_seen": 1323302912
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006049590755897931,
+      "loss": 0.0725,
+      "theoretical_loss": 3.584027083541222,
+      "tokens_seen": 1323565056
+    },
+    {
+      "epoch": 0.4,
+      "objective/train/advantage_avg": -0.0004341166641097516,
+      "objective/train/docs_used": 484332,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4634686708450317,
+      "objective/train/original_loss": 1.463468313217163,
+      "objective/train/theoretical_loss": 3.583963501621533,
+      "objective/train/tokens_used": 1344287200,
+      "objective/train/value_avg": -0.007572174072265625,
+      "objective/train/value_loss": 0.000452632550150156,
+      "objective/train/value_max": -7.033348083496094e-05,
+      "objective/train/value_min": -0.37451171875,
+      "objective/train/value_reward_corr": 0.6120818484782071,
+      "objective/train/value_std": 0.01380157470703125,
+      "objective/train/weight_avg": 0.9997588396072388,
+      "objective/train/weighted_lm_loss": 1.463590383529663,
+      "objective/train/weights_max": 1.2186625003814697,
+      "objective/train/weights_min": 0.3854755163192749,
+      "theoretical_loss": 3.583963501621533,
+      "tokens_seen": 1323827200
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006048788316482106,
+      "loss": 0.0722,
+      "theoretical_loss": 3.583963501621533,
+      "tokens_seen": 1323827200
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006047985877066282,
+      "loss": 0.0714,
+      "theoretical_loss": 3.583899935815616,
+      "tokens_seen": 1324089344
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006047183437650458,
+      "loss": 0.0715,
+      "theoretical_loss": 3.583836386116197,
+      "tokens_seen": 1324351488
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006046380998234633,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5837728525160086,
+      "tokens_seen": 1324613632
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006045578558818809,
+      "loss": 0.0764,
+      "theoretical_loss": 3.5837093350077875,
+      "tokens_seen": 1324875776
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006044776119402985,
+      "loss": 0.071,
+      "theoretical_loss": 3.5836458335842747,
+      "tokens_seen": 1325137920
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006043973679987161,
+      "loss": 0.0728,
+      "theoretical_loss": 3.5835823482382163,
+      "tokens_seen": 1325400064
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006043171240571336,
+      "loss": 0.0712,
+      "theoretical_loss": 3.583518878962364,
+      "tokens_seen": 1325662208
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006042368801155514,
+      "loss": 0.0709,
+      "theoretical_loss": 3.583455425749472,
+      "tokens_seen": 1325924352
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006041566361739689,
+      "loss": 0.0726,
+      "theoretical_loss": 3.583391988592301,
+      "tokens_seen": 1326186496
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006040763922323864,
+      "loss": 0.0748,
+      "theoretical_loss": 3.5833285674836164,
+      "tokens_seen": 1326448640
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006039961482908041,
+      "loss": 0.0723,
+      "theoretical_loss": 3.583265162416187,
+      "tokens_seen": 1326710784
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006039159043492216,
+      "loss": 0.0723,
+      "theoretical_loss": 3.583201773382788,
+      "tokens_seen": 1326972928
+    },
+    {
+      "epoch": 0.4,
+      "objective/train/advantage_avg": 0.0011210001539438963,
+      "objective/train/docs_used": 485470,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5233768224716187,
+      "objective/train/original_loss": 1.5233769416809082,
+      "objective/train/theoretical_loss": 3.5831700848765924,
+      "objective/train/tokens_used": 1347564000,
+      "objective/train/value_avg": -0.0078277587890625,
+      "objective/train/value_loss": 0.0003868304193019867,
+      "objective/train/value_max": -7.31348991394043e-05,
+      "objective/train/value_min": -0.60107421875,
+      "objective/train/value_reward_corr": 0.5873188182429645,
+      "objective/train/value_std": 0.0148773193359375,
+      "objective/train/weight_avg": 1.0012885332107544,
+      "objective/train/weighted_lm_loss": 1.5247997045516968,
+      "objective/train/weights_max": 1.3386691808700562,
+      "objective/train/weights_min": 0.3721534311771393,
+      "theoretical_loss": 3.5831700848765924,
+      "tokens_seen": 1327104000
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006038356604076392,
+      "loss": 0.0732,
+      "theoretical_loss": 3.583138400376197,
+      "tokens_seen": 1327235072
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006037554164660568,
+      "loss": 0.071,
+      "theoretical_loss": 3.583075043389199,
+      "tokens_seen": 1327497216
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006036751725244744,
+      "loss": 0.073,
+      "theoretical_loss": 3.583011702414581,
+      "tokens_seen": 1327759360
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.000603594928582892,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5829483774451374,
+      "tokens_seen": 1328021504
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006035146846413096,
+      "loss": 0.0711,
+      "theoretical_loss": 3.582885068473665,
+      "tokens_seen": 1328283648
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006034344406997272,
+      "loss": 0.0726,
+      "theoretical_loss": 3.582821775492966,
+      "tokens_seen": 1328545792
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006033541967581448,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5827584984958474,
+      "tokens_seen": 1328807936
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006032739528165624,
+      "loss": 0.0733,
+      "theoretical_loss": 3.582695237475121,
+      "tokens_seen": 1329070080
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006031937088749799,
+      "loss": 0.071,
+      "theoretical_loss": 3.582631992423603,
+      "tokens_seen": 1329332224
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006031134649333976,
+      "loss": 0.0708,
+      "theoretical_loss": 3.582568763334115,
+      "tokens_seen": 1329594368
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006030332209918151,
+      "loss": 0.0684,
+      "theoretical_loss": 3.582505550199481,
+      "tokens_seen": 1329856512
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006029529770502326,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5824423530125324,
+      "tokens_seen": 1330118656
+    },
+    {
+      "epoch": 0.4,
+      "objective/train/advantage_avg": 0.0005511230556294322,
+      "objective/train/docs_used": 486514,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4543170928955078,
+      "objective/train/original_loss": 1.4543169736862183,
+      "objective/train/theoretical_loss": 3.582379171766104,
+      "objective/train/tokens_used": 1350840800,
+      "objective/train/value_avg": -0.00664520263671875,
+      "objective/train/value_loss": 0.0001335861161351204,
+      "objective/train/value_max": -9.918212890625e-05,
+      "objective/train/value_min": -0.250244140625,
+      "objective/train/value_reward_corr": 0.6410735725075329,
+      "objective/train/value_std": 0.01018524169921875,
+      "objective/train/weight_avg": 1.0006134510040283,
+      "objective/train/weighted_lm_loss": 1.4551430940628052,
+      "objective/train/weights_max": 1.1343940496444702,
+      "objective/train/weights_min": 0.3683467209339142,
+      "theoretical_loss": 3.582379171766104,
+      "tokens_seen": 1330380800
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006028727331086503,
+      "loss": 0.0718,
+      "theoretical_loss": 3.582379171766104,
+      "tokens_seen": 1330380800
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006027924891670679,
+      "loss": 0.0722,
+      "theoretical_loss": 3.582316006453034,
+      "tokens_seen": 1330642944
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006027122452254855,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5822528570661683,
+      "tokens_seen": 1330905088
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006026320012839031,
+      "loss": 0.0731,
+      "theoretical_loss": 3.582189723598354,
+      "tokens_seen": 1331167232
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006025517573423207,
+      "loss": 0.0722,
+      "theoretical_loss": 3.582126606042446,
+      "tokens_seen": 1331429376
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006024715134007382,
+      "loss": 0.072,
+      "theoretical_loss": 3.5820635043913005,
+      "tokens_seen": 1331691520
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006023912694591559,
+      "loss": 0.0732,
+      "theoretical_loss": 3.582000418637781,
+      "tokens_seen": 1331953664
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006023110255175734,
+      "loss": 0.0686,
+      "theoretical_loss": 3.581937348774755,
+      "tokens_seen": 1332215808
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.000602230781575991,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5818742947950932,
+      "tokens_seen": 1332477952
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006021505376344086,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5818112566916724,
+      "tokens_seen": 1332740096
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006020702936928261,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5817482344573746,
+      "tokens_seen": 1333002240
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006019900497512439,
+      "loss": 0.0728,
+      "theoretical_loss": 3.5816852280850835,
+      "tokens_seen": 1333264384
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006019098058096614,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5816222375676903,
+      "tokens_seen": 1333526528
+    },
+    {
+      "epoch": 0.4,
+      "objective/train/advantage_avg": 0.0011593915987759829,
+      "objective/train/docs_used": 487665,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3809312582015991,
+      "objective/train/original_loss": 1.3809311389923096,
+      "objective/train/theoretical_loss": 3.58159074825236,
+      "objective/train/tokens_used": 1354117600,
+      "objective/train/value_avg": -0.00823211669921875,
+      "objective/train/value_loss": 0.000309052033117041,
+      "objective/train/value_max": -0.00010889768600463867,
+      "objective/train/value_min": -0.958984375,
+      "objective/train/value_reward_corr": 0.7228285026400678,
+      "objective/train/value_std": 0.0185089111328125,
+      "objective/train/weight_avg": 1.001293659210205,
+      "objective/train/weighted_lm_loss": 1.3820528984069824,
+      "objective/train/weights_max": 1.5919711589813232,
+      "objective/train/weights_min": 0.38147324323654175,
+      "theoretical_loss": 3.58159074825236,
+      "tokens_seen": 1333657600
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.000601829561868079,
+      "loss": 0.0721,
+      "theoretical_loss": 3.58155926289809,
+      "tokens_seen": 1333788672
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006017493179264966,
+      "loss": 0.0687,
+      "theoretical_loss": 3.581496304069181,
+      "tokens_seen": 1334050816
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006016690739849141,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5814333610738673,
+      "tokens_seen": 1334312960
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006015888300433317,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5813704339050583,
+      "tokens_seen": 1334575104
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006015085861017493,
+      "loss": 0.0717,
+      "theoretical_loss": 3.581307522555666,
+      "tokens_seen": 1334837248
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006014283421601669,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5812446270186085,
+      "tokens_seen": 1335099392
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006013480982185844,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5811817472868075,
+      "tokens_seen": 1335361536
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006012678542770022,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5811188833531897,
+      "tokens_seen": 1335623680
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006011876103354197,
+      "loss": 0.068,
+      "theoretical_loss": 3.5810560352106866,
+      "tokens_seen": 1335885824
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006011073663938372,
+      "loss": 0.0686,
+      "theoretical_loss": 3.580993202852234,
+      "tokens_seen": 1336147968
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 0.0006010271224522549,
+      "loss": 0.0675,
+      "theoretical_loss": 3.580930386270772,
+      "tokens_seen": 1336410112
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006009468785106724,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5808675854592464,
+      "tokens_seen": 1336672256
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": -8.30290955491364e-05,
+      "objective/train/docs_used": 488954,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2475985288619995,
+      "objective/train/original_loss": 1.247598648071289,
+      "objective/train/theoretical_loss": 3.5808048004106054,
+      "objective/train/tokens_used": 1357394400,
+      "objective/train/value_avg": -0.00824737548828125,
+      "objective/train/value_loss": 0.00022322498261928558,
+      "objective/train/value_max": -4.1961669921875e-05,
+      "objective/train/value_min": -0.374755859375,
+      "objective/train/value_reward_corr": 0.6647625277203212,
+      "objective/train/value_std": 0.0136566162109375,
+      "objective/train/weight_avg": 1.0000206232070923,
+      "objective/train/weighted_lm_loss": 1.2482742071151733,
+      "objective/train/weights_max": 1.1859358549118042,
+      "objective/train/weights_min": 0.3940271735191345,
+      "theoretical_loss": 3.5808048004106054,
+      "tokens_seen": 1336934400
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006008666345690901,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5808048004106054,
+      "tokens_seen": 1336934400
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006007863906275076,
+      "loss": 0.071,
+      "theoretical_loss": 3.5807420311178033,
+      "tokens_seen": 1337196544
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006007061466859252,
+      "loss": 0.071,
+      "theoretical_loss": 3.580679277573799,
+      "tokens_seen": 1337458688
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006006259027443429,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5806165397715546,
+      "tokens_seen": 1337720832
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006005456588027604,
+      "loss": 0.0717,
+      "theoretical_loss": 3.580553817704039,
+      "tokens_seen": 1337982976
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000600465414861178,
+      "loss": 0.0696,
+      "theoretical_loss": 3.580491111364223,
+      "tokens_seen": 1338245120
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006003851709195956,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5804284207450836,
+      "tokens_seen": 1338507264
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006003049269780132,
+      "loss": 0.0682,
+      "theoretical_loss": 3.580365745839602,
+      "tokens_seen": 1338769408
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006002246830364307,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5803030866407637,
+      "tokens_seen": 1339031552
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006001444390948484,
+      "loss": 0.0701,
+      "theoretical_loss": 3.580240443141559,
+      "tokens_seen": 1339293696
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0006000641951532659,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5801778153349817,
+      "tokens_seen": 1339555840
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005999839512116834,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5801152032140315,
+      "tokens_seen": 1339817984
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005999037072701011,
+      "loss": 0.0698,
+      "theoretical_loss": 3.580052606771712,
+      "tokens_seen": 1340080128
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": 0.0011167412158101797,
+      "objective/train/docs_used": 490192,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4070236682891846,
+      "objective/train/original_loss": 1.4070236682891846,
+      "objective/train/theoretical_loss": 3.580021314427854,
+      "objective/train/tokens_used": 1360671200,
+      "objective/train/value_avg": -0.00506591796875,
+      "objective/train/value_loss": 0.00011695722059812397,
+      "objective/train/value_max": -6.157159805297852e-05,
+      "objective/train/value_min": -0.1492919921875,
+      "objective/train/value_reward_corr": 0.4796502469239035,
+      "objective/train/value_std": 0.006679534912109375,
+      "objective/train/weight_avg": 1.0011686086654663,
+      "objective/train/weighted_lm_loss": 1.408930778503418,
+      "objective/train/weights_max": 1.1452467441558838,
+      "objective/train/weights_min": 0.3836681544780731,
+      "theoretical_loss": 3.580021314427854,
+      "tokens_seen": 1340211200
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005998234633285187,
+      "loss": 0.0697,
+      "theoretical_loss": 3.579990026001031,
+      "tokens_seen": 1340342272
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005997432193869364,
+      "loss": 0.072,
+      "theoretical_loss": 3.579927460895002,
+      "tokens_seen": 1340604416
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005996629754453539,
+      "loss": 0.074,
+      "theoretical_loss": 3.5798649114466405,
+      "tokens_seen": 1340866560
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005995827315037715,
+      "loss": 0.0703,
+      "theoretical_loss": 3.579802377648969,
+      "tokens_seen": 1341128704
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005995024875621891,
+      "loss": 0.0724,
+      "theoretical_loss": 3.579739859495013,
+      "tokens_seen": 1341390848
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005994222436206067,
+      "loss": 0.072,
+      "theoretical_loss": 3.5796773569778026,
+      "tokens_seen": 1341652992
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005993419996790242,
+      "loss": 0.0712,
+      "theoretical_loss": 3.579614870090374,
+      "tokens_seen": 1341915136
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005992617557374418,
+      "loss": 0.075,
+      "theoretical_loss": 3.5795523988257654,
+      "tokens_seen": 1342177280
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005991815117958594,
+      "loss": 0.072,
+      "theoretical_loss": 3.5794899431770215,
+      "tokens_seen": 1342439424
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000599101267854277,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5794275031371896,
+      "tokens_seen": 1342701568
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005990210239126947,
+      "loss": 0.0694,
+      "theoretical_loss": 3.579365078699323,
+      "tokens_seen": 1342963712
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005989407799711122,
+      "loss": 0.072,
+      "theoretical_loss": 3.579302669856479,
+      "tokens_seen": 1343225856
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": 0.000923873158171773,
+      "objective/train/docs_used": 490931,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.553824543952942,
+      "objective/train/original_loss": 1.553824543952942,
+      "objective/train/theoretical_loss": 3.5792402766017197,
+      "objective/train/tokens_used": 1363948000,
+      "objective/train/value_avg": -0.007419586181640625,
+      "objective/train/value_loss": 0.0002518415276426822,
+      "objective/train/value_max": -8.481740951538086e-05,
+      "objective/train/value_min": -0.54833984375,
+      "objective/train/value_reward_corr": 0.6071947482415825,
+      "objective/train/value_std": 0.01244354248046875,
+      "objective/train/weight_avg": 1.0010340213775635,
+      "objective/train/weighted_lm_loss": 1.5558077096939087,
+      "objective/train/weights_max": 1.2336021661758423,
+      "objective/train/weights_min": 0.3683885335922241,
+      "theoretical_loss": 3.5792402766017197,
+      "tokens_seen": 1343488000
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005988605360295298,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5792402766017197,
+      "tokens_seen": 1343488000
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005987802920879474,
+      "loss": 0.0693,
+      "theoretical_loss": 3.57917789892811,
+      "tokens_seen": 1343750144
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005987000481463649,
+      "loss": 0.0692,
+      "theoretical_loss": 3.579115536828721,
+      "tokens_seen": 1344012288
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005986198042047825,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5790531902966274,
+      "tokens_seen": 1344274432
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005985395602632001,
+      "loss": 0.0698,
+      "theoretical_loss": 3.578990859324909,
+      "tokens_seen": 1344536576
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005984593163216177,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5789285439066494,
+      "tokens_seen": 1344798720
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005983790723800354,
+      "loss": 0.0701,
+      "theoretical_loss": 3.578866244034937,
+      "tokens_seen": 1345060864
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000598298828438453,
+      "loss": 0.0733,
+      "theoretical_loss": 3.5788039597028636,
+      "tokens_seen": 1345323008
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005982185844968705,
+      "loss": 0.0677,
+      "theoretical_loss": 3.5787416909035272,
+      "tokens_seen": 1345585152
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005981383405552881,
+      "loss": 0.0707,
+      "theoretical_loss": 3.578679437630029,
+      "tokens_seen": 1345847296
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005980580966137057,
+      "loss": 0.072,
+      "theoretical_loss": 3.5786171998754748,
+      "tokens_seen": 1346109440
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005979778526721232,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5785549776329746,
+      "tokens_seen": 1346371584
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005978976087305409,
+      "loss": 0.0697,
+      "theoretical_loss": 3.578492770895643,
+      "tokens_seen": 1346633728
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": 0.0012662590015679598,
+      "objective/train/docs_used": 492127,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3743919134140015,
+      "objective/train/original_loss": 1.3743919134140015,
+      "objective/train/theoretical_loss": 3.5784616733392656,
+      "objective/train/tokens_used": 1367224800,
+      "objective/train/value_avg": -0.0101776123046875,
+      "objective/train/value_loss": 0.0002954881056211889,
+      "objective/train/value_max": -7.033348083496094e-05,
+      "objective/train/value_min": -0.70166015625,
+      "objective/train/value_reward_corr": 0.7162222528307167,
+      "objective/train/value_std": 0.0204315185546875,
+      "objective/train/weight_avg": 1.0014090538024902,
+      "objective/train/weighted_lm_loss": 1.3758933544158936,
+      "objective/train/weights_max": 1.8491084575653076,
+      "objective/train/weights_min": 0.37020933628082275,
+      "theoretical_loss": 3.5784616733392656,
+      "tokens_seen": 1346764800
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005978173647889584,
+      "loss": 0.0675,
+      "theoretical_loss": 3.5784305796566,
+      "tokens_seen": 1346895872
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000597737120847376,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5783684039089687,
+      "tokens_seen": 1347158016
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005976568769057937,
+      "loss": 0.069,
+      "theoretical_loss": 3.578306243645876,
+      "tokens_seen": 1347420160
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005975766329642112,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5782440988604547,
+      "tokens_seen": 1347682304
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005974963890226288,
+      "loss": 0.07,
+      "theoretical_loss": 3.5781819695458417,
+      "tokens_seen": 1347944448
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005974161450810464,
+      "loss": 0.0718,
+      "theoretical_loss": 3.578119855695178,
+      "tokens_seen": 1348206592
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000597335901139464,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5780577573016084,
+      "tokens_seen": 1348468736
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005972556571978816,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5779956743582835,
+      "tokens_seen": 1348730880
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005971754132562992,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5779336068583563,
+      "tokens_seen": 1348993024
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005970951693147167,
+      "loss": 0.0716,
+      "theoretical_loss": 3.577871554794986,
+      "tokens_seen": 1349255168
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005970149253731343,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5778095181613354,
+      "tokens_seen": 1349517312
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000596934681431552,
+      "loss": 0.0713,
+      "theoretical_loss": 3.577747496950572,
+      "tokens_seen": 1349779456
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": 0.00041352870175614953,
+      "objective/train/docs_used": 493399,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4604887962341309,
+      "objective/train/original_loss": 1.46048903465271,
+      "objective/train/theoretical_loss": 3.577685491155866,
+      "objective/train/tokens_used": 1370501600,
+      "objective/train/value_avg": -0.006885528564453125,
+      "objective/train/value_loss": 0.0001873795554274693,
+      "objective/train/value_max": -9.608268737792969e-05,
+      "objective/train/value_min": -0.3330078125,
+      "objective/train/value_reward_corr": 0.638973367373221,
+      "objective/train/value_std": 0.0112152099609375,
+      "objective/train/weight_avg": 1.0004981756210327,
+      "objective/train/weighted_lm_loss": 1.4614195823669434,
+      "objective/train/weights_max": 1.1102087497711182,
+      "objective/train/weights_min": 0.3691481649875641,
+      "theoretical_loss": 3.577685491155866,
+      "tokens_seen": 1350041600
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005968544374899695,
+      "loss": 0.0717,
+      "theoretical_loss": 3.577685491155866,
+      "tokens_seen": 1350041600
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005967741935483872,
+      "loss": 0.0715,
+      "theoretical_loss": 3.577623500770394,
+      "tokens_seen": 1350303744
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005966939496068047,
+      "loss": 0.0705,
+      "theoretical_loss": 3.577561525787337,
+      "tokens_seen": 1350565888
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005966137056652223,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5774995661998785,
+      "tokens_seen": 1350828032
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005965334617236399,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5774376220012085,
+      "tokens_seen": 1351090176
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005964532177820575,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5773756931845186,
+      "tokens_seen": 1351352320
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000596372973840475,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5773137797430077,
+      "tokens_seen": 1351614464
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005962927298988926,
+      "loss": 0.0718,
+      "theoretical_loss": 3.577251881669877,
+      "tokens_seen": 1351876608
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005962124859573102,
+      "loss": 0.0756,
+      "theoretical_loss": 3.5771899989583336,
+      "tokens_seen": 1352138752
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005961322420157277,
+      "loss": 0.0694,
+      "theoretical_loss": 3.577128131601587,
+      "tokens_seen": 1352400896
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005960519980741455,
+      "loss": 0.0677,
+      "theoretical_loss": 3.5770662795928527,
+      "tokens_seen": 1352663040
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000595971754132563,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5770044429253494,
+      "tokens_seen": 1352925184
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005958915101909807,
+      "loss": 0.0701,
+      "theoretical_loss": 3.576942621592301,
+      "tokens_seen": 1353187328
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": -8.931478078011423e-05,
+      "objective/train/docs_used": 494475,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3972373008728027,
+      "objective/train/original_loss": 1.3972374200820923,
+      "objective/train/theoretical_loss": 3.5769117166740805,
+      "objective/train/tokens_used": 1373778400,
+      "objective/train/value_avg": -0.00982666015625,
+      "objective/train/value_loss": 0.00028727849712595344,
+      "objective/train/value_max": -4.297494888305664e-05,
+      "objective/train/value_min": -0.77197265625,
+      "objective/train/value_reward_corr": 0.7621639892533103,
+      "objective/train/value_std": 0.0201416015625,
+      "objective/train/weight_avg": 1.0000455379486084,
+      "objective/train/weighted_lm_loss": 1.3973900079727173,
+      "objective/train/weights_max": 1.478626012802124,
+      "objective/train/weights_min": 0.3716852366924286,
+      "theoretical_loss": 3.5769117166740805,
+      "tokens_seen": 1353318400
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005958112662493982,
+      "loss": 0.071,
+      "theoretical_loss": 3.576880815586935,
+      "tokens_seen": 1353449472
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005957310223078157,
+      "loss": 0.0714,
+      "theoretical_loss": 3.576819024902483,
+      "tokens_seen": 1353711616
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005956507783662334,
+      "loss": 0.074,
+      "theoretical_loss": 3.576757249532183,
+      "tokens_seen": 1353973760
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005955705344246509,
+      "loss": 0.0697,
+      "theoretical_loss": 3.576695489469274,
+      "tokens_seen": 1354235904
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005954902904830685,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5766337447070016,
+      "tokens_seen": 1354498048
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005954100465414862,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5765720152386153,
+      "tokens_seen": 1354760192
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005953298025999038,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5765103010573682,
+      "tokens_seen": 1355022336
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005952495586583213,
+      "loss": 0.071,
+      "theoretical_loss": 3.576448602156518,
+      "tokens_seen": 1355284480
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005951693147167389,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5763869185293276,
+      "tokens_seen": 1355546624
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005950890707751565,
+      "loss": 0.0729,
+      "theoretical_loss": 3.576325250169062,
+      "tokens_seen": 1355808768
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000595008826833574,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5762635970689933,
+      "tokens_seen": 1356070912
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005949285828919917,
+      "loss": 0.0688,
+      "theoretical_loss": 3.576201959222396,
+      "tokens_seen": 1356333056
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": 0.0009401834686286747,
+      "objective/train/docs_used": 495730,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5935755968093872,
+      "objective/train/original_loss": 1.5935757160186768,
+      "objective/train/theoretical_loss": 3.576140336622548,
+      "objective/train/tokens_used": 1377055200,
+      "objective/train/value_avg": -0.01068878173828125,
+      "objective/train/value_loss": 0.0004906124086119235,
+      "objective/train/value_max": -7.253885269165039e-05,
+      "objective/train/value_min": -0.935546875,
+      "objective/train/value_reward_corr": 0.8641266897975199,
+      "objective/train/value_std": 0.038726806640625,
+      "objective/train/weight_avg": 1.0011733770370483,
+      "objective/train/weighted_lm_loss": 1.594736933708191,
+      "objective/train/weights_max": 1.7491079568862915,
+      "objective/train/weights_min": 0.3695010542869568,
+      "theoretical_loss": 3.576140336622548,
+      "tokens_seen": 1356595200
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005948483389504092,
+      "loss": 0.07,
+      "theoretical_loss": 3.576140336622548,
+      "tokens_seen": 1356595200
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005947680950088269,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5760787292627345,
+      "tokens_seen": 1356857344
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005946878510672444,
+      "loss": 0.0726,
+      "theoretical_loss": 3.576017137136242,
+      "tokens_seen": 1357119488
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000594607607125662,
+      "loss": 0.071,
+      "theoretical_loss": 3.5759555602363635,
+      "tokens_seen": 1357381632
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005945273631840797,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5758939985563942,
+      "tokens_seen": 1357643776
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005944471192424972,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5758324520896347,
+      "tokens_seen": 1357905920
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005943668753009148,
+      "loss": 0.0717,
+      "theoretical_loss": 3.57577092082939,
+      "tokens_seen": 1358168064
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005942866313593324,
+      "loss": 0.071,
+      "theoretical_loss": 3.5757094047689684,
+      "tokens_seen": 1358430208
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.00059420638741775,
+      "loss": 0.0695,
+      "theoretical_loss": 3.575647903901684,
+      "tokens_seen": 1358692352
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005941261434761675,
+      "loss": 0.0715,
+      "theoretical_loss": 3.575586418220853,
+      "tokens_seen": 1358954496
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005940458995345851,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5755249477197983,
+      "tokens_seen": 1359216640
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005939656555930027,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5754634923918447,
+      "tokens_seen": 1359478784
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005938854116514203,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5754020522303227,
+      "tokens_seen": 1359740928
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": 0.0009771112818270922,
+      "objective/train/docs_used": 496861,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3605506420135498,
+      "objective/train/original_loss": 1.3605504035949707,
+      "objective/train/theoretical_loss": 3.5753713378348904,
+      "objective/train/tokens_used": 1380332000,
+      "objective/train/value_avg": -0.0107574462890625,
+      "objective/train/value_loss": 0.0002261370827909559,
+      "objective/train/value_max": -7.659196853637695e-05,
+      "objective/train/value_min": -0.440673828125,
+      "objective/train/value_reward_corr": 0.701051009387965,
+      "objective/train/value_std": 0.0171356201171875,
+      "objective/train/weight_avg": 1.001084804534912,
+      "objective/train/weighted_lm_loss": 1.3605451583862305,
+      "objective/train/weights_max": 1.1667248010635376,
+      "objective/train/weights_min": 0.5305235981941223,
+      "theoretical_loss": 3.5753713378348904,
+      "tokens_seen": 1359872000
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000593805167709838,
+      "loss": 0.0697,
+      "theoretical_loss": 3.575340627228566,
+      "tokens_seen": 1360003072
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005937249237682555,
+      "loss": 0.0711,
+      "theoretical_loss": 3.575279217379914,
+      "tokens_seen": 1360265216
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005936446798266731,
+      "loss": 0.0737,
+      "theoretical_loss": 3.575217822677709,
+      "tokens_seen": 1360527360
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005935644358850907,
+      "loss": 0.0698,
+      "theoretical_loss": 3.575156443115297,
+      "tokens_seen": 1360789504
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005934841919435082,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5750950786860307,
+      "tokens_seen": 1361051648
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005934039480019259,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5750337293832644,
+      "tokens_seen": 1361313792
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005933237040603434,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5749723952003576,
+      "tokens_seen": 1361575936
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000593243460118761,
+      "loss": 0.0741,
+      "theoretical_loss": 3.5749110761306744,
+      "tokens_seen": 1361838080
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005931632161771787,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5748497721675823,
+      "tokens_seen": 1362100224
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005930829722355963,
+      "loss": 0.0722,
+      "theoretical_loss": 3.574788483304453,
+      "tokens_seen": 1362362368
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005930027282940138,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5747272095346636,
+      "tokens_seen": 1362624512
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005929224843524315,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5746659508515943,
+      "tokens_seen": 1362886656
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": 0.0009679661598056555,
+      "objective/train/docs_used": 498147,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4032081365585327,
+      "objective/train/original_loss": 1.4032080173492432,
+      "objective/train/theoretical_loss": 3.5746047072486293,
+      "objective/train/tokens_used": 1383608800,
+      "objective/train/value_avg": -0.006671905517578125,
+      "objective/train/value_loss": 0.00034899654565379024,
+      "objective/train/value_max": -7.486343383789062e-05,
+      "objective/train/value_min": -0.732421875,
+      "objective/train/value_reward_corr": 0.5755574112170757,
+      "objective/train/value_std": 0.0115966796875,
+      "objective/train/weight_avg": 1.0011147260665894,
+      "objective/train/weighted_lm_loss": 1.4046744108200073,
+      "objective/train/weights_max": 1.3812626600265503,
+      "objective/train/weights_min": 0.3844299912452698,
+      "theoretical_loss": 3.5746047072486293,
+      "tokens_seen": 1363148800
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000592842240410849,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5746047072486293,
+      "tokens_seen": 1363148800
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005927619964692665,
+      "loss": 0.0691,
+      "theoretical_loss": 3.574543478719158,
+      "tokens_seen": 1363410944
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005926817525276842,
+      "loss": 0.07,
+      "theoretical_loss": 3.5744822652565724,
+      "tokens_seen": 1363673088
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005926015085861017,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5744210668542706,
+      "tokens_seen": 1363935232
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005925212646445193,
+      "loss": 0.0722,
+      "theoretical_loss": 3.574359883505653,
+      "tokens_seen": 1364197376
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000592441020702937,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5742987152041255,
+      "tokens_seen": 1364459520
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005923607767613546,
+      "loss": 0.0688,
+      "theoretical_loss": 3.574237561943098,
+      "tokens_seen": 1364721664
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005922805328197721,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5741764237159837,
+      "tokens_seen": 1364983808
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005922002888781897,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5741153005162003,
+      "tokens_seen": 1365245952
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005921200449366073,
+      "loss": 0.073,
+      "theoretical_loss": 3.57405419233717,
+      "tokens_seen": 1365508096
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005920398009950249,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5739930991723194,
+      "tokens_seen": 1365770240
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005919595570534425,
+      "loss": 0.0736,
+      "theoretical_loss": 3.5739320210150787,
+      "tokens_seen": 1366032384
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.00059187931311186,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5738709578588814,
+      "tokens_seen": 1366294528
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": -0.0008065802976489067,
+      "objective/train/docs_used": 499217,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3456385135650635,
+      "objective/train/original_loss": 1.345638632774353,
+      "objective/train/theoretical_loss": 3.573840431904124,
+      "objective/train/tokens_used": 1386885600,
+      "objective/train/value_avg": -0.007415771484375,
+      "objective/train/value_loss": 0.00022081122733652592,
+      "objective/train/value_max": -5.650520324707031e-05,
+      "objective/train/value_min": -0.2288818359375,
+      "objective/train/value_reward_corr": 0.6760963179560076,
+      "objective/train/value_std": 0.01129913330078125,
+      "objective/train/weight_avg": 0.9992921948432922,
+      "objective/train/weighted_lm_loss": 1.3444547653198242,
+      "objective/train/weights_max": 1.128661870956421,
+      "objective/train/weights_min": 0.3694700598716736,
+      "theoretical_loss": 3.573840431904124,
+      "tokens_seen": 1366425600
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005917990691702777,
+      "loss": 0.0688,
+      "theoretical_loss": 3.573809909697167,
+      "tokens_seen": 1366556672
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005917188252286952,
+      "loss": 0.0701,
+      "theoretical_loss": 3.573748876523379,
+      "tokens_seen": 1366818816
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005916385812871128,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5736878583309624,
+      "tokens_seen": 1367080960
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005915583373455305,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5736268551133694,
+      "tokens_seen": 1367343104
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.000591478093403948,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5735658668640538,
+      "tokens_seen": 1367605248
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005913978494623656,
+      "loss": 0.0699,
+      "theoretical_loss": 3.573504893576476,
+      "tokens_seen": 1367867392
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005913176055207832,
+      "loss": 0.0699,
+      "theoretical_loss": 3.573443935244099,
+      "tokens_seen": 1368129536
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005912373615792008,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5733829918603903,
+      "tokens_seen": 1368391680
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005911571176376183,
+      "loss": 0.0728,
+      "theoretical_loss": 3.573322063418821,
+      "tokens_seen": 1368653824
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005910768736960359,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5732611499128666,
+      "tokens_seen": 1368915968
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005909966297544535,
+      "loss": 0.07,
+      "theoretical_loss": 3.5732002513360075,
+      "tokens_seen": 1369178112
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 0.0005909163858128712,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5731393676817267,
+      "tokens_seen": 1369440256
+    },
+    {
+      "epoch": 0.41,
+      "objective/train/advantage_avg": 0.00028625421691685915,
+      "objective/train/docs_used": 500422,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4607715606689453,
+      "objective/train/original_loss": 1.4607714414596558,
+      "objective/train/theoretical_loss": 3.573078498943513,
+      "objective/train/tokens_used": 1390162400,
+      "objective/train/value_avg": -0.0100250244140625,
+      "objective/train/value_loss": 0.00033545782207511365,
+      "objective/train/value_max": -5.519390106201172e-05,
+      "objective/train/value_min": -0.65869140625,
+      "objective/train/value_reward_corr": 0.753396177125751,
+      "objective/train/value_std": 0.0200042724609375,
+      "objective/train/weight_avg": 1.000441312789917,
+      "objective/train/weighted_lm_loss": 1.4615808725357056,
+      "objective/train/weights_max": 1.7887964248657227,
+      "objective/train/weights_min": 0.40708115696907043,
+      "theoretical_loss": 3.573078498943513,
+      "tokens_seen": 1369702400
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005908361418712888,
+      "loss": 0.071,
+      "theoretical_loss": 3.573078498943513,
+      "tokens_seen": 1369702400
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005907558979297063,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5730176451148568,
+      "tokens_seen": 1369964544
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000590675653988124,
+      "loss": 0.0699,
+      "theoretical_loss": 3.572956806189256,
+      "tokens_seen": 1370226688
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005905954100465415,
+      "loss": 0.071,
+      "theoretical_loss": 3.5728959821602095,
+      "tokens_seen": 1370488832
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000590515166104959,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5728351730212218,
+      "tokens_seen": 1370750976
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005904349221633767,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5727743787658017,
+      "tokens_seen": 1371013120
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005903546782217942,
+      "loss": 0.0746,
+      "theoretical_loss": 3.572713599387461,
+      "tokens_seen": 1371275264
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005902744342802118,
+      "loss": 0.0694,
+      "theoretical_loss": 3.572652834879716,
+      "tokens_seen": 1371537408
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005901941903386295,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5725920852360877,
+      "tokens_seen": 1371799552
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005901139463970471,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5725313504501006,
+      "tokens_seen": 1372061696
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005900337024554646,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5724706305152827,
+      "tokens_seen": 1372323840
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005899534585138823,
+      "loss": 0.0668,
+      "theoretical_loss": 3.572409925425167,
+      "tokens_seen": 1372585984
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005898732145722998,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5723492351732906,
+      "tokens_seen": 1372848128
+    },
+    {
+      "epoch": 0.42,
+      "objective/train/advantage_avg": 8.659496234031394e-05,
+      "objective/train/docs_used": 501474,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2795002460479736,
+      "objective/train/original_loss": 1.2795000076293945,
+      "objective/train/theoretical_loss": 3.5723188956096736,
+      "objective/train/tokens_used": 1393439200,
+      "objective/train/value_avg": -0.0092010498046875,
+      "objective/train/value_loss": 0.0001278216950595379,
+      "objective/train/value_max": -8.821487426757812e-05,
+      "objective/train/value_min": -0.24072265625,
+      "objective/train/value_reward_corr": 0.8686309210454173,
+      "objective/train/value_std": 0.0202178955078125,
+      "objective/train/weight_avg": 1.0001461505889893,
+      "objective/train/weighted_lm_loss": 1.279416561126709,
+      "objective/train/weights_max": 1.1410603523254395,
+      "objective/train/weights_min": 0.38203245401382446,
+      "theoretical_loss": 3.5723188956096736,
+      "tokens_seen": 1372979200
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005897929706307173,
+      "loss": 0.0701,
+      "theoretical_loss": 3.572288559753194,
+      "tokens_seen": 1373110272
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000589712726689135,
+      "loss": 0.0748,
+      "theoretical_loss": 3.5722278991584218,
+      "tokens_seen": 1373372416
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005896324827475525,
+      "loss": 0.0711,
+      "theoretical_loss": 3.572167253382523,
+      "tokens_seen": 1373634560
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005895522388059702,
+      "loss": 0.0741,
+      "theoretical_loss": 3.5721066224190503,
+      "tokens_seen": 1373896704
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005894719948643878,
+      "loss": 0.0701,
+      "theoretical_loss": 3.572046006261561,
+      "tokens_seen": 1374158848
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005893917509228054,
+      "loss": 0.0678,
+      "theoretical_loss": 3.5719854049036153,
+      "tokens_seen": 1374420992
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000589311506981223,
+      "loss": 0.0694,
+      "theoretical_loss": 3.571924818338779,
+      "tokens_seen": 1374683136
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005892312630396405,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5718642465606214,
+      "tokens_seen": 1374945280
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005891510190980581,
+      "loss": 0.0685,
+      "theoretical_loss": 3.571803689562714,
+      "tokens_seen": 1375207424
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005890707751564757,
+      "loss": 0.0721,
+      "theoretical_loss": 3.571743147338635,
+      "tokens_seen": 1375469568
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005889905312148933,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5716826198819653,
+      "tokens_seen": 1375731712
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005889102872733108,
+      "loss": 0.0734,
+      "theoretical_loss": 3.57162210718629,
+      "tokens_seen": 1375993856
+    },
+    {
+      "epoch": 0.42,
+      "objective/train/advantage_avg": 0.0006503405747935176,
+      "objective/train/docs_used": 502650,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2630902528762817,
+      "objective/train/original_loss": 1.2630900144577026,
+      "objective/train/theoretical_loss": 3.5715616092451983,
+      "objective/train/tokens_used": 1396716000,
+      "objective/train/value_avg": -0.00872039794921875,
+      "objective/train/value_loss": 0.00023866714036557823,
+      "objective/train/value_max": -8.749961853027344e-05,
+      "objective/train/value_min": -0.23876953125,
+      "objective/train/value_reward_corr": 0.7916243639009954,
+      "objective/train/value_std": 0.01435089111328125,
+      "objective/train/weight_avg": 1.000759482383728,
+      "objective/train/weighted_lm_loss": 1.2627712488174438,
+      "objective/train/weights_max": 1.1882543563842773,
+      "objective/train/weights_min": 0.3765632212162018,
+      "theoretical_loss": 3.5715616092451983,
+      "tokens_seen": 1376256000
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005888300433317285,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5715616092451983,
+      "tokens_seen": 1376256000
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000588749799390146,
+      "loss": 0.0675,
+      "theoretical_loss": 3.5715011260522824,
+      "tokens_seen": 1376518144
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005886695554485636,
+      "loss": 0.0701,
+      "theoretical_loss": 3.57144065760114,
+      "tokens_seen": 1376780288
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005885893115069813,
+      "loss": 0.071,
+      "theoretical_loss": 3.5713802038853726,
+      "tokens_seen": 1377042432
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005885090675653988,
+      "loss": 0.0675,
+      "theoretical_loss": 3.5713197648985844,
+      "tokens_seen": 1377304576
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005884288236238165,
+      "loss": 0.0723,
+      "theoretical_loss": 3.571259340634385,
+      "tokens_seen": 1377566720
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000588348579682234,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5711989310863874,
+      "tokens_seen": 1377828864
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005882683357406516,
+      "loss": 0.0664,
+      "theoretical_loss": 3.571138536248209,
+      "tokens_seen": 1378091008
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005881880917990692,
+      "loss": 0.0715,
+      "theoretical_loss": 3.57107815611347,
+      "tokens_seen": 1378353152
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005881078478574867,
+      "loss": 0.072,
+      "theoretical_loss": 3.571017790675796,
+      "tokens_seen": 1378615296
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005880276039159043,
+      "loss": 0.0712,
+      "theoretical_loss": 3.570957439928815,
+      "tokens_seen": 1378877440
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000587947359974322,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5708971038661614,
+      "tokens_seen": 1379139584
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005878671160327396,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5708367824814715,
+      "tokens_seen": 1379401728
+    },
+    {
+      "epoch": 0.42,
+      "objective/train/advantage_avg": 0.0012014803942292929,
+      "objective/train/docs_used": 503826,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4277607202529907,
+      "objective/train/original_loss": 1.4277606010437012,
+      "objective/train/theoretical_loss": 3.5708066272913754,
+      "objective/train/tokens_used": 1399992800,
+      "objective/train/value_avg": -0.00450897216796875,
+      "objective/train/value_loss": 6.611572462134063e-05,
+      "objective/train/value_max": -2.4139881134033203e-05,
+      "objective/train/value_min": -0.134033203125,
+      "objective/train/value_reward_corr": 0.5657798775306075,
+      "objective/train/value_std": 0.006134033203125,
+      "objective/train/weight_avg": 1.0012338161468506,
+      "objective/train/weighted_lm_loss": 1.4301097393035889,
+      "objective/train/weights_max": 1.1097402572631836,
+      "objective/train/weights_min": 0.7171305418014526,
+      "theoretical_loss": 3.5708066272913754,
+      "tokens_seen": 1379532800
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005877868720911571,
+      "loss": 0.0707,
+      "theoretical_loss": 3.570776475768386,
+      "tokens_seen": 1379663872
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005877066281495748,
+      "loss": 0.0714,
+      "theoretical_loss": 3.57071618372055,
+      "tokens_seen": 1379926016
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005876263842079923,
+      "loss": 0.0719,
+      "theoretical_loss": 3.570655906331612,
+      "tokens_seen": 1380188160
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005875461402664098,
+      "loss": 0.0708,
+      "theoretical_loss": 3.570595643595225,
+      "tokens_seen": 1380450304
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005874658963248275,
+      "loss": 0.0708,
+      "theoretical_loss": 3.570535395505045,
+      "tokens_seen": 1380712448
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000587385652383245,
+      "loss": 0.0682,
+      "theoretical_loss": 3.570475162054734,
+      "tokens_seen": 1380974592
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005873054084416626,
+      "loss": 0.0705,
+      "theoretical_loss": 3.570414943237956,
+      "tokens_seen": 1381236736
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005872251645000803,
+      "loss": 0.0706,
+      "theoretical_loss": 3.570354739048379,
+      "tokens_seen": 1381498880
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005871449205584979,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5702945494796765,
+      "tokens_seen": 1381761024
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005870646766169155,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5702343745255236,
+      "tokens_seen": 1382023168
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005869844326753331,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5701742141796022,
+      "tokens_seen": 1382285312
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005869041887337506,
+      "loss": 0.0683,
+      "theoretical_loss": 3.570114068435595,
+      "tokens_seen": 1382547456
+    },
+    {
+      "epoch": 0.42,
+      "objective/train/advantage_avg": 0.0009983739582821727,
+      "objective/train/docs_used": 505055,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3534469604492188,
+      "objective/train/original_loss": 1.3534469604492188,
+      "objective/train/theoretical_loss": 3.570053937287192,
+      "objective/train/tokens_used": 1403269600,
+      "objective/train/value_avg": -0.01267242431640625,
+      "objective/train/value_loss": 0.00016829556261654943,
+      "objective/train/value_max": -9.459257125854492e-05,
+      "objective/train/value_min": -0.450927734375,
+      "objective/train/value_reward_corr": 0.8653068777879614,
+      "objective/train/value_std": 0.0235748291015625,
+      "objective/train/weight_avg": 1.0010778903961182,
+      "objective/train/weighted_lm_loss": 1.354577898979187,
+      "objective/train/weights_max": 1.1801592111587524,
+      "objective/train/weights_min": 0.3766666650772095,
+      "theoretical_loss": 3.570053937287192,
+      "tokens_seen": 1382809600
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005868239447921682,
+      "loss": 0.0718,
+      "theoretical_loss": 3.570053937287192,
+      "tokens_seen": 1382809600
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005867437008505858,
+      "loss": 0.0704,
+      "theoretical_loss": 3.569993820728084,
+      "tokens_seen": 1383071744
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005866634569090033,
+      "loss": 0.0687,
+      "theoretical_loss": 3.569933718751967,
+      "tokens_seen": 1383333888
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000586583212967421,
+      "loss": 0.0705,
+      "theoretical_loss": 3.569873631352542,
+      "tokens_seen": 1383596032
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005865029690258386,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5698135585235122,
+      "tokens_seen": 1383858176
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005864227250842562,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5697535002585856,
+      "tokens_seen": 1384120320
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005863424811426738,
+      "loss": 0.0703,
+      "theoretical_loss": 3.569693456551474,
+      "tokens_seen": 1384382464
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005862622372010913,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5696334273958925,
+      "tokens_seen": 1384644608
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005861819932595089,
+      "loss": 0.0702,
+      "theoretical_loss": 3.569573412785561,
+      "tokens_seen": 1384906752
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005861017493179265,
+      "loss": 0.0718,
+      "theoretical_loss": 3.569513412714203,
+      "tokens_seen": 1385168896
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005860215053763441,
+      "loss": 0.073,
+      "theoretical_loss": 3.569453427175546,
+      "tokens_seen": 1385431040
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005859412614347616,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5693934561633203,
+      "tokens_seen": 1385693184
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005858610174931793,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5693334996712625,
+      "tokens_seen": 1385955328
+    },
+    {
+      "epoch": 0.42,
+      "objective/train/advantage_avg": 0.0001342852774541825,
+      "objective/train/docs_used": 506209,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5595440864562988,
+      "objective/train/original_loss": 1.5595442056655884,
+      "objective/train/theoretical_loss": 3.5693035268683393,
+      "objective/train/tokens_used": 1406546400,
+      "objective/train/value_avg": -0.007114410400390625,
+      "objective/train/value_loss": 0.00042897695675492287,
+      "objective/train/value_max": -7.31348991394043e-05,
+      "objective/train/value_min": -0.84033203125,
+      "objective/train/value_reward_corr": 0.7246249964333472,
+      "objective/train/value_std": 0.0205078125,
+      "objective/train/weight_avg": 1.0003262758255005,
+      "objective/train/weighted_lm_loss": 1.5597079992294312,
+      "objective/train/weights_max": 1.8627805709838867,
+      "objective/train/weights_min": 0.39913442730903625,
+      "theoretical_loss": 3.5693035268683393,
+      "tokens_seen": 1386086400
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005857807735515968,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5692735576931103,
+      "tokens_seen": 1386217472
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005857005296100145,
+      "loss": 0.0675,
+      "theoretical_loss": 3.569213630222607,
+      "tokens_seen": 1386479616
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005856202856684321,
+      "loss": 0.07,
+      "theoretical_loss": 3.5691537172535,
+      "tokens_seen": 1386741760
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005855400417268496,
+      "loss": 0.0676,
+      "theoretical_loss": 3.569093818779539,
+      "tokens_seen": 1387003904
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005854597977852673,
+      "loss": 0.0737,
+      "theoretical_loss": 3.5690339347944784,
+      "tokens_seen": 1387266048
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005853795538436848,
+      "loss": 0.0684,
+      "theoretical_loss": 3.568974065292077,
+      "tokens_seen": 1387528192
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005852993099021024,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5689142102660973,
+      "tokens_seen": 1387790336
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.00058521906596052,
+      "loss": 0.068,
+      "theoretical_loss": 3.568854369710305,
+      "tokens_seen": 1388052480
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005851388220189375,
+      "loss": 0.0671,
+      "theoretical_loss": 3.5687945436184703,
+      "tokens_seen": 1388314624
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005850585780773551,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5687347319843665,
+      "tokens_seen": 1388576768
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005849783341357728,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5686749348017726,
+      "tokens_seen": 1388838912
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005848980901941904,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5686151520644684,
+      "tokens_seen": 1389101056
+    },
+    {
+      "epoch": 0.42,
+      "objective/train/advantage_avg": 0.0010544790420681238,
+      "objective/train/docs_used": 507487,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3682609796524048,
+      "objective/train/original_loss": 1.3682608604431152,
+      "objective/train/theoretical_loss": 3.56855538376624,
+      "objective/train/tokens_used": 1409823200,
+      "objective/train/value_avg": -0.00675201416015625,
+      "objective/train/value_loss": 8.866895223036408e-05,
+      "objective/train/value_max": -4.5418739318847656e-05,
+      "objective/train/value_min": -0.2037353515625,
+      "objective/train/value_reward_corr": 0.83755308145317,
+      "objective/train/value_std": 0.012359619140625,
+      "objective/train/weight_avg": 1.0010948181152344,
+      "objective/train/weighted_lm_loss": 1.3702481985092163,
+      "objective/train/weights_max": 1.1099433898925781,
+      "objective/train/weights_min": 0.3734019696712494,
+      "theoretical_loss": 3.56855538376624,
+      "tokens_seen": 1389363200
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005848178462526079,
+      "loss": 0.0704,
+      "theoretical_loss": 3.56855538376624,
+      "tokens_seen": 1389363200
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005847376023110256,
+      "loss": 0.0684,
+      "theoretical_loss": 3.568495629900877,
+      "tokens_seen": 1389625344
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005846573583694431,
+      "loss": 0.0729,
+      "theoretical_loss": 3.5684358904621725,
+      "tokens_seen": 1389887488
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005845771144278606,
+      "loss": 0.0676,
+      "theoretical_loss": 3.5683761654439223,
+      "tokens_seen": 1390149632
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005844968704862783,
+      "loss": 0.067,
+      "theoretical_loss": 3.5683164548399287,
+      "tokens_seen": 1390411776
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005844166265446958,
+      "loss": 0.072,
+      "theoretical_loss": 3.568256758643995,
+      "tokens_seen": 1390673920
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005843363826031135,
+      "loss": 0.0671,
+      "theoretical_loss": 3.5681970768499305,
+      "tokens_seen": 1390936064
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000584256138661531,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5681374094515466,
+      "tokens_seen": 1391198208
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005841758947199487,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5680777564426602,
+      "tokens_seen": 1391460352
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005840956507783663,
+      "loss": 0.0667,
+      "theoretical_loss": 3.56801811781709,
+      "tokens_seen": 1391722496
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005840154068367838,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5679584935686615,
+      "tokens_seen": 1391984640
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005839351628952014,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5678988836912007,
+      "tokens_seen": 1392246784
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000583854918953619,
+      "loss": 0.0693,
+      "theoretical_loss": 3.567839288178539,
+      "tokens_seen": 1392508928
+    },
+    {
+      "epoch": 0.42,
+      "objective/train/advantage_avg": 0.0001814004499465227,
+      "objective/train/docs_used": 508733,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.275675892829895,
+      "objective/train/original_loss": 1.2756757736206055,
+      "objective/train/theoretical_loss": 3.567809495807081,
+      "objective/train/tokens_used": 1413100000,
+      "objective/train/value_avg": -0.00782012939453125,
+      "objective/train/value_loss": 0.00013020160258747637,
+      "objective/train/value_max": -0.00010722875595092773,
+      "objective/train/value_min": -0.270751953125,
+      "objective/train/value_reward_corr": 0.7759416009441611,
+      "objective/train/value_std": 0.01282501220703125,
+      "objective/train/weight_avg": 1.0002455711364746,
+      "objective/train/weighted_lm_loss": 1.275399088859558,
+      "objective/train/weights_max": 1.1240557432174683,
+      "objective/train/weights_min": 0.7806023359298706,
+      "theoretical_loss": 3.567809495807081,
+      "tokens_seen": 1392640000
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005837746750120366,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5677797070245125,
+      "tokens_seen": 1392771072
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005836944310704541,
+      "loss": 0.0688,
+      "theoretical_loss": 3.567720140222959,
+      "tokens_seen": 1393033216
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005836141871288718,
+      "loss": 0.0713,
+      "theoretical_loss": 3.567660587767722,
+      "tokens_seen": 1393295360
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005835339431872894,
+      "loss": 0.0694,
+      "theoretical_loss": 3.567601049652648,
+      "tokens_seen": 1393557504
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.000583453699245707,
+      "loss": 0.069,
+      "theoretical_loss": 3.567541525871587,
+      "tokens_seen": 1393819648
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005833734553041246,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5674820164183934,
+      "tokens_seen": 1394081792
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005832932113625421,
+      "loss": 0.0718,
+      "theoretical_loss": 3.567422521286925,
+      "tokens_seen": 1394343936
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005832129674209598,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5673630404710432,
+      "tokens_seen": 1394606080
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005831327234793773,
+      "loss": 0.0704,
+      "theoretical_loss": 3.567303573964614,
+      "tokens_seen": 1394868224
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005830524795377949,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5672441217615063,
+      "tokens_seen": 1395130368
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005829722355962125,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5671846838555936,
+      "tokens_seen": 1395392512
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005828919916546301,
+      "loss": 0.0706,
+      "theoretical_loss": 3.567125260240752,
+      "tokens_seen": 1395654656
+    },
+    {
+      "epoch": 0.42,
+      "objective/train/advantage_avg": -0.0010937025072053075,
+      "objective/train/docs_used": 509897,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2824382781982422,
+      "objective/train/original_loss": 1.2824382781982422,
+      "objective/train/theoretical_loss": 3.5670658509108626,
+      "objective/train/tokens_used": 1416376800,
+      "objective/train/value_avg": -0.00798797607421875,
+      "objective/train/value_loss": 0.0003030859516002238,
+      "objective/train/value_max": -9.763240814208984e-05,
+      "objective/train/value_min": -0.5146484375,
+      "objective/train/value_reward_corr": 0.868826981243496,
+      "objective/train/value_std": 0.0196075439453125,
+      "objective/train/weight_avg": 0.9990524053573608,
+      "objective/train/weighted_lm_loss": 1.2816001176834106,
+      "objective/train/weights_max": 1.214937686920166,
+      "objective/train/weights_min": 0.6117776036262512,
+      "theoretical_loss": 3.5670658509108626,
+      "tokens_seen": 1395916800
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005828117477130476,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5670658509108626,
+      "tokens_seen": 1395916800
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005827315037714653,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5670064558598096,
+      "tokens_seen": 1396178944
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005826512598298829,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5669470750814813,
+      "tokens_seen": 1396441088
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005825710158883004,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5668877085697694,
+      "tokens_seen": 1396703232
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005824907719467181,
+      "loss": 0.0731,
+      "theoretical_loss": 3.5668283563185703,
+      "tokens_seen": 1396965376
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005824105280051356,
+      "loss": 0.0695,
+      "theoretical_loss": 3.566769018321782,
+      "tokens_seen": 1397227520
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005823302840635532,
+      "loss": 0.0731,
+      "theoretical_loss": 3.5667096945733086,
+      "tokens_seen": 1397489664
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005822500401219708,
+      "loss": 0.0695,
+      "theoretical_loss": 3.566650385067057,
+      "tokens_seen": 1397751808
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005821697961803883,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5665910897969377,
+      "tokens_seen": 1398013952
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005820895522388059,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5665318087568645,
+      "tokens_seen": 1398276096
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005820093082972236,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5664725419407564,
+      "tokens_seen": 1398538240
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005819290643556412,
+      "loss": 0.0704,
+      "theoretical_loss": 3.566413289342535,
+      "tokens_seen": 1398800384
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005818488204140588,
+      "loss": 0.0703,
+      "theoretical_loss": 3.566354050956126,
+      "tokens_seen": 1399062528
+    },
+    {
+      "epoch": 0.42,
+      "objective/train/advantage_avg": 0.000327935500536114,
+      "objective/train/docs_used": 511124,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4136905670166016,
+      "objective/train/original_loss": 1.4136905670166016,
+      "objective/train/theoretical_loss": 3.5663244370904543,
+      "objective/train/tokens_used": 1419653600,
+      "objective/train/value_avg": -0.0080413818359375,
+      "objective/train/value_loss": 0.00012046914343954995,
+      "objective/train/value_max": -0.00011593103408813477,
+      "objective/train/value_min": -0.3408203125,
+      "objective/train/value_reward_corr": 0.7663032114662961,
+      "objective/train/value_std": 0.013214111328125,
+      "objective/train/weight_avg": 1.0003875494003296,
+      "objective/train/weighted_lm_loss": 1.4141696691513062,
+      "objective/train/weights_max": 1.1405415534973145,
+      "objective/train/weights_min": 0.8112593293190002,
+      "theoretical_loss": 3.5663244370904543,
+      "tokens_seen": 1399193600
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005817685764724764,
+      "loss": 0.0721,
+      "theoretical_loss": 3.566294826775459,
+      "tokens_seen": 1399324672
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005816883325308939,
+      "loss": 0.0682,
+      "theoretical_loss": 3.566235616794466,
+      "tokens_seen": 1399586816
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005816080885893115,
+      "loss": 0.0698,
+      "theoretical_loss": 3.566176421007085,
+      "tokens_seen": 1399848960
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005815278446477291,
+      "loss": 0.0691,
+      "theoretical_loss": 3.566117239407256,
+      "tokens_seen": 1400111104
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005814476007061466,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5660580719889237,
+      "tokens_seen": 1400373248
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005813673567645643,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5659989187460353,
+      "tokens_seen": 1400635392
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005812871128229819,
+      "loss": 0.0741,
+      "theoretical_loss": 3.5659397796725427,
+      "tokens_seen": 1400897536
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005812068688813995,
+      "loss": 0.0672,
+      "theoretical_loss": 3.565880654762402,
+      "tokens_seen": 1401159680
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005811266249398171,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5658215440095717,
+      "tokens_seen": 1401421824
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005810463809982346,
+      "loss": 0.0699,
+      "theoretical_loss": 3.565762447408015,
+      "tokens_seen": 1401683968
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005809661370566522,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5657033649516974,
+      "tokens_seen": 1401946112
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 0.0005808858931150698,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5656442966345905,
+      "tokens_seen": 1402208256
+    },
+    {
+      "epoch": 0.42,
+      "objective/train/advantage_avg": 0.0005382169038057327,
+      "objective/train/docs_used": 512300,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5377522706985474,
+      "objective/train/original_loss": 1.537752389907837,
+      "objective/train/theoretical_loss": 3.565585242450667,
+      "objective/train/tokens_used": 1422930400,
+      "objective/train/value_avg": -0.006023406982421875,
+      "objective/train/value_loss": 0.0002111362264258787,
+      "objective/train/value_max": -6.711483001708984e-05,
+      "objective/train/value_min": -0.61865234375,
+      "objective/train/value_reward_corr": 0.7056385404610528,
+      "objective/train/value_std": 0.013519287109375,
+      "objective/train/weight_avg": 1.0006349086761475,
+      "objective/train/weighted_lm_loss": 1.538962721824646,
+      "objective/train/weights_max": 1.5594542026519775,
+      "objective/train/weights_min": 0.37144145369529724,
+      "theoretical_loss": 3.565585242450667,
+      "tokens_seen": 1402470400
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005808056491734874,
+      "loss": 0.0743,
+      "theoretical_loss": 3.565585242450667,
+      "tokens_seen": 1402470400
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000580725405231905,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5655262023939054,
+      "tokens_seen": 1402732544
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005806451612903226,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5654671764582866,
+      "tokens_seen": 1402994688
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005805649173487401,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5654081646377955,
+      "tokens_seen": 1403256832
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005804846734071579,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5653491669264215,
+      "tokens_seen": 1403518976
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005804044294655754,
+      "loss": 0.071,
+      "theoretical_loss": 3.565290183318156,
+      "tokens_seen": 1403781120
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005803241855239929,
+      "loss": 0.0706,
+      "theoretical_loss": 3.565231213806995,
+      "tokens_seen": 1404043264
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005802439415824106,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5651722583869394,
+      "tokens_seen": 1404305408
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005801636976408281,
+      "loss": 0.0724,
+      "theoretical_loss": 3.565113317051991,
+      "tokens_seen": 1404567552
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005800834536992457,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5650543897961584,
+      "tokens_seen": 1404829696
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005800032097576633,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5649954766134515,
+      "tokens_seen": 1405091840
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005799229658160809,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5649365774978845,
+      "tokens_seen": 1405353984
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005798427218744984,
+      "loss": 0.07,
+      "theoretical_loss": 3.5648776924434755,
+      "tokens_seen": 1405616128
+    },
+    {
+      "epoch": 0.43,
+      "objective/train/advantage_avg": 0.0011828236747533083,
+      "objective/train/docs_used": 513528,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.369804859161377,
+      "objective/train/original_loss": 1.369804859161377,
+      "objective/train/theoretical_loss": 3.5648482551873366,
+      "objective/train/tokens_used": 1426207200,
+      "objective/train/value_avg": -0.00914764404296875,
+      "objective/train/value_loss": 0.00022888151579536498,
+      "objective/train/value_max": -0.00011962652206420898,
+      "objective/train/value_min": -0.6591796875,
+      "objective/train/value_reward_corr": 0.6871334576079523,
+      "objective/train/value_std": 0.01441192626953125,
+      "objective/train/weight_avg": 1.0012850761413574,
+      "objective/train/weighted_lm_loss": 1.371004581451416,
+      "objective/train/weights_max": 1.3404009342193604,
+      "objective/train/weights_min": 0.3783082664012909,
+      "theoretical_loss": 3.5648482551873366,
+      "tokens_seen": 1405747200
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005797624779329161,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5648188214442467,
+      "tokens_seen": 1405878272
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005796822339913337,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5647599644942227,
+      "tokens_seen": 1406140416
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005796019900497512,
+      "loss": 0.0713,
+      "theoretical_loss": 3.564701121587434,
+      "tokens_seen": 1406402560
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005795217461081689,
+      "loss": 0.067,
+      "theoretical_loss": 3.5646422927179113,
+      "tokens_seen": 1406664704
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005794415021665864,
+      "loss": 0.0693,
+      "theoretical_loss": 3.564583477879692,
+      "tokens_seen": 1406926848
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005793612582250041,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5645246770668164,
+      "tokens_seen": 1407188992
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005792810142834216,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5644658902733273,
+      "tokens_seen": 1407451136
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005792007703418391,
+      "loss": 0.0714,
+      "theoretical_loss": 3.564407117493272,
+      "tokens_seen": 1407713280
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005791205264002569,
+      "loss": 0.0695,
+      "theoretical_loss": 3.564348358720702,
+      "tokens_seen": 1407975424
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005790402824586744,
+      "loss": 0.0732,
+      "theoretical_loss": 3.564289613949671,
+      "tokens_seen": 1408237568
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000578960038517092,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5642308831742384,
+      "tokens_seen": 1408499712
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005788797945755096,
+      "loss": 0.0708,
+      "theoretical_loss": 3.564172166388465,
+      "tokens_seen": 1408761856
+    },
+    {
+      "epoch": 0.43,
+      "objective/train/advantage_avg": 0.000899914710316807,
+      "objective/train/docs_used": 514827,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.595373511314392,
+      "objective/train/original_loss": 1.5953733921051025,
+      "objective/train/theoretical_loss": 3.5641134635864153,
+      "objective/train/tokens_used": 1429484000,
+      "objective/train/value_avg": -0.007175445556640625,
+      "objective/train/value_loss": 0.00010781797755043954,
+      "objective/train/value_max": -6.35385513305664e-05,
+      "objective/train/value_min": -0.2237548828125,
+      "objective/train/value_reward_corr": 0.6808584583299038,
+      "objective/train/value_std": 0.00986480712890625,
+      "objective/train/weight_avg": 1.000953197479248,
+      "objective/train/weighted_lm_loss": 1.5966917276382446,
+      "objective/train/weights_max": 1.1473335027694702,
+      "objective/train/weights_min": 0.6740972399711609,
+      "theoretical_loss": 3.5641134635864153,
+      "tokens_seen": 1409024000
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005787995506339272,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5641134635864153,
+      "tokens_seen": 1409024000
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005787193066923447,
+      "loss": 0.0691,
+      "theoretical_loss": 3.56405477476216,
+      "tokens_seen": 1409286144
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005786390627507623,
+      "loss": 0.0711,
+      "theoretical_loss": 3.563996099909771,
+      "tokens_seen": 1409548288
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005785588188091799,
+      "loss": 0.0728,
+      "theoretical_loss": 3.5639374390233245,
+      "tokens_seen": 1409810432
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005784785748675974,
+      "loss": 0.0706,
+      "theoretical_loss": 3.563878792096901,
+      "tokens_seen": 1410072576
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005783983309260151,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5638201591245826,
+      "tokens_seen": 1410334720
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005783180869844327,
+      "loss": 0.0696,
+      "theoretical_loss": 3.563761540100457,
+      "tokens_seen": 1410596864
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005782378430428504,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5637029350186156,
+      "tokens_seen": 1410859008
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005781575991012679,
+      "loss": 0.0715,
+      "theoretical_loss": 3.563644343873152,
+      "tokens_seen": 1411121152
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005780773551596854,
+      "loss": 0.0742,
+      "theoretical_loss": 3.5635857666581643,
+      "tokens_seen": 1411383296
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005779971112181031,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5635272033677534,
+      "tokens_seen": 1411645440
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005779168672765206,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5634686539960247,
+      "tokens_seen": 1411907584
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005778366233349382,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5634101185370874,
+      "tokens_seen": 1412169728
+    },
+    {
+      "epoch": 0.43,
+      "objective/train/advantage_avg": 0.0011907402658835053,
+      "objective/train/docs_used": 516049,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4614386558532715,
+      "objective/train/original_loss": 1.4614386558532715,
+      "objective/train/theoretical_loss": 3.563380856023075,
+      "objective/train/tokens_used": 1432760800,
+      "objective/train/value_avg": -0.0091552734375,
+      "objective/train/value_loss": 0.00017796926840674132,
+      "objective/train/value_max": -5.9664249420166016e-05,
+      "objective/train/value_min": -0.1868896484375,
+      "objective/train/value_reward_corr": 0.6773004524352086,
+      "objective/train/value_std": 0.01293182373046875,
+      "objective/train/weight_avg": 1.0012750625610352,
+      "objective/train/weighted_lm_loss": 1.4639524221420288,
+      "objective/train/weights_max": 1.1568671464920044,
+      "objective/train/weights_min": 0.37346750497817993,
+      "theoretical_loss": 3.563380856023075,
+      "tokens_seen": 1412300800
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005777563793933558,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5633515969850533,
+      "tokens_seen": 1412431872
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005776761354517734,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5632930893340378,
+      "tokens_seen": 1412694016
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000577595891510191,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5632345955781606,
+      "tokens_seen": 1412956160
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005775156475686087,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5631761157115456,
+      "tokens_seen": 1413218304
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005774354036270262,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5631176497283175,
+      "tokens_seen": 1413480448
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005773551596854437,
+      "loss": 0.0722,
+      "theoretical_loss": 3.563059197622608,
+      "tokens_seen": 1413742592
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005772749157438614,
+      "loss": 0.0722,
+      "theoretical_loss": 3.56300075938855,
+      "tokens_seen": 1414004736
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005771946718022789,
+      "loss": 0.072,
+      "theoretical_loss": 3.5629423350202813,
+      "tokens_seen": 1414266880
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005771144278606965,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5628839245119424,
+      "tokens_seen": 1414529024
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005770341839191141,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5628255278576777,
+      "tokens_seen": 1414791168
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005769539399775317,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5627671450516347,
+      "tokens_seen": 1415053312
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005768736960359494,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5627087760879657,
+      "tokens_seen": 1415315456
+    },
+    {
+      "epoch": 0.43,
+      "objective/train/advantage_avg": 0.0009819098049774766,
+      "objective/train/docs_used": 517252,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.468359112739563,
+      "objective/train/original_loss": 1.4683589935302734,
+      "objective/train/theoretical_loss": 3.5626504209608254,
+      "objective/train/tokens_used": 1436037600,
+      "objective/train/value_avg": -0.006824493408203125,
+      "objective/train/value_loss": 0.00015597307356074452,
+      "objective/train/value_max": -7.200241088867188e-05,
+      "objective/train/value_min": -0.5869140625,
+      "objective/train/value_reward_corr": 0.6682466763435388,
+      "objective/train/value_std": 0.0118560791015625,
+      "objective/train/weight_avg": 1.001055121421814,
+      "objective/train/weighted_lm_loss": 1.4700490236282349,
+      "objective/train/weights_max": 1.603075385093689,
+      "objective/train/weights_min": 0.37400925159454346,
+      "theoretical_loss": 3.5626504209608254,
+      "tokens_seen": 1415577600
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005767934520943669,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5626504209608254,
+      "tokens_seen": 1415577600
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005767132081527845,
+      "loss": 0.0677,
+      "theoretical_loss": 3.5625920796643724,
+      "tokens_seen": 1415839744
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005766329642112021,
+      "loss": 0.074,
+      "theoretical_loss": 3.5625337521927687,
+      "tokens_seen": 1416101888
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005765527202696197,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5624754385401802,
+      "tokens_seen": 1416364032
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005764724763280372,
+      "loss": 0.0704,
+      "theoretical_loss": 3.562417138700776,
+      "tokens_seen": 1416626176
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005763922323864549,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5623588526687295,
+      "tokens_seen": 1416888320
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005763119884448724,
+      "loss": 0.0683,
+      "theoretical_loss": 3.562300580438216,
+      "tokens_seen": 1417150464
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005762317445032899,
+      "loss": 0.0731,
+      "theoretical_loss": 3.562242322003416,
+      "tokens_seen": 1417412608
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005761515005617077,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5621840773585127,
+      "tokens_seen": 1417674752
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005760712566201252,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5621258464976924,
+      "tokens_seen": 1417936896
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005759910126785428,
+      "loss": 0.069,
+      "theoretical_loss": 3.5620676294151465,
+      "tokens_seen": 1418199040
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005759107687369604,
+      "loss": 0.0696,
+      "theoretical_loss": 3.562009426105069,
+      "tokens_seen": 1418461184
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000575830524795378,
+      "loss": 0.0714,
+      "theoretical_loss": 3.561951236561656,
+      "tokens_seen": 1418723328
+    },
+    {
+      "epoch": 0.43,
+      "objective/train/advantage_avg": 0.00033617427106946707,
+      "objective/train/docs_used": 518371,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4012682437896729,
+      "objective/train/original_loss": 1.4012682437896729,
+      "objective/train/theoretical_loss": 3.5619221469506375,
+      "objective/train/tokens_used": 1439314400,
+      "objective/train/value_avg": -0.006404876708984375,
+      "objective/train/value_loss": 0.0001416690502082929,
+      "objective/train/value_max": -6.014108657836914e-05,
+      "objective/train/value_min": -0.350341796875,
+      "objective/train/value_reward_corr": 0.7427841776207642,
+      "objective/train/value_std": 0.013153076171875,
+      "objective/train/weight_avg": 1.000402808189392,
+      "objective/train/weighted_lm_loss": 1.4016001224517822,
+      "objective/train/weights_max": 1.1936334371566772,
+      "objective/train/weights_min": 0.38919898867607117,
+      "theoretical_loss": 3.5619221469506375,
+      "tokens_seen": 1418854400
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005757502808537955,
+      "loss": 0.0728,
+      "theoretical_loss": 3.56189306077911,
+      "tokens_seen": 1418985472
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005756700369122131,
+      "loss": 0.072,
+      "theoretical_loss": 3.561834898751635,
+      "tokens_seen": 1419247616
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005755897929706307,
+      "loss": 0.0709,
+      "theoretical_loss": 3.561776750473439,
+      "tokens_seen": 1419509760
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005755095490290483,
+      "loss": 0.073,
+      "theoretical_loss": 3.561718615938733,
+      "tokens_seen": 1419771904
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000575429305087466,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5616604951417328,
+      "tokens_seen": 1420034048
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005753490611458835,
+      "loss": 0.0708,
+      "theoretical_loss": 3.561602388076656,
+      "tokens_seen": 1420296192
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005752688172043012,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5615442947377254,
+      "tokens_seen": 1420558336
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005751885732627187,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5614862151191664,
+      "tokens_seen": 1420820480
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005751083293211362,
+      "loss": 0.0721,
+      "theoretical_loss": 3.561428149215208,
+      "tokens_seen": 1421082624
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005750280853795539,
+      "loss": 0.0694,
+      "theoretical_loss": 3.561370097020083,
+      "tokens_seen": 1421344768
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005749478414379714,
+      "loss": 0.0712,
+      "theoretical_loss": 3.561312058528026,
+      "tokens_seen": 1421606912
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000574867597496389,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5612540337332783,
+      "tokens_seen": 1421869056
+    },
+    {
+      "epoch": 0.43,
+      "objective/train/advantage_avg": 0.0010773309040814638,
+      "objective/train/docs_used": 519513,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5569469928741455,
+      "objective/train/original_loss": 1.556947112083435,
+      "objective/train/theoretical_loss": 3.5611960226300816,
+      "objective/train/tokens_used": 1442591200,
+      "objective/train/value_avg": -0.007076263427734375,
+      "objective/train/value_loss": 0.00012166771193733439,
+      "objective/train/value_max": -4.9114227294921875e-05,
+      "objective/train/value_min": -0.2486572265625,
+      "objective/train/value_reward_corr": 0.7132574980049912,
+      "objective/train/value_std": 0.011260986328125,
+      "objective/train/weight_avg": 1.001137375831604,
+      "objective/train/weighted_lm_loss": 1.5595605373382568,
+      "objective/train/weights_max": 1.176132321357727,
+      "objective/train/weights_min": 0.8112593293190002,
+      "theoretical_loss": 3.5611960226300816,
+      "tokens_seen": 1422131200
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005747873535548066,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5611960226300816,
+      "tokens_seen": 1422131200
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005747071096132242,
+      "loss": 0.0713,
+      "theoretical_loss": 3.561138025212683,
+      "tokens_seen": 1422393344
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005746268656716417,
+      "loss": 0.0699,
+      "theoretical_loss": 3.561080041475332,
+      "tokens_seen": 1422655488
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005745466217300594,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5610220714122827,
+      "tokens_seen": 1422917632
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000574466377788477,
+      "loss": 0.0697,
+      "theoretical_loss": 3.560964115017791,
+      "tokens_seen": 1423179776
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005743861338468946,
+      "loss": 0.0701,
+      "theoretical_loss": 3.560906172286118,
+      "tokens_seen": 1423441920
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005743058899053122,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5608482432115265,
+      "tokens_seen": 1423704064
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005742256459637297,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5607903277882853,
+      "tokens_seen": 1423966208
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005741454020221474,
+      "loss": 0.0714,
+      "theoretical_loss": 3.560732426010664,
+      "tokens_seen": 1424228352
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005740651580805649,
+      "loss": 0.0727,
+      "theoretical_loss": 3.560674537872937,
+      "tokens_seen": 1424490496
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005739849141389825,
+      "loss": 0.0694,
+      "theoretical_loss": 3.560616663369382,
+      "tokens_seen": 1424752640
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005739046701974002,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5605588024942803,
+      "tokens_seen": 1425014784
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005738244262558177,
+      "loss": 0.0708,
+      "theoretical_loss": 3.560500955241916,
+      "tokens_seen": 1425276928
+    },
+    {
+      "epoch": 0.43,
+      "objective/train/advantage_avg": 0.0009654366294853389,
+      "objective/train/docs_used": 520731,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3049242496490479,
+      "objective/train/original_loss": 1.3049242496490479,
+      "objective/train/theoretical_loss": 3.5604720367224756,
+      "objective/train/tokens_used": 1445868000,
+      "objective/train/value_avg": -0.007049560546875,
+      "objective/train/value_loss": 0.00032756561995483935,
+      "objective/train/value_max": -3.534555435180664e-05,
+      "objective/train/value_min": -0.97900390625,
+      "objective/train/value_reward_corr": 0.657158229851628,
+      "objective/train/value_std": 0.01511383056640625,
+      "objective/train/weight_avg": 1.0011066198349,
+      "objective/train/weighted_lm_loss": 1.3057200908660889,
+      "objective/train/weights_max": 1.2917603254318237,
+      "objective/train/weights_min": 0.36948272585868835,
+      "theoretical_loss": 3.5604720367224756,
+      "tokens_seen": 1425408000
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005737441823142353,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5604431216065775,
+      "tokens_seen": 1425539072
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005736639383726529,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5603853015825564,
+      "tokens_seen": 1425801216
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005735836944310705,
+      "loss": 0.0692,
+      "theoretical_loss": 3.560327495164147,
+      "tokens_seen": 1426063360
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000573503450489488,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5602697023456473,
+      "tokens_seen": 1426325504
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005734232065479057,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5602119231213605,
+      "tokens_seen": 1426587648
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005733429626063232,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5601541574855906,
+      "tokens_seen": 1426849792
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005732627186647407,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5600964054326463,
+      "tokens_seen": 1427111936
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005731824747231584,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5600386669568405,
+      "tokens_seen": 1427374080
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000573102230781576,
+      "loss": 0.0753,
+      "theoretical_loss": 3.559980942052488,
+      "tokens_seen": 1427636224
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005730219868399937,
+      "loss": 0.0696,
+      "theoretical_loss": 3.559923230713907,
+      "tokens_seen": 1427898368
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005729417428984112,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5598655329354214,
+      "tokens_seen": 1428160512
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005728614989568288,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5598078487113556,
+      "tokens_seen": 1428422656
+    },
+    {
+      "epoch": 0.43,
+      "objective/train/advantage_avg": 0.0017608533380553126,
+      "objective/train/docs_used": 521864,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.409778118133545,
+      "objective/train/original_loss": 1.409778118133545,
+      "objective/train/theoretical_loss": 3.55975017803604,
+      "objective/train/tokens_used": 1449144800,
+      "objective/train/value_avg": -0.0076446533203125,
+      "objective/train/value_loss": 0.00018956906569655985,
+      "objective/train/value_max": -5.692243576049805e-05,
+      "objective/train/value_min": -0.73388671875,
+      "objective/train/value_reward_corr": 0.678791158290969,
+      "objective/train/value_std": 0.013763427734375,
+      "objective/train/weight_avg": 1.001844882965088,
+      "objective/train/weighted_lm_loss": 1.4125566482543945,
+      "objective/train/weights_max": 1.1568920612335205,
+      "objective/train/weights_min": 0.36942073702812195,
+      "theoretical_loss": 3.55975017803604,
+      "tokens_seen": 1428684800
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005727812550152464,
+      "loss": 0.0731,
+      "theoretical_loss": 3.55975017803604,
+      "tokens_seen": 1428684800
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005727010110736639,
+      "loss": 0.0751,
+      "theoretical_loss": 3.5596925209038055,
+      "tokens_seen": 1428946944
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005726207671320815,
+      "loss": 0.0721,
+      "theoretical_loss": 3.55963487730899,
+      "tokens_seen": 1429209088
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005725405231904991,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5595772472459313,
+      "tokens_seen": 1429471232
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005724602792489167,
+      "loss": 0.0707,
+      "theoretical_loss": 3.559519630708973,
+      "tokens_seen": 1429733376
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005723800353073343,
+      "loss": 0.071,
+      "theoretical_loss": 3.5594620276924607,
+      "tokens_seen": 1429995520
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000572299791365752,
+      "loss": 0.0703,
+      "theoretical_loss": 3.559404438190745,
+      "tokens_seen": 1430257664
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005722195474241695,
+      "loss": 0.0705,
+      "theoretical_loss": 3.559346862198178,
+      "tokens_seen": 1430519808
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000572139303482587,
+      "loss": 0.0718,
+      "theoretical_loss": 3.559289299709116,
+      "tokens_seen": 1430781952
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005720590595410047,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5592317507179194,
+      "tokens_seen": 1431044096
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005719788155994222,
+      "loss": 0.072,
+      "theoretical_loss": 3.5591742152189507,
+      "tokens_seen": 1431306240
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005718985716578399,
+      "loss": 0.0694,
+      "theoretical_loss": 3.559116693206577,
+      "tokens_seen": 1431568384
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005718183277162574,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5590591846751685,
+      "tokens_seen": 1431830528
+    },
+    {
+      "epoch": 0.43,
+      "objective/train/advantage_avg": 0.0009012027876451612,
+      "objective/train/docs_used": 522998,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3030763864517212,
+      "objective/train/original_loss": 1.3030766248703003,
+      "objective/train/theoretical_loss": 3.5590304354630673,
+      "objective/train/tokens_used": 1452421600,
+      "objective/train/value_avg": -0.00653076171875,
+      "objective/train/value_loss": 0.00013989253784529865,
+      "objective/train/value_max": -8.350610733032227e-05,
+      "objective/train/value_min": -0.347900390625,
+      "objective/train/value_reward_corr": 0.6877286111560412,
+      "objective/train/value_std": 0.01165008544921875,
+      "objective/train/weight_avg": 1.0009663105010986,
+      "objective/train/weighted_lm_loss": 1.304845929145813,
+      "objective/train/weights_max": 1.297332525253296,
+      "objective/train/weights_min": 0.37335923314094543,
+      "theoretical_loss": 3.5590304354630673,
+      "tokens_seen": 1431961600
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000571738083774675,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5590016896190977,
+      "tokens_seen": 1432092672
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005716578398330927,
+      "loss": 0.07,
+      "theoretical_loss": 3.5589442080327416,
+      "tokens_seen": 1432354816
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005715775958915102,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5588867399104798,
+      "tokens_seen": 1432616960
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005714973519499278,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5588292852466967,
+      "tokens_seen": 1432879104
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005714171080083454,
+      "loss": 0.0669,
+      "theoretical_loss": 3.558771844035779,
+      "tokens_seen": 1433141248
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000571336864066763,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5587144162721156,
+      "tokens_seen": 1433403392
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005712566201251805,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5586570019501016,
+      "tokens_seen": 1433665536
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005711763761835982,
+      "loss": 0.074,
+      "theoretical_loss": 3.558599601064133,
+      "tokens_seen": 1433927680
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005710961322420157,
+      "loss": 0.072,
+      "theoretical_loss": 3.5585422136086104,
+      "tokens_seen": 1434189824
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005710158883004333,
+      "loss": 0.072,
+      "theoretical_loss": 3.558484839577937,
+      "tokens_seen": 1434451968
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.000570935644358851,
+      "loss": 0.071,
+      "theoretical_loss": 3.55842747896652,
+      "tokens_seen": 1434714112
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005708554004172685,
+      "loss": 0.068,
+      "theoretical_loss": 3.55837013176877,
+      "tokens_seen": 1434976256
+    },
+    {
+      "epoch": 0.43,
+      "objective/train/advantage_avg": 0.0011591583024710417,
+      "objective/train/docs_used": 524207,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4704546928405762,
+      "objective/train/original_loss": 1.470454454421997,
+      "objective/train/theoretical_loss": 3.5583127979791005,
+      "objective/train/tokens_used": 1455698400,
+      "objective/train/value_avg": -0.005870819091796875,
+      "objective/train/value_loss": 0.00017229585500899702,
+      "objective/train/value_max": -8.219480514526367e-05,
+      "objective/train/value_min": -0.263671875,
+      "objective/train/value_reward_corr": 0.5902054650726603,
+      "objective/train/value_std": 0.01013946533203125,
+      "objective/train/weight_avg": 1.0012328624725342,
+      "objective/train/weighted_lm_loss": 1.4722777605056763,
+      "objective/train/weights_max": 1.1787195205688477,
+      "objective/train/weights_min": 0.3683807849884033,
+      "theoretical_loss": 3.5583127979791005,
+      "tokens_seen": 1435238400
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 0.0005707751564756861,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5583127979791005,
+      "tokens_seen": 1435238400
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005706949125341037,
+      "loss": 0.0712,
+      "theoretical_loss": 3.558255477591928,
+      "tokens_seen": 1435500544
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005706146685925213,
+      "loss": 0.0695,
+      "theoretical_loss": 3.558198170601674,
+      "tokens_seen": 1435762688
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005705344246509389,
+      "loss": 0.0695,
+      "theoretical_loss": 3.558140877002761,
+      "tokens_seen": 1436024832
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005704541807093565,
+      "loss": 0.0728,
+      "theoretical_loss": 3.5580835967896167,
+      "tokens_seen": 1436286976
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000570373936767774,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5580263299566712,
+      "tokens_seen": 1436549120
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005702936928261916,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5579690764983587,
+      "tokens_seen": 1436811264
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005702134488846092,
+      "loss": 0.0669,
+      "theoretical_loss": 3.557911836409115,
+      "tokens_seen": 1437073408
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005701332049430268,
+      "loss": 0.0694,
+      "theoretical_loss": 3.557854609683382,
+      "tokens_seen": 1437335552
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005700529610014445,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5577973963156024,
+      "tokens_seen": 1437597696
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000569972717059862,
+      "loss": 0.0693,
+      "theoretical_loss": 3.557740196300224,
+      "tokens_seen": 1437859840
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005698924731182796,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5576830096316963,
+      "tokens_seen": 1438121984
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005698122291766972,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5576258363044735,
+      "tokens_seen": 1438384128
+    },
+    {
+      "epoch": 0.44,
+      "objective/train/advantage_avg": 0.001051856903359294,
+      "objective/train/docs_used": 525418,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4021375179290771,
+      "objective/train/original_loss": 1.4021375179290771,
+      "objective/train/theoretical_loss": 3.5575972546421184,
+      "objective/train/tokens_used": 1458975200,
+      "objective/train/value_avg": -0.00624847412109375,
+      "objective/train/value_loss": 0.00013370224041864276,
+      "objective/train/value_max": -2.8431415557861328e-05,
+      "objective/train/value_min": -0.401123046875,
+      "objective/train/value_reward_corr": 0.6913211097079406,
+      "objective/train/value_std": 0.0120086669921875,
+      "objective/train/weight_avg": 1.001117467880249,
+      "objective/train/weighted_lm_loss": 1.4049772024154663,
+      "objective/train/weights_max": 1.2057886123657227,
+      "objective/train/weights_min": 0.7243527770042419,
+      "theoretical_loss": 3.5575972546421184,
+      "tokens_seen": 1438515200
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005697319852351147,
+      "loss": 0.073,
+      "theoretical_loss": 3.5575686763130117,
+      "tokens_seen": 1438646272
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005696517412935323,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5575115296517725,
+      "tokens_seen": 1438908416
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005695714973519499,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5574543963152188,
+      "tokens_seen": 1439170560
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005694912534103675,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5573972762978174,
+      "tokens_seen": 1439432704
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000569411009468785,
+      "loss": 0.074,
+      "theoretical_loss": 3.5573401695940383,
+      "tokens_seen": 1439694848
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005693307655272028,
+      "loss": 0.0705,
+      "theoretical_loss": 3.557283076198356,
+      "tokens_seen": 1439956992
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005692505215856203,
+      "loss": 0.0725,
+      "theoretical_loss": 3.557225996105246,
+      "tokens_seen": 1440219136
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005691702776440379,
+      "loss": 0.0734,
+      "theoretical_loss": 3.5571689293091895,
+      "tokens_seen": 1440481280
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005690900337024555,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5571118758046696,
+      "tokens_seen": 1440743424
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000569009789760873,
+      "loss": 0.0683,
+      "theoretical_loss": 3.5570548355861726,
+      "tokens_seen": 1441005568
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005689295458192907,
+      "loss": 0.0734,
+      "theoretical_loss": 3.5569978086481884,
+      "tokens_seen": 1441267712
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005688493018777082,
+      "loss": 0.0682,
+      "theoretical_loss": 3.556940794985211,
+      "tokens_seen": 1441529856
+    },
+    {
+      "epoch": 0.44,
+      "objective/train/advantage_avg": 0.0003429764765314758,
+      "objective/train/docs_used": 526671,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.310866355895996,
+      "objective/train/original_loss": 1.310866355895996,
+      "objective/train/theoretical_loss": 3.5568837945917364,
+      "objective/train/tokens_used": 1462252000,
+      "objective/train/value_avg": -0.005947113037109375,
+      "objective/train/value_loss": 0.0002256840671179816,
+      "objective/train/value_max": -4.267692565917969e-05,
+      "objective/train/value_min": -0.268798828125,
+      "objective/train/value_reward_corr": 0.5540173147606325,
+      "objective/train/value_std": 0.00830841064453125,
+      "objective/train/weight_avg": 1.0004374980926514,
+      "objective/train/weighted_lm_loss": 1.3115313053131104,
+      "objective/train/weights_max": 1.1662676334381104,
+      "objective/train/weights_min": 0.3735102415084839,
+      "theoretical_loss": 3.5568837945917364,
+      "tokens_seen": 1441792000
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005687690579361258,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5568837945917364,
+      "tokens_seen": 1441792000
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005686888139945435,
+      "loss": 0.0733,
+      "theoretical_loss": 3.556826807462264,
+      "tokens_seen": 1442054144
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000568608570052961,
+      "loss": 0.071,
+      "theoretical_loss": 3.5567698335912983,
+      "tokens_seen": 1442316288
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005685283261113786,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5567128729733444,
+      "tokens_seen": 1442578432
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005684480821697962,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5566559256029118,
+      "tokens_seen": 1442840576
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005683678382282138,
+      "loss": 0.0687,
+      "theoretical_loss": 3.556598991474515,
+      "tokens_seen": 1443102720
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005682875942866313,
+      "loss": 0.0709,
+      "theoretical_loss": 3.556542070582669,
+      "tokens_seen": 1443364864
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000568207350345049,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5564851629218928,
+      "tokens_seen": 1443627008
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005681271064034665,
+      "loss": 0.0687,
+      "theoretical_loss": 3.55642826848671,
+      "tokens_seen": 1443889152
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000568046862461884,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5563713872716467,
+      "tokens_seen": 1444151296
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005679666185203018,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5563145192712318,
+      "tokens_seen": 1444413440
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005678863745787193,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5562576644799977,
+      "tokens_seen": 1444675584
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000567806130637137,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5562008228924804,
+      "tokens_seen": 1444937728
+    },
+    {
+      "epoch": 0.44,
+      "objective/train/advantage_avg": 0.0019295045640319586,
+      "objective/train/docs_used": 528002,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3192483186721802,
+      "objective/train/original_loss": 1.3192481994628906,
+      "objective/train/theoretical_loss": 3.556172407048409,
+      "objective/train/tokens_used": 1465528800,
+      "objective/train/value_avg": -0.00946807861328125,
+      "objective/train/value_loss": 0.0003503127081785351,
+      "objective/train/value_max": -8.416175842285156e-05,
+      "objective/train/value_min": -0.499267578125,
+      "objective/train/value_reward_corr": 0.6022576954413182,
+      "objective/train/value_std": 0.015655517578125,
+      "objective/train/weight_avg": 1.002081036567688,
+      "objective/train/weighted_lm_loss": 1.3213697671890259,
+      "objective/train/weights_max": 1.389220118522644,
+      "objective/train/weights_min": 0.37109869718551636,
+      "theoretical_loss": 3.556172407048409,
+      "tokens_seen": 1445068800
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005677258866955545,
+      "loss": 0.0699,
+      "theoretical_loss": 3.556143994503219,
+      "tokens_seen": 1445199872
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005676456427539721,
+      "loss": 0.0672,
+      "theoretical_loss": 3.5560871793067554,
+      "tokens_seen": 1445462016
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005675653988123897,
+      "loss": 0.0694,
+      "theoretical_loss": 3.556030377297635,
+      "tokens_seen": 1445724160
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005674851548708073,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5559735884704073,
+      "tokens_seen": 1445986304
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005674049109292248,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5559168128196235,
+      "tokens_seen": 1446248448
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005673246669876424,
+      "loss": 0.0696,
+      "theoretical_loss": 3.555860050339839,
+      "tokens_seen": 1446510592
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.00056724442304606,
+      "loss": 0.0704,
+      "theoretical_loss": 3.555803301025613,
+      "tokens_seen": 1446772736
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005671641791044776,
+      "loss": 0.0695,
+      "theoretical_loss": 3.555746564871506,
+      "tokens_seen": 1447034880
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005670839351628953,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5556898418720837,
+      "tokens_seen": 1447297024
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005670036912213128,
+      "loss": 0.0674,
+      "theoretical_loss": 3.555633132021914,
+      "tokens_seen": 1447559168
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005669234472797304,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5555764353155688,
+      "tokens_seen": 1447821312
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000566843203338148,
+      "loss": 0.0701,
+      "theoretical_loss": 3.555519751747622,
+      "tokens_seen": 1448083456
+    },
+    {
+      "epoch": 0.44,
+      "objective/train/advantage_avg": 0.0014572564978152514,
+      "objective/train/docs_used": 529166,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.432005763053894,
+      "objective/train/original_loss": 1.432005763053894,
+      "objective/train/theoretical_loss": 3.5554630813126513,
+      "objective/train/tokens_used": 1468805600,
+      "objective/train/value_avg": -0.00937652587890625,
+      "objective/train/value_loss": 0.00045499225961975753,
+      "objective/train/value_max": -0.0001366138458251953,
+      "objective/train/value_min": -0.97021484375,
+      "objective/train/value_reward_corr": 0.7644421545995016,
+      "objective/train/value_std": 0.0241241455078125,
+      "objective/train/weight_avg": 1.0016591548919678,
+      "objective/train/weighted_lm_loss": 1.4349960088729858,
+      "objective/train/weights_max": 1.7320685386657715,
+      "objective/train/weights_min": 0.369626522064209,
+      "theoretical_loss": 3.5554630813126513,
+      "tokens_seen": 1448345600
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005667629593965655,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5554630813126513,
+      "tokens_seen": 1448345600
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005666827154549832,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5554064240052385,
+      "tokens_seen": 1448607744
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005666024715134007,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5553497798199674,
+      "tokens_seen": 1448869888
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005665222275718183,
+      "loss": 0.0724,
+      "theoretical_loss": 3.555293148751426,
+      "tokens_seen": 1449132032
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000566441983630236,
+      "loss": 0.0719,
+      "theoretical_loss": 3.555236530794204,
+      "tokens_seen": 1449394176
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005663617396886536,
+      "loss": 0.07,
+      "theoretical_loss": 3.5551799259428964,
+      "tokens_seen": 1449656320
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005662814957470711,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5551233341920994,
+      "tokens_seen": 1449918464
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005662012518054887,
+      "loss": 0.0717,
+      "theoretical_loss": 3.555066755536414,
+      "tokens_seen": 1450180608
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005661210078639063,
+      "loss": 0.0737,
+      "theoretical_loss": 3.555010189970443,
+      "tokens_seen": 1450442752
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005660407639223238,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5549536374887936,
+      "tokens_seen": 1450704896
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005659605199807415,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5548970980860757,
+      "tokens_seen": 1450967040
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000565880276039159,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5548405717569023,
+      "tokens_seen": 1451229184
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005658000320975766,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5547840584958896,
+      "tokens_seen": 1451491328
+    },
+    {
+      "epoch": 0.44,
+      "objective/train/advantage_avg": -0.004492069128900766,
+      "objective/train/docs_used": 530451,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.529040813446045,
+      "objective/train/original_loss": 1.529041051864624,
+      "objective/train/theoretical_loss": 3.5547558067642617,
+      "objective/train/tokens_used": 1472082400,
+      "objective/train/value_avg": -0.007659912109375,
+      "objective/train/value_loss": 0.0013438891619443893,
+      "objective/train/value_max": -7.253885269165039e-05,
+      "objective/train/value_min": -0.2027587890625,
+      "objective/train/value_reward_corr": 0.6128778178033915,
+      "objective/train/value_std": 0.01111602783203125,
+      "objective/train/weight_avg": 0.9961403012275696,
+      "objective/train/weighted_lm_loss": 1.5288517475128174,
+      "objective/train/weights_max": 1.1460858583450317,
+      "objective/train/weights_min": 0.79996258020401,
+      "theoretical_loss": 3.5547558067642617,
+      "tokens_seen": 1451622400
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005657197881559943,
+      "loss": 0.0724,
+      "theoretical_loss": 3.554727558297657,
+      "tokens_seen": 1451753472
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005656395442144118,
+      "loss": 0.0701,
+      "theoretical_loss": 3.554671071156828,
+      "tokens_seen": 1452015616
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005655593002728294,
+      "loss": 0.0708,
+      "theoretical_loss": 3.554614597068027,
+      "tokens_seen": 1452277760
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000565479056331247,
+      "loss": 0.0716,
+      "theoretical_loss": 3.554558136025884,
+      "tokens_seen": 1452539904
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005653988123896646,
+      "loss": 0.0721,
+      "theoretical_loss": 3.554501688025031,
+      "tokens_seen": 1452802048
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005653185684480822,
+      "loss": 0.0703,
+      "theoretical_loss": 3.554445253060103,
+      "tokens_seen": 1453064192
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005652383245064998,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5543888311257397,
+      "tokens_seen": 1453326336
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005651580805649173,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5543324222165813,
+      "tokens_seen": 1453588480
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005650778366233349,
+      "loss": 0.0693,
+      "theoretical_loss": 3.554276026327274,
+      "tokens_seen": 1453850624
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005649975926817526,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5542196434524653,
+      "tokens_seen": 1454112768
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005649173487401701,
+      "loss": 0.0721,
+      "theoretical_loss": 3.554163273586806,
+      "tokens_seen": 1454374912
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005648371047985878,
+      "loss": 0.0698,
+      "theoretical_loss": 3.554106916724951,
+      "tokens_seen": 1454637056
+    },
+    {
+      "epoch": 0.44,
+      "objective/train/advantage_avg": 0.0006120117614045739,
+      "objective/train/docs_used": 531642,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4688389301300049,
+      "objective/train/original_loss": 1.4688386917114258,
+      "objective/train/theoretical_loss": 3.5540505728615583,
+      "objective/train/tokens_used": 1475359200,
+      "objective/train/value_avg": -0.00594329833984375,
+      "objective/train/value_loss": 0.0004463873337954283,
+      "objective/train/value_max": -6.35385513305664e-05,
+      "objective/train/value_min": -0.556640625,
+      "objective/train/value_reward_corr": 0.45180356973675206,
+      "objective/train/value_std": 0.0121307373046875,
+      "objective/train/weight_avg": 1.0007911920547485,
+      "objective/train/weighted_lm_loss": 1.469335675239563,
+      "objective/train/weights_max": 1.7006314992904663,
+      "objective/train/weights_min": 0.22718757390975952,
+      "theoretical_loss": 3.5540505728615583,
+      "tokens_seen": 1454899200
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005647568608570053,
+      "loss": 0.07,
+      "theoretical_loss": 3.5540505728615583,
+      "tokens_seen": 1454899200
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005646766169154229,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5539942419912878,
+      "tokens_seen": 1455161344
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005645963729738405,
+      "loss": 0.0703,
+      "theoretical_loss": 3.553937924108804,
+      "tokens_seen": 1455423488
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005645161290322581,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5538816192087728,
+      "tokens_seen": 1455685632
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005644358850906756,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5538253272858658,
+      "tokens_seen": 1455947776
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005643556411490932,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5537690483347557,
+      "tokens_seen": 1456209920
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005642753972075108,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5537127823501184,
+      "tokens_seen": 1456472064
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005641951532659285,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5536565293266342,
+      "tokens_seen": 1456734208
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005641149093243461,
+      "loss": 0.0754,
+      "theoretical_loss": 3.553600289258986,
+      "tokens_seen": 1456996352
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005640346653827636,
+      "loss": 0.0712,
+      "theoretical_loss": 3.553544062141859,
+      "tokens_seen": 1457258496
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005639544214411813,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5534878479699423,
+      "tokens_seen": 1457520640
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005638741774995988,
+      "loss": 0.0731,
+      "theoretical_loss": 3.5534316467379288,
+      "tokens_seen": 1457782784
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005637939335580163,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5533754584405126,
+      "tokens_seen": 1458044928
+    },
+    {
+      "epoch": 0.44,
+      "objective/train/advantage_avg": 0.00047437482862733305,
+      "objective/train/docs_used": 532886,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3345245122909546,
+      "objective/train/original_loss": 1.3345245122909546,
+      "objective/train/theoretical_loss": 3.553347369140622,
+      "objective/train/tokens_used": 1478636000,
+      "objective/train/value_avg": -0.00730133056640625,
+      "objective/train/value_loss": 0.00017969703185372055,
+      "objective/train/value_max": -5.227327346801758e-05,
+      "objective/train/value_min": -0.3125,
+      "objective/train/value_reward_corr": 0.6709926357986172,
+      "objective/train/value_std": 0.01160430908203125,
+      "objective/train/weight_avg": 1.000558853149414,
+      "objective/train/weighted_lm_loss": 1.335375189781189,
+      "objective/train/weights_max": 1.3342957496643066,
+      "objective/train/weights_min": 0.37208810448646545,
+      "theoretical_loss": 3.553347369140622,
+      "tokens_seen": 1458176000
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000563713689616434,
+      "loss": 0.0698,
+      "theoretical_loss": 3.553319283072393,
+      "tokens_seen": 1458307072
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005636334456748515,
+      "loss": 0.0725,
+      "theoretical_loss": 3.553263120628271,
+      "tokens_seen": 1458569216
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005635532017332691,
+      "loss": 0.0708,
+      "theoretical_loss": 3.553206971102852,
+      "tokens_seen": 1458831360
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005634729577916868,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5531508344908436,
+      "tokens_seen": 1459093504
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005633927138501044,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5530947107869557,
+      "tokens_seen": 1459355648
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005633124699085219,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5530385999859035,
+      "tokens_seen": 1459617792
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005632322259669395,
+      "loss": 0.071,
+      "theoretical_loss": 3.5529825020824033,
+      "tokens_seen": 1459879936
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005631519820253571,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5529264170711756,
+      "tokens_seen": 1460142080
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005630717380837746,
+      "loss": 0.0715,
+      "theoretical_loss": 3.552870344946944,
+      "tokens_seen": 1460404224
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005629914941421923,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5528142857044345,
+      "tokens_seen": 1460666368
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005629112502006098,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5527582393383765,
+      "tokens_seen": 1460928512
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005628310062590275,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5527022058435036,
+      "tokens_seen": 1461190656
+    },
+    {
+      "epoch": 0.44,
+      "objective/train/advantage_avg": 0.0010733278468251228,
+      "objective/train/docs_used": 534190,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3928134441375732,
+      "objective/train/original_loss": 1.3928134441375732,
+      "objective/train/theoretical_loss": 3.552646185214551,
+      "objective/train/tokens_used": 1481912800,
+      "objective/train/value_avg": -0.00820159912109375,
+      "objective/train/value_loss": 0.00018559301679488271,
+      "objective/train/value_max": -8.821487426757812e-05,
+      "objective/train/value_min": -0.344482421875,
+      "objective/train/value_reward_corr": 0.6832127209726154,
+      "objective/train/value_std": 0.01428985595703125,
+      "objective/train/weight_avg": 1.0011622905731201,
+      "objective/train/weighted_lm_loss": 1.393998146057129,
+      "objective/train/weights_max": 1.2060829401016235,
+      "objective/train/weights_min": 0.6082649827003479,
+      "theoretical_loss": 3.552646185214551,
+      "tokens_seen": 1461452800
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000562750762317445,
+      "loss": 0.0703,
+      "theoretical_loss": 3.552646185214551,
+      "tokens_seen": 1461452800
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005626705183758626,
+      "loss": 0.0699,
+      "theoretical_loss": 3.552590177446257,
+      "tokens_seen": 1461714944
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005625902744342803,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5525341825333645,
+      "tokens_seen": 1461977088
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005625100304926978,
+      "loss": 0.0728,
+      "theoretical_loss": 3.552478200470618,
+      "tokens_seen": 1462239232
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005624297865511154,
+      "loss": 0.072,
+      "theoretical_loss": 3.552422231252766,
+      "tokens_seen": 1462501376
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000562349542609533,
+      "loss": 0.0698,
+      "theoretical_loss": 3.552366274874559,
+      "tokens_seen": 1462763520
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005622692986679506,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5523103313307516,
+      "tokens_seen": 1463025664
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005621890547263681,
+      "loss": 0.0729,
+      "theoretical_loss": 3.5522544006161016,
+      "tokens_seen": 1463287808
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005621088107847857,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5521984827253688,
+      "tokens_seen": 1463549952
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005620285668432034,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5521425776533175,
+      "tokens_seen": 1463812096
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005619483229016209,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5520866853947135,
+      "tokens_seen": 1464074240
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005618680789600386,
+      "loss": 0.0688,
+      "theoretical_loss": 3.5520308059443275,
+      "tokens_seen": 1464336384
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005617878350184561,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5519749392969313,
+      "tokens_seen": 1464598528
+    },
+    {
+      "epoch": 0.44,
+      "objective/train/advantage_avg": 0.000920631573535502,
+      "objective/train/docs_used": 534938,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3900043964385986,
+      "objective/train/original_loss": 1.3900043964385986,
+      "objective/train/theoretical_loss": 3.5519470107727216,
+      "objective/train/tokens_used": 1485189600,
+      "objective/train/value_avg": -0.00879669189453125,
+      "objective/train/value_loss": 0.0005025153513997793,
+      "objective/train/value_max": -9.763240814208984e-05,
+      "objective/train/value_min": -0.7587890625,
+      "objective/train/value_reward_corr": 0.6102940318855143,
+      "objective/train/value_std": 0.0169677734375,
+      "objective/train/weight_avg": 1.0011345148086548,
+      "objective/train/weighted_lm_loss": 1.391117811203003,
+      "objective/train/weights_max": 2.0519468784332275,
+      "objective/train/weights_min": 0.3684299886226654,
+      "theoretical_loss": 3.5519470107727216,
+      "tokens_seen": 1464729600
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005617075910768738,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5519190854473006,
+      "tokens_seen": 1464860672
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005616273471352913,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5518632443902156,
+      "tokens_seen": 1465122816
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005615471031937089,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5518074161204565,
+      "tokens_seen": 1465384960
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005614668592521265,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5517516006328096,
+      "tokens_seen": 1465647104
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.000561386615310544,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5516957979220627,
+      "tokens_seen": 1465909248
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005613063713689616,
+      "loss": 0.071,
+      "theoretical_loss": 3.551640007983007,
+      "tokens_seen": 1466171392
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005612261274273793,
+      "loss": 0.0683,
+      "theoretical_loss": 3.551584230810436,
+      "tokens_seen": 1466433536
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005611458834857969,
+      "loss": 0.0709,
+      "theoretical_loss": 3.551528466399148,
+      "tokens_seen": 1466695680
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005610656395442144,
+      "loss": 0.0711,
+      "theoretical_loss": 3.551472714743942,
+      "tokens_seen": 1466957824
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005609853956026321,
+      "loss": 0.069,
+      "theoretical_loss": 3.551416975839623,
+      "tokens_seen": 1467219968
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005609051516610496,
+      "loss": 0.07,
+      "theoretical_loss": 3.551361249680996,
+      "tokens_seen": 1467482112
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005608249077194671,
+      "loss": 0.069,
+      "theoretical_loss": 3.5513055362628707,
+      "tokens_seen": 1467744256
+    },
+    {
+      "epoch": 0.44,
+      "objective/train/advantage_avg": 0.0007431964622810483,
+      "objective/train/docs_used": 535989,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3608063459396362,
+      "objective/train/original_loss": 1.3608062267303467,
+      "objective/train/theoretical_loss": 3.5512498355800597,
+      "objective/train/tokens_used": 1488466400,
+      "objective/train/value_avg": -0.0081329345703125,
+      "objective/train/value_loss": 0.0004426943778526038,
+      "objective/train/value_max": -4.6133995056152344e-05,
+      "objective/train/value_min": -0.53125,
+      "objective/train/value_reward_corr": 0.5491775466311715,
+      "objective/train/value_std": 0.015716552734375,
+      "objective/train/weight_avg": 1.0009335279464722,
+      "objective/train/weighted_lm_loss": 1.3623521327972412,
+      "objective/train/weights_max": 1.7010573148727417,
+      "objective/train/weights_min": 0.36846092343330383,
+      "theoretical_loss": 3.5512498355800597,
+      "tokens_seen": 1468006400
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005607446637778848,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5512498355800597,
+      "tokens_seen": 1468006400
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 0.0005606644198363023,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5511941476273785,
+      "tokens_seen": 1468268544
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005605841758947199,
+      "loss": 0.0687,
+      "theoretical_loss": 3.551138472399646,
+      "tokens_seen": 1468530688
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005605039319531376,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5510828098916836,
+      "tokens_seen": 1468792832
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005604236880115552,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5510271600983154,
+      "tokens_seen": 1469054976
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005603434440699728,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5509715230143692,
+      "tokens_seen": 1469317120
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005602632001283903,
+      "loss": 0.0701,
+      "theoretical_loss": 3.550915898634676,
+      "tokens_seen": 1469579264
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005601829561868079,
+      "loss": 0.0678,
+      "theoretical_loss": 3.550860286954069,
+      "tokens_seen": 1469841408
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005601027122452255,
+      "loss": 0.068,
+      "theoretical_loss": 3.5508046879673856,
+      "tokens_seen": 1470103552
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005600224683036431,
+      "loss": 0.0721,
+      "theoretical_loss": 3.550749101669465,
+      "tokens_seen": 1470365696
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005599422243620606,
+      "loss": 0.0678,
+      "theoretical_loss": 3.5506935280551497,
+      "tokens_seen": 1470627840
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005598619804204783,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5506379671192865,
+      "tokens_seen": 1470889984
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005597817364788959,
+      "loss": 0.0711,
+      "theoretical_loss": 3.550582418856723,
+      "tokens_seen": 1471152128
+    },
+    {
+      "epoch": 0.45,
+      "objective/train/advantage_avg": 0.0005571788060478866,
+      "objective/train/docs_used": 537210,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3439960479736328,
+      "objective/train/original_loss": 1.3439958095550537,
+      "objective/train/theoretical_loss": 3.55055464947632,
+      "objective/train/tokens_used": 1491743200,
+      "objective/train/value_avg": -0.00795745849609375,
+      "objective/train/value_loss": 0.00031730145565234125,
+      "objective/train/value_max": -6.657838821411133e-05,
+      "objective/train/value_min": -0.343505859375,
+      "objective/train/value_reward_corr": 0.6187407854038999,
+      "objective/train/value_std": 0.01346588134765625,
+      "objective/train/weight_avg": 1.0006963014602661,
+      "objective/train/weighted_lm_loss": 1.345666766166687,
+      "objective/train/weights_max": 1.2610359191894531,
+      "objective/train/weights_min": 0.3889971375465393,
+      "theoretical_loss": 3.55055464947632,
+      "tokens_seen": 1471283200
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005597014925373134,
+      "loss": 0.0692,
+      "theoretical_loss": 3.550526883262312,
+      "tokens_seen": 1471414272
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005596212485957311,
+      "loss": 0.0711,
+      "theoretical_loss": 3.550471360330907,
+      "tokens_seen": 1471676416
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005595410046541486,
+      "loss": 0.0691,
+      "theoretical_loss": 3.550415850057367,
+      "tokens_seen": 1471938560
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005594607607125662,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5503603524365523,
+      "tokens_seen": 1472200704
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005593805167709838,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5503048674633266,
+      "tokens_seen": 1472462848
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005593002728294014,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5502493951325564,
+      "tokens_seen": 1472724992
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005592200288878189,
+      "loss": 0.0739,
+      "theoretical_loss": 3.550193935439112,
+      "tokens_seen": 1472987136
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005591397849462365,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5501384883778666,
+      "tokens_seen": 1473249280
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005590595410046541,
+      "loss": 0.0729,
+      "theoretical_loss": 3.5500830539436956,
+      "tokens_seen": 1473511424
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005589792970630718,
+      "loss": 0.0694,
+      "theoretical_loss": 3.550027632131477,
+      "tokens_seen": 1473773568
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005588990531214894,
+      "loss": 0.0709,
+      "theoretical_loss": 3.549972222936094,
+      "tokens_seen": 1474035712
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005588188091799069,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5499168263524297,
+      "tokens_seen": 1474297856
+    },
+    {
+      "epoch": 0.45,
+      "objective/train/advantage_avg": 0.0007825845386832952,
+      "objective/train/docs_used": 538469,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.366462230682373,
+      "objective/train/original_loss": 1.366462230682373,
+      "objective/train/theoretical_loss": 3.549861442375373,
+      "objective/train/tokens_used": 1495020000,
+      "objective/train/value_avg": -0.005584716796875,
+      "objective/train/value_loss": 0.00012209409032948315,
+      "objective/train/value_max": -4.9114227294921875e-05,
+      "objective/train/value_min": -0.50048828125,
+      "objective/train/value_reward_corr": 0.7176460320886918,
+      "objective/train/value_std": 0.01194000244140625,
+      "objective/train/weight_avg": 1.0008388757705688,
+      "objective/train/weighted_lm_loss": 1.3676908016204834,
+      "objective/train/weights_max": 1.2191088199615479,
+      "objective/train/weights_min": 0.372718870639801,
+      "theoretical_loss": 3.549861442375373,
+      "tokens_seen": 1474560000
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005587385652383246,
+      "loss": 0.0717,
+      "theoretical_loss": 3.549861442375373,
+      "tokens_seen": 1474560000
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005586583212967421,
+      "loss": 0.068,
+      "theoretical_loss": 3.5498060709998143,
+      "tokens_seen": 1474822144
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005585780773551596,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5497507122206473,
+      "tokens_seen": 1475084288
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005584978334135773,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5496953660327684,
+      "tokens_seen": 1475346432
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005584175894719948,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5496400324310775,
+      "tokens_seen": 1475608576
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005583373455304124,
+      "loss": 0.0668,
+      "theoretical_loss": 3.5495847114104766,
+      "tokens_seen": 1475870720
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005582571015888301,
+      "loss": 0.0719,
+      "theoretical_loss": 3.549529402965873,
+      "tokens_seen": 1476132864
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005581768576472477,
+      "loss": 0.0699,
+      "theoretical_loss": 3.549474107092173,
+      "tokens_seen": 1476395008
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005580966137056652,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5494188237842894,
+      "tokens_seen": 1476657152
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005580163697640829,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5493635530371366,
+      "tokens_seen": 1476919296
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005579361258225004,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5493082948456314,
+      "tokens_seen": 1477181440
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005578558818809179,
+      "loss": 0.0723,
+      "theoretical_loss": 3.549253049204695,
+      "tokens_seen": 1477443584
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005577756379393356,
+      "loss": 0.0679,
+      "theoretical_loss": 3.549197816109251,
+      "tokens_seen": 1477705728
+    },
+    {
+      "epoch": 0.45,
+      "objective/train/advantage_avg": 0.0006546726217493415,
+      "objective/train/docs_used": 539614,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3523414134979248,
+      "objective/train/original_loss": 1.3523411750793457,
+      "objective/train/theoretical_loss": 3.549170204264502,
+      "objective/train/tokens_used": 1498296800,
+      "objective/train/value_avg": -0.0098419189453125,
+      "objective/train/value_loss": 0.00043507429654709995,
+      "objective/train/value_max": -5.829334259033203e-05,
+      "objective/train/value_min": -0.2685546875,
+      "objective/train/value_reward_corr": 0.6779971392819804,
+      "objective/train/value_std": 0.017822265625,
+      "objective/train/weight_avg": 1.0008503198623657,
+      "objective/train/weighted_lm_loss": 1.3519009351730347,
+      "objective/train/weights_max": 1.2550468444824219,
+      "objective/train/weights_min": 0.39393097162246704,
+      "theoretical_loss": 3.549170204264502,
+      "tokens_seen": 1477836800
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005576953939977531,
+      "loss": 0.0694,
+      "theoretical_loss": 3.549142595554224,
+      "tokens_seen": 1477967872
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005576151500561709,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5490873875345446,
+      "tokens_seen": 1478230016
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005575349061145884,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5490321920451446,
+      "tokens_seen": 1478492160
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.000557454662173006,
+      "loss": 0.0703,
+      "theoretical_loss": 3.54897700908096,
+      "tokens_seen": 1478754304
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005573744182314236,
+      "loss": 0.0725,
+      "theoretical_loss": 3.548921838636927,
+      "tokens_seen": 1479016448
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005572941742898411,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5488666807079885,
+      "tokens_seen": 1479278592
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005572139303482587,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5488115352890874,
+      "tokens_seen": 1479540736
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005571336864066763,
+      "loss": 0.071,
+      "theoretical_loss": 3.5487564023751714,
+      "tokens_seen": 1479802880
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005570534424650939,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5487012819611894,
+      "tokens_seen": 1480065024
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005569731985235114,
+      "loss": 0.0716,
+      "theoretical_loss": 3.548646174042095,
+      "tokens_seen": 1480327168
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005568929545819291,
+      "loss": 0.0729,
+      "theoretical_loss": 3.5485910786128434,
+      "tokens_seen": 1480589312
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005568127106403467,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5485359956683933,
+      "tokens_seen": 1480851456
+    },
+    {
+      "epoch": 0.45,
+      "objective/train/advantage_avg": 0.0007702909060753882,
+      "objective/train/docs_used": 540822,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.462890863418579,
+      "objective/train/original_loss": 1.462890625,
+      "objective/train/theoretical_loss": 3.548480925203706,
+      "objective/train/tokens_used": 1501573600,
+      "objective/train/value_avg": -0.01175689697265625,
+      "objective/train/value_loss": 0.0004899628693237901,
+      "objective/train/value_max": -9.173154830932617e-05,
+      "objective/train/value_min": -0.472900390625,
+      "objective/train/value_reward_corr": 0.6486511891603499,
+      "objective/train/value_std": 0.0182952880859375,
+      "objective/train/weight_avg": 1.0009918212890625,
+      "objective/train/weighted_lm_loss": 1.4647504091262817,
+      "objective/train/weights_max": 1.530020833015442,
+      "objective/train/weights_min": 0.37112417817115784,
+      "theoretical_loss": 3.548480925203706,
+      "tokens_seen": 1481113600
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005567324666987642,
+      "loss": 0.0724,
+      "theoretical_loss": 3.548480925203706,
+      "tokens_seen": 1481113600
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005566522227571819,
+      "loss": 0.0712,
+      "theoretical_loss": 3.548425867213747,
+      "tokens_seen": 1481375744
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005565719788155994,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5483708216934833,
+      "tokens_seen": 1481637888
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005564917348740171,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5483157886378844,
+      "tokens_seen": 1481900032
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005564114909324346,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5482607680419243,
+      "tokens_seen": 1482162176
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005563312469908522,
+      "loss": 0.0706,
+      "theoretical_loss": 3.548205759900579,
+      "tokens_seen": 1482424320
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005562510030492698,
+      "loss": 0.0708,
+      "theoretical_loss": 3.548150764208828,
+      "tokens_seen": 1482686464
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005561707591076873,
+      "loss": 0.07,
+      "theoretical_loss": 3.5480957809616527,
+      "tokens_seen": 1482948608
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.000556090515166105,
+      "loss": 0.0726,
+      "theoretical_loss": 3.548040810154038,
+      "tokens_seen": 1483210752
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005560102712245226,
+      "loss": 0.074,
+      "theoretical_loss": 3.5479858517809717,
+      "tokens_seen": 1483472896
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005559300272829402,
+      "loss": 0.0721,
+      "theoretical_loss": 3.547930905837445,
+      "tokens_seen": 1483735040
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005558497833413577,
+      "loss": 0.0748,
+      "theoretical_loss": 3.547875972318451,
+      "tokens_seen": 1483997184
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005557695393997754,
+      "loss": 0.0691,
+      "theoretical_loss": 3.547821051218987,
+      "tokens_seen": 1484259328
+    },
+    {
+      "epoch": 0.45,
+      "objective/train/advantage_avg": 0.00014683479093946517,
+      "objective/train/docs_used": 542078,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4164698123931885,
+      "objective/train/original_loss": 1.4164698123931885,
+      "objective/train/theoretical_loss": 3.5477935953250155,
+      "objective/train/tokens_used": 1504850400,
+      "objective/train/value_avg": -0.006717681884765625,
+      "objective/train/value_loss": 0.000441345531726256,
+      "objective/train/value_max": -4.363059997558594e-05,
+      "objective/train/value_min": -0.91796875,
+      "objective/train/value_reward_corr": 0.6474719816157412,
+      "objective/train/value_std": 0.01513671875,
+      "objective/train/weight_avg": 1.000320315361023,
+      "objective/train/weighted_lm_loss": 1.4166218042373657,
+      "objective/train/weights_max": 1.835614800453186,
+      "objective/train/weights_min": 0.23298361897468567,
+      "theoretical_loss": 3.5477935953250155,
+      "tokens_seen": 1484390400
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005556892954581929,
+      "loss": 0.071,
+      "theoretical_loss": 3.5477661425340514,
+      "tokens_seen": 1484521472
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005556090515166104,
+      "loss": 0.0718,
+      "theoretical_loss": 3.547711246258647,
+      "tokens_seen": 1484783616
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005555288075750281,
+      "loss": 0.0732,
+      "theoretical_loss": 3.547656362387779,
+      "tokens_seen": 1485045760
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005554485636334456,
+      "loss": 0.0733,
+      "theoretical_loss": 3.5476014909164553,
+      "tokens_seen": 1485307904
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005553683196918632,
+      "loss": 0.0696,
+      "theoretical_loss": 3.547546631839687,
+      "tokens_seen": 1485570048
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005552880757502809,
+      "loss": 0.0704,
+      "theoretical_loss": 3.547491785152488,
+      "tokens_seen": 1485832192
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005552078318086985,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5474369508498755,
+      "tokens_seen": 1486094336
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005551275878671161,
+      "loss": 0.0716,
+      "theoretical_loss": 3.547382128926868,
+      "tokens_seen": 1486356480
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005550473439255337,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5473273193784896,
+      "tokens_seen": 1486618624
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005549670999839512,
+      "loss": 0.0697,
+      "theoretical_loss": 3.547272522199764,
+      "tokens_seen": 1486880768
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005548868560423688,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5472177373857208,
+      "tokens_seen": 1487142912
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005548066121007864,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5471629649313905,
+      "tokens_seen": 1487405056
+    },
+    {
+      "epoch": 0.45,
+      "objective/train/advantage_avg": -0.0009506873902864754,
+      "objective/train/docs_used": 543335,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3594411611557007,
+      "objective/train/original_loss": 1.3594409227371216,
+      "objective/train/theoretical_loss": 3.547108204831807,
+      "objective/train/tokens_used": 1508127200,
+      "objective/train/value_avg": -0.0094451904296875,
+      "objective/train/value_loss": 0.000862235261593014,
+      "objective/train/value_max": -8.612871170043945e-05,
+      "objective/train/value_min": -0.98388671875,
+      "objective/train/value_reward_corr": 0.6495069759170901,
+      "objective/train/value_std": 0.0212860107421875,
+      "objective/train/weight_avg": 0.9994162917137146,
+      "objective/train/weighted_lm_loss": 1.3578301668167114,
+      "objective/train/weights_max": 1.7093836069107056,
+      "objective/train/weights_min": 0.23241551220417023,
+      "theoretical_loss": 3.547108204831807,
+      "tokens_seen": 1487667200
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005547263681592039,
+      "loss": 0.0688,
+      "theoretical_loss": 3.547108204831807,
+      "tokens_seen": 1487667200
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005546461242176217,
+      "loss": 0.0726,
+      "theoretical_loss": 3.547053457082008,
+      "tokens_seen": 1487929344
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005545658802760392,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5469987216770322,
+      "tokens_seen": 1488191488
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005544856363344568,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5469439986119227,
+      "tokens_seen": 1488453632
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005544053923928744,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5468892878817253,
+      "tokens_seen": 1488715776
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005543251484512919,
+      "loss": 0.0714,
+      "theoretical_loss": 3.546834589481488,
+      "tokens_seen": 1488977920
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005542449045097095,
+      "loss": 0.07,
+      "theoretical_loss": 3.5467799034062617,
+      "tokens_seen": 1489240064
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005541646605681271,
+      "loss": 0.0688,
+      "theoretical_loss": 3.546725229651101,
+      "tokens_seen": 1489502208
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005540844166265447,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5466705682110633,
+      "tokens_seen": 1489764352
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005540041726849623,
+      "loss": 0.0716,
+      "theoretical_loss": 3.546615919081207,
+      "tokens_seen": 1490026496
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.00055392392874338,
+      "loss": 0.0679,
+      "theoretical_loss": 3.546561282256596,
+      "tokens_seen": 1490288640
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005538436848017975,
+      "loss": 0.0678,
+      "theoretical_loss": 3.546506657732295,
+      "tokens_seen": 1490550784
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005537634408602151,
+      "loss": 0.0713,
+      "theoretical_loss": 3.546452045503372,
+      "tokens_seen": 1490812928
+    },
+    {
+      "epoch": 0.45,
+      "objective/train/advantage_avg": 0.001109226606786251,
+      "objective/train/docs_used": 544526,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4318077564239502,
+      "objective/train/original_loss": 1.4318077564239502,
+      "objective/train/theoretical_loss": 3.546424743998138,
+      "objective/train/tokens_used": 1511404000,
+      "objective/train/value_avg": -0.006862640380859375,
+      "objective/train/value_loss": 0.0001357761793769896,
+      "objective/train/value_max": -7.140636444091797e-05,
+      "objective/train/value_min": -0.798828125,
+      "objective/train/value_reward_corr": 0.6639085992244163,
+      "objective/train/value_std": 0.01186370849609375,
+      "objective/train/weight_avg": 1.0011723041534424,
+      "objective/train/weighted_lm_loss": 1.4328300952911377,
+      "objective/train/weights_max": 1.399935007095337,
+      "objective/train/weights_min": 0.37172776460647583,
+      "theoretical_loss": 3.546424743998138,
+      "tokens_seen": 1490944000
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005536831969186327,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5463974455648994,
+      "tokens_seen": 1491075072
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005536029529770502,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5463428579119505,
+      "tokens_seen": 1491337216
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005535227090354679,
+      "loss": 0.0693,
+      "theoretical_loss": 3.546288282539602,
+      "tokens_seen": 1491599360
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005534424650938854,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5462337194429336,
+      "tokens_seen": 1491861504
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.000553362221152303,
+      "loss": 0.0723,
+      "theoretical_loss": 3.546179168617028,
+      "tokens_seen": 1492123648
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005532819772107206,
+      "loss": 0.069,
+      "theoretical_loss": 3.54612463005697,
+      "tokens_seen": 1492385792
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005532017332691381,
+      "loss": 0.073,
+      "theoretical_loss": 3.546070103757849,
+      "tokens_seen": 1492647936
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005531214893275557,
+      "loss": 0.0682,
+      "theoretical_loss": 3.546015589714755,
+      "tokens_seen": 1492910080
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005530412453859734,
+      "loss": 0.0714,
+      "theoretical_loss": 3.545961087922782,
+      "tokens_seen": 1493172224
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.000552961001444391,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5459065983770266,
+      "tokens_seen": 1493434368
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005528807575028085,
+      "loss": 0.0668,
+      "theoretical_loss": 3.5458521210725893,
+      "tokens_seen": 1493696512
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005528005135612262,
+      "loss": 0.069,
+      "theoretical_loss": 3.545797656004571,
+      "tokens_seen": 1493958656
+    },
+    {
+      "epoch": 0.45,
+      "objective/train/advantage_avg": 0.0012045002076774836,
+      "objective/train/docs_used": 545793,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2948263883590698,
+      "objective/train/original_loss": 1.2948261499404907,
+      "objective/train/theoretical_loss": 3.545743203168077,
+      "objective/train/tokens_used": 1514680800,
+      "objective/train/value_avg": -0.0071563720703125,
+      "objective/train/value_loss": 0.0002446006110403687,
+      "objective/train/value_max": -7.843971252441406e-05,
+      "objective/train/value_min": -0.2481689453125,
+      "objective/train/value_reward_corr": 0.6621380128230953,
+      "objective/train/value_std": 0.0140380859375,
+      "objective/train/weight_avg": 1.0013121366500854,
+      "objective/train/weighted_lm_loss": 1.2966434955596924,
+      "objective/train/weights_max": 1.1413390636444092,
+      "objective/train/weights_min": 0.3839024305343628,
+      "theoretical_loss": 3.545743203168077,
+      "tokens_seen": 1494220800
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005527202696196437,
+      "loss": 0.0681,
+      "theoretical_loss": 3.545743203168077,
+      "tokens_seen": 1494220800
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005526400256780613,
+      "loss": 0.0693,
+      "theoretical_loss": 3.545688762558216,
+      "tokens_seen": 1494482944
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005525597817364789,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5456343341700984,
+      "tokens_seen": 1494745088
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005524795377948964,
+      "loss": 0.0701,
+      "theoretical_loss": 3.545579917998838,
+      "tokens_seen": 1495007232
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005523992938533142,
+      "loss": 0.069,
+      "theoretical_loss": 3.5455255140395505,
+      "tokens_seen": 1495269376
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005523190499117317,
+      "loss": 0.071,
+      "theoretical_loss": 3.5454711222873554,
+      "tokens_seen": 1495531520
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005522388059701493,
+      "loss": 0.0712,
+      "theoretical_loss": 3.545416742737375,
+      "tokens_seen": 1495793664
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005521585620285669,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5453623753847343,
+      "tokens_seen": 1496055808
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005520783180869845,
+      "loss": 0.0687,
+      "theoretical_loss": 3.54530802022456,
+      "tokens_seen": 1496317952
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.000551998074145402,
+      "loss": 0.0745,
+      "theoretical_loss": 3.545253677251983,
+      "tokens_seen": 1496580096
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005519178302038196,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5451993464621365,
+      "tokens_seen": 1496842240
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005518375862622372,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5451450278501566,
+      "tokens_seen": 1497104384
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005517573423206547,
+      "loss": 0.0737,
+      "theoretical_loss": 3.545090721411182,
+      "tokens_seen": 1497366528
+    },
+    {
+      "epoch": 0.45,
+      "objective/train/advantage_avg": 9.254212636733428e-05,
+      "objective/train/docs_used": 547043,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3230732679367065,
+      "objective/train/original_loss": 1.323073387145996,
+      "objective/train/theoretical_loss": 3.545063572755053,
+      "objective/train/tokens_used": 1517957600,
+      "objective/train/value_avg": -0.0081787109375,
+      "objective/train/value_loss": 0.000149422645336017,
+      "objective/train/value_max": -5.519390106201172e-05,
+      "objective/train/value_min": -0.55078125,
+      "objective/train/value_reward_corr": 0.7170197682857611,
+      "objective/train/value_std": 0.01326751708984375,
+      "objective/train/weight_avg": 1.0001667737960815,
+      "objective/train/weighted_lm_loss": 1.3230582475662231,
+      "objective/train/weights_max": 1.6947335004806519,
+      "objective/train/weights_min": 0.6550559401512146,
+      "theoretical_loss": 3.545063572755053,
+      "tokens_seen": 1497497600
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005516770983790724,
+      "loss": 0.071,
+      "theoretical_loss": 3.545036427140354,
+      "tokens_seen": 1497628672
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.00055159685443749,
+      "loss": 0.0727,
+      "theoretical_loss": 3.544982145032817,
+      "tokens_seen": 1497890816
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005515166104959077,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5449278750837188,
+      "tokens_seen": 1498152960
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005514363665543252,
+      "loss": 0.0734,
+      "theoretical_loss": 3.5448736172882085,
+      "tokens_seen": 1498415104
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005513561226127427,
+      "loss": 0.069,
+      "theoretical_loss": 3.5448193716414393,
+      "tokens_seen": 1498677248
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005512758786711604,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5447651381385668,
+      "tokens_seen": 1498939392
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005511956347295779,
+      "loss": 0.0707,
+      "theoretical_loss": 3.544710916774749,
+      "tokens_seen": 1499201536
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005511153907879955,
+      "loss": 0.07,
+      "theoretical_loss": 3.5446567075451463,
+      "tokens_seen": 1499463680
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005510351468464131,
+      "loss": 0.0713,
+      "theoretical_loss": 3.544602510444924,
+      "tokens_seen": 1499725824
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005509549029048307,
+      "loss": 0.0701,
+      "theoretical_loss": 3.544548325469247,
+      "tokens_seen": 1499987968
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005508746589632483,
+      "loss": 0.0738,
+      "theoretical_loss": 3.5444941526132863,
+      "tokens_seen": 1500250112
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005507944150216659,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5444399918722134,
+      "tokens_seen": 1500512256
+    },
+    {
+      "epoch": 0.45,
+      "objective/train/advantage_avg": -5.096226232126355e-05,
+      "objective/train/docs_used": 548178,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.424842357635498,
+      "objective/train/original_loss": 1.424842357635498,
+      "objective/train/theoretical_loss": 3.5443858432412028,
+      "objective/train/tokens_used": 1521234400,
+      "objective/train/value_avg": -0.007396697998046875,
+      "objective/train/value_loss": 0.0001681852008914575,
+      "objective/train/value_max": -4.869699478149414e-05,
+      "objective/train/value_min": -0.2218017578125,
+      "objective/train/value_reward_corr": 0.7244656461983854,
+      "objective/train/value_std": 0.01325225830078125,
+      "objective/train/weight_avg": 1.0000311136245728,
+      "objective/train/weighted_lm_loss": 1.4248158931732178,
+      "objective/train/weights_max": 1.162267804145813,
+      "objective/train/weights_min": 0.7341392636299133,
+      "theoretical_loss": 3.5443858432412028,
+      "tokens_seen": 1500774400
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005507141710800835,
+      "loss": 0.0683,
+      "theoretical_loss": 3.5443858432412028,
+      "tokens_seen": 1500774400
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.000550633927138501,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5443317067154325,
+      "tokens_seen": 1501036544
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 0.0005505536831969187,
+      "loss": 0.0718,
+      "theoretical_loss": 3.544277582290083,
+      "tokens_seen": 1501298688
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005504734392553362,
+      "loss": 0.0713,
+      "theoretical_loss": 3.544223469960337,
+      "tokens_seen": 1501560832
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005503931953137538,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5441693697213816,
+      "tokens_seen": 1501822976
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005503129513721714,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5441152815684043,
+      "tokens_seen": 1502085120
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005502327074305889,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5440612054965968,
+      "tokens_seen": 1502347264
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005501524634890067,
+      "loss": 0.0722,
+      "theoretical_loss": 3.544007141501154,
+      "tokens_seen": 1502609408
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005500722195474242,
+      "loss": 0.0697,
+      "theoretical_loss": 3.543953089577272,
+      "tokens_seen": 1502871552
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005499919756058418,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5438990497201512,
+      "tokens_seen": 1503133696
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005499117316642594,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5438450219249935,
+      "tokens_seen": 1503395840
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000549831487722677,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5437910061870044,
+      "tokens_seen": 1503657984
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005497512437810945,
+      "loss": 0.0709,
+      "theoretical_loss": 3.543737002501392,
+      "tokens_seen": 1503920128
+    },
+    {
+      "epoch": 0.46,
+      "objective/train/advantage_avg": 0.0003146275703329593,
+      "objective/train/docs_used": 549351,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3942029476165771,
+      "objective/train/original_loss": 1.3942029476165771,
+      "objective/train/theoretical_loss": 3.54371000517673,
+      "objective/train/tokens_used": 1524511200,
+      "objective/train/value_avg": -0.006191253662109375,
+      "objective/train/value_loss": 0.00012938177678734064,
+      "objective/train/value_max": -5.346536636352539e-05,
+      "objective/train/value_min": -0.2010498046875,
+      "objective/train/value_reward_corr": 0.6965117820617344,
+      "objective/train/value_std": 0.01032257080078125,
+      "objective/train/weight_avg": 1.0003749132156372,
+      "objective/train/weighted_lm_loss": 1.3946812152862549,
+      "objective/train/weights_max": 1.142175316810608,
+      "objective/train/weights_min": 0.3683358430862427,
+      "theoretical_loss": 3.54371000517673,
+      "tokens_seen": 1504051200
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005496709998395121,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5436830108633663,
+      "tokens_seen": 1504182272
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005495907558979297,
+      "loss": 0.068,
+      "theoretical_loss": 3.5436290312681415,
+      "tokens_seen": 1504444416
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005495105119563472,
+      "loss": 0.0685,
+      "theoretical_loss": 3.543575063710933,
+      "tokens_seen": 1504706560
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000549430268014765,
+      "loss": 0.0703,
+      "theoretical_loss": 3.543521108186961,
+      "tokens_seen": 1504968704
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005493500240731825,
+      "loss": 0.0732,
+      "theoretical_loss": 3.543467164691445,
+      "tokens_seen": 1505230848
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005492697801316001,
+      "loss": 0.0678,
+      "theoretical_loss": 3.5434132332196113,
+      "tokens_seen": 1505492992
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005491895361900177,
+      "loss": 0.0698,
+      "theoretical_loss": 3.543359313766686,
+      "tokens_seen": 1505755136
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005491092922484352,
+      "loss": 0.0704,
+      "theoretical_loss": 3.543305406327899,
+      "tokens_seen": 1506017280
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005490290483068528,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5432515108984832,
+      "tokens_seen": 1506279424
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005489488043652704,
+      "loss": 0.0673,
+      "theoretical_loss": 3.543197627473673,
+      "tokens_seen": 1506541568
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000548868560423688,
+      "loss": 0.0713,
+      "theoretical_loss": 3.543143756048708,
+      "tokens_seen": 1506803712
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005487883164821056,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5430898966188265,
+      "tokens_seen": 1507065856
+    },
+    {
+      "epoch": 0.46,
+      "objective/train/advantage_avg": 0.0005419878871180117,
+      "objective/train/docs_used": 550589,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4911402463912964,
+      "objective/train/original_loss": 1.491140365600586,
+      "objective/train/theoretical_loss": 3.543036049179274,
+      "objective/train/tokens_used": 1527788000,
+      "objective/train/value_avg": -0.0102386474609375,
+      "objective/train/value_loss": 0.00021217716857790947,
+      "objective/train/value_max": -0.00013446807861328125,
+      "objective/train/value_min": -0.215576171875,
+      "objective/train/value_reward_corr": 0.7336762770578672,
+      "objective/train/value_std": 0.01538848876953125,
+      "objective/train/weight_avg": 1.0006433725357056,
+      "objective/train/weighted_lm_loss": 1.4918407201766968,
+      "objective/train/weights_max": 1.1528605222702026,
+      "objective/train/weights_min": 0.382254034280777,
+      "theoretical_loss": 3.543036049179274,
+      "tokens_seen": 1507328000
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005487080725405232,
+      "loss": 0.0707,
+      "theoretical_loss": 3.543036049179274,
+      "tokens_seen": 1507328000
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005486278285989408,
+      "loss": 0.0731,
+      "theoretical_loss": 3.5429822137252955,
+      "tokens_seen": 1507590144
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005485475846573585,
+      "loss": 0.0696,
+      "theoretical_loss": 3.54292839025214,
+      "tokens_seen": 1507852288
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000548467340715776,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5428745787550593,
+      "tokens_seen": 1508114432
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005483870967741935,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5428207792293076,
+      "tokens_seen": 1508376576
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005483068528326112,
+      "loss": 0.0698,
+      "theoretical_loss": 3.542766991670142,
+      "tokens_seen": 1508638720
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005482266088910287,
+      "loss": 0.0705,
+      "theoretical_loss": 3.542713216072821,
+      "tokens_seen": 1508900864
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005481463649494463,
+      "loss": 0.0722,
+      "theoretical_loss": 3.542659452432608,
+      "tokens_seen": 1509163008
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005480661210078639,
+      "loss": 0.0704,
+      "theoretical_loss": 3.542605700744768,
+      "tokens_seen": 1509425152
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005479858770662815,
+      "loss": 0.0693,
+      "theoretical_loss": 3.542551961004568,
+      "tokens_seen": 1509687296
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000547905633124699,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5424982332072794,
+      "tokens_seen": 1509949440
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005478253891831167,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5424445173481756,
+      "tokens_seen": 1510211584
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005477451452415343,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5423908134225304,
+      "tokens_seen": 1510473728
+    },
+    {
+      "epoch": 0.46,
+      "objective/train/advantage_avg": 0.0009087743819691241,
+      "objective/train/docs_used": 551665,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4157395362854004,
+      "objective/train/original_loss": 1.4157395362854004,
+      "objective/train/theoretical_loss": 3.54236396593328,
+      "objective/train/tokens_used": 1531064800,
+      "objective/train/value_avg": -0.00485992431640625,
+      "objective/train/value_loss": 8.174213144229725e-05,
+      "objective/train/value_max": -3.0934810638427734e-05,
+      "objective/train/value_min": -0.196044921875,
+      "objective/train/value_reward_corr": 0.6648075748225439,
+      "objective/train/value_std": 0.0084686279296875,
+      "objective/train/weight_avg": 1.0009490251541138,
+      "objective/train/weighted_lm_loss": 1.417286992073059,
+      "objective/train/weights_max": 1.1846336126327515,
+      "objective/train/weights_min": 0.6891753673553467,
+      "theoretical_loss": 3.54236396593328,
+      "tokens_seen": 1510604800
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005476649012999519,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5423371214256245,
+      "tokens_seen": 1510735872
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005475846573583695,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5422834413527378,
+      "tokens_seen": 1510998016
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000547504413416787,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5422297731991548,
+      "tokens_seen": 1511260160
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005474241694752047,
+      "loss": 0.0687,
+      "theoretical_loss": 3.542176116960162,
+      "tokens_seen": 1511522304
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005473439255336222,
+      "loss": 0.0717,
+      "theoretical_loss": 3.542122472631048,
+      "tokens_seen": 1511784448
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005472636815920397,
+      "loss": 0.0718,
+      "theoretical_loss": 3.542068840207105,
+      "tokens_seen": 1512046592
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005471834376504575,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5420152196836288,
+      "tokens_seen": 1512308736
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000547103193708875,
+      "loss": 0.0682,
+      "theoretical_loss": 3.541961611055915,
+      "tokens_seen": 1512570880
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005470229497672926,
+      "loss": 0.07,
+      "theoretical_loss": 3.5419080143192643,
+      "tokens_seen": 1512833024
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005469427058257102,
+      "loss": 0.0704,
+      "theoretical_loss": 3.541854429468979,
+      "tokens_seen": 1513095168
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005468624618841278,
+      "loss": 0.071,
+      "theoretical_loss": 3.5418008565003647,
+      "tokens_seen": 1513357312
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005467822179425453,
+      "loss": 0.0718,
+      "theoretical_loss": 3.541747295408729,
+      "tokens_seen": 1513619456
+    },
+    {
+      "epoch": 0.46,
+      "objective/train/advantage_avg": 0.0005864388658665121,
+      "objective/train/docs_used": 552897,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2440524101257324,
+      "objective/train/original_loss": 1.2440524101257324,
+      "objective/train/theoretical_loss": 3.541693746189383,
+      "objective/train/tokens_used": 1534341600,
+      "objective/train/value_avg": -0.0083770751953125,
+      "objective/train/value_loss": 0.0004337378195486963,
+      "objective/train/value_max": -5.8770179748535156e-05,
+      "objective/train/value_min": -0.755859375,
+      "objective/train/value_reward_corr": 0.6793665458707917,
+      "objective/train/value_std": 0.017364501953125,
+      "objective/train/weight_avg": 1.000771403312683,
+      "objective/train/weighted_lm_loss": 1.244329571723938,
+      "objective/train/weights_max": 1.7658472061157227,
+      "objective/train/weights_min": 0.37000179290771484,
+      "theoretical_loss": 3.541693746189383,
+      "tokens_seen": 1513881600
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005467019740009629,
+      "loss": 0.07,
+      "theoretical_loss": 3.541693746189383,
+      "tokens_seen": 1513881600
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005466217300593805,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5416402088376397,
+      "tokens_seen": 1514143744
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000546541486117798,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5415866833488154,
+      "tokens_seen": 1514405888
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005464612421762158,
+      "loss": 0.0707,
+      "theoretical_loss": 3.541533169718228,
+      "tokens_seen": 1514668032
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005463809982346333,
+      "loss": 0.0701,
+      "theoretical_loss": 3.541479667941199,
+      "tokens_seen": 1514930176
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000546300754293051,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5414261780130527,
+      "tokens_seen": 1515192320
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005462205103514685,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5413726999291155,
+      "tokens_seen": 1515454464
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000546140266409886,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5413192336847166,
+      "tokens_seen": 1515716608
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005460600224683037,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5412657792751876,
+      "tokens_seen": 1515978752
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005459797785267212,
+      "loss": 0.0698,
+      "theoretical_loss": 3.541212336695863,
+      "tokens_seen": 1516240896
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005458995345851388,
+      "loss": 0.0703,
+      "theoretical_loss": 3.54115890594208,
+      "tokens_seen": 1516503040
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005458192906435564,
+      "loss": 0.0677,
+      "theoretical_loss": 3.5411054870091787,
+      "tokens_seen": 1516765184
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000545739046701974,
+      "loss": 0.07,
+      "theoretical_loss": 3.541052079892502,
+      "tokens_seen": 1517027328
+    },
+    {
+      "epoch": 0.46,
+      "objective/train/advantage_avg": 0.0009824709268286824,
+      "objective/train/docs_used": 554075,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4332348108291626,
+      "objective/train/original_loss": 1.4332348108291626,
+      "objective/train/theoretical_loss": 3.5410253807637924,
+      "objective/train/tokens_used": 1537618400,
+      "objective/train/value_avg": -0.006805419921875,
+      "objective/train/value_loss": 0.00014581528375856578,
+      "objective/train/value_max": -9.173154830932617e-05,
+      "objective/train/value_min": -0.273193359375,
+      "objective/train/value_reward_corr": 0.6536788814320686,
+      "objective/train/value_std": 0.01123809814453125,
+      "objective/train/weight_avg": 1.0010511875152588,
+      "objective/train/weighted_lm_loss": 1.4342068433761597,
+      "objective/train/weights_max": 1.1899962425231934,
+      "objective/train/weights_min": 0.38768166303634644,
+      "theoretical_loss": 3.5410253807637924,
+      "tokens_seen": 1517158400
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005456588027603916,
+      "loss": 0.0708,
+      "theoretical_loss": 3.540998684587394,
+      "tokens_seen": 1517289472
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005455785588188093,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5409453010892022,
+      "tokens_seen": 1517551616
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005454983148772268,
+      "loss": 0.0696,
+      "theoretical_loss": 3.540891929393278,
+      "tokens_seen": 1517813760
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005454180709356443,
+      "loss": 0.0703,
+      "theoretical_loss": 3.540838569494974,
+      "tokens_seen": 1518075904
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000545337826994062,
+      "loss": 0.0677,
+      "theoretical_loss": 3.540785221389646,
+      "tokens_seen": 1518338048
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005452575830524795,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5407318850726517,
+      "tokens_seen": 1518600192
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005451773391108972,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5406785605393525,
+      "tokens_seen": 1518862336
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005450970951693147,
+      "loss": 0.0714,
+      "theoretical_loss": 3.540625247785111,
+      "tokens_seen": 1519124480
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005450168512277323,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5405719468052945,
+      "tokens_seen": 1519386624
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.00054493660728615,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5405186575952716,
+      "tokens_seen": 1519648768
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005448563633445675,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5404653801504127,
+      "tokens_seen": 1519910912
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005447761194029851,
+      "loss": 0.0713,
+      "theoretical_loss": 3.540412114466093,
+      "tokens_seen": 1520173056
+    },
+    {
+      "epoch": 0.46,
+      "objective/train/advantage_avg": 0.001241142163053155,
+      "objective/train/docs_used": 555119,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.443703293800354,
+      "objective/train/original_loss": 1.443703532218933,
+      "objective/train/theoretical_loss": 3.5403588605376886,
+      "objective/train/tokens_used": 1540895200,
+      "objective/train/value_avg": -0.01012420654296875,
+      "objective/train/value_loss": 0.00023096564109437168,
+      "objective/train/value_max": -6.109476089477539e-05,
+      "objective/train/value_min": -0.2802734375,
+      "objective/train/value_reward_corr": 0.8002673033354555,
+      "objective/train/value_std": 0.019866943359375,
+      "objective/train/weight_avg": 1.001347541809082,
+      "objective/train/weighted_lm_loss": 1.4455903768539429,
+      "objective/train/weights_max": 1.1403095722198486,
+      "objective/train/weights_min": 0.36870554089546204,
+      "theoretical_loss": 3.5403588605376886,
+      "tokens_seen": 1520435200
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005446958754614027,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5403588605376886,
+      "tokens_seen": 1520435200
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005446156315198203,
+      "loss": 0.0682,
+      "theoretical_loss": 3.540305618360578,
+      "tokens_seen": 1520697344
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005445353875782378,
+      "loss": 0.0733,
+      "theoretical_loss": 3.540252387930144,
+      "tokens_seen": 1520959488
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005444551436366555,
+      "loss": 0.0689,
+      "theoretical_loss": 3.540199169241771,
+      "tokens_seen": 1521221632
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000544374899695073,
+      "loss": 0.0711,
+      "theoretical_loss": 3.540145962290845,
+      "tokens_seen": 1521483776
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005442946557534905,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5400927670727573,
+      "tokens_seen": 1521745920
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005442144118119083,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5400395835828986,
+      "tokens_seen": 1522008064
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005441341678703258,
+      "loss": 0.0726,
+      "theoretical_loss": 3.539986411816665,
+      "tokens_seen": 1522270208
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005440539239287434,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5399332517694533,
+      "tokens_seen": 1522532352
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000543973679987161,
+      "loss": 0.0686,
+      "theoretical_loss": 3.539880103436664,
+      "tokens_seen": 1522794496
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005438934360455786,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5398269668136986,
+      "tokens_seen": 1523056640
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005438131921039962,
+      "loss": 0.0677,
+      "theoretical_loss": 3.539773841895964,
+      "tokens_seen": 1523318784
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005437329481624137,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5397207286788666,
+      "tokens_seen": 1523580928
+    },
+    {
+      "epoch": 0.46,
+      "objective/train/advantage_avg": 0.001080079935491085,
+      "objective/train/docs_used": 556367,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4472988843917847,
+      "objective/train/original_loss": 1.4472988843917847,
+      "objective/train/theoretical_loss": 3.5396941764566225,
+      "objective/train/tokens_used": 1544172000,
+      "objective/train/value_avg": -0.006328582763671875,
+      "objective/train/value_loss": 0.00018607915262691677,
+      "objective/train/value_max": -0.00011414289474487305,
+      "objective/train/value_min": -0.2391357421875,
+      "objective/train/value_reward_corr": 0.6196534618081238,
+      "objective/train/value_std": 0.01094818115234375,
+      "objective/train/weight_avg": 1.0011615753173828,
+      "objective/train/weighted_lm_loss": 1.449265956878662,
+      "objective/train/weights_max": 1.1805914640426636,
+      "objective/train/weights_min": 0.3763621747493744,
+      "theoretical_loss": 3.5396941764566225,
+      "tokens_seen": 1523712000
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005436527042208313,
+      "loss": 0.0736,
+      "theoretical_loss": 3.539667627157818,
+      "tokens_seen": 1523843072
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005435724602792489,
+      "loss": 0.0704,
+      "theoretical_loss": 3.53961453732823,
+      "tokens_seen": 1524105216
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005434922163376666,
+      "loss": 0.0697,
+      "theoretical_loss": 3.53956145918552,
+      "tokens_seen": 1524367360
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005434119723960841,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5395083927251045,
+      "tokens_seen": 1524629504
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005433317284545018,
+      "loss": 0.0737,
+      "theoretical_loss": 3.539455337942405,
+      "tokens_seen": 1524891648
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005432514845129193,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5394022948328447,
+      "tokens_seen": 1525153792
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005431712405713368,
+      "loss": 0.069,
+      "theoretical_loss": 3.5393492633918497,
+      "tokens_seen": 1525415936
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005430909966297545,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5392962436148485,
+      "tokens_seen": 1525678080
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000543010752688172,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5392432354972723,
+      "tokens_seen": 1525940224
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005429305087465896,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5391902390345544,
+      "tokens_seen": 1526202368
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005428502648050072,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5391372542221315,
+      "tokens_seen": 1526464512
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005427700208634248,
+      "loss": 0.0692,
+      "theoretical_loss": 3.539084281055443,
+      "tokens_seen": 1526726656
+    },
+    {
+      "epoch": 0.46,
+      "objective/train/advantage_avg": 0.0010113557800650597,
+      "objective/train/docs_used": 557485,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3238643407821655,
+      "objective/train/original_loss": 1.323864459991455,
+      "objective/train/theoretical_loss": 3.5390313195299283,
+      "objective/train/tokens_used": 1547448800,
+      "objective/train/value_avg": -0.00528717041015625,
+      "objective/train/value_loss": 0.0001374837738694623,
+      "objective/train/value_max": -6.711483001708984e-05,
+      "objective/train/value_min": -0.223876953125,
+      "objective/train/value_reward_corr": 0.6038054569835163,
+      "objective/train/value_std": 0.00836944580078125,
+      "objective/train/weight_avg": 1.001071572303772,
+      "objective/train/weighted_lm_loss": 1.3252609968185425,
+      "objective/train/weights_max": 1.0974810123443604,
+      "objective/train/weights_min": 0.368222713470459,
+      "theoretical_loss": 3.5390313195299283,
+      "tokens_seen": 1526988800
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005426897769218424,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5390313195299283,
+      "tokens_seen": 1526988800
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005426095329802601,
+      "loss": 0.0671,
+      "theoretical_loss": 3.538978369641033,
+      "tokens_seen": 1527250944
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005425292890386776,
+      "loss": 0.0698,
+      "theoretical_loss": 3.538925431384203,
+      "tokens_seen": 1527513088
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005424490450970952,
+      "loss": 0.0703,
+      "theoretical_loss": 3.538872504754888,
+      "tokens_seen": 1527775232
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005423688011555128,
+      "loss": 0.0698,
+      "theoretical_loss": 3.538819589748539,
+      "tokens_seen": 1528037376
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005422885572139303,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5387666863606104,
+      "tokens_seen": 1528299520
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000542208313272348,
+      "loss": 0.072,
+      "theoretical_loss": 3.5387137945865588,
+      "tokens_seen": 1528561664
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005421280693307655,
+      "loss": 0.0703,
+      "theoretical_loss": 3.538660914421844,
+      "tokens_seen": 1528823808
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005420478253891831,
+      "loss": 0.072,
+      "theoretical_loss": 3.5386080458619276,
+      "tokens_seen": 1529085952
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005419675814476008,
+      "loss": 0.0709,
+      "theoretical_loss": 3.538555188902274,
+      "tokens_seen": 1529348096
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005418873375060183,
+      "loss": 0.0701,
+      "theoretical_loss": 3.53850234353835,
+      "tokens_seen": 1529610240
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005418070935644359,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5384495097656252,
+      "tokens_seen": 1529872384
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005417268496228535,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5383966875795716,
+      "tokens_seen": 1530134528
+    },
+    {
+      "epoch": 0.46,
+      "objective/train/advantage_avg": 0.0014327875105664134,
+      "objective/train/docs_used": 558616,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3117499351501465,
+      "objective/train/original_loss": 1.3117499351501465,
+      "objective/train/theoretical_loss": 3.5383702808301325,
+      "objective/train/tokens_used": 1550725600,
+      "objective/train/value_avg": -0.00811767578125,
+      "objective/train/value_loss": 0.0003342017880640924,
+      "objective/train/value_max": -9.03010368347168e-05,
+      "objective/train/value_min": -0.2138671875,
+      "objective/train/value_reward_corr": 0.653793128617924,
+      "objective/train/value_std": 0.01491546630859375,
+      "objective/train/weight_avg": 1.0015684366226196,
+      "objective/train/weighted_lm_loss": 1.3147151470184326,
+      "objective/train/weights_max": 1.1417571306228638,
+      "objective/train/weights_min": 0.13709571957588196,
+      "theoretical_loss": 3.5383702808301325,
+      "tokens_seen": 1530265600
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005416466056812711,
+      "loss": 0.0719,
+      "theoretical_loss": 3.538343876975664,
+      "tokens_seen": 1530396672
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005415663617396886,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5382910779493795,
+      "tokens_seen": 1530658816
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005414861177981063,
+      "loss": 0.07,
+      "theoretical_loss": 3.538238290496198,
+      "tokens_seen": 1530920960
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005414058738565238,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5381855146116017,
+      "tokens_seen": 1531183104
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005413256299149413,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5381327502910747,
+      "tokens_seen": 1531445248
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000541245385973359,
+      "loss": 0.071,
+      "theoretical_loss": 3.538079997530105,
+      "tokens_seen": 1531707392
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005411651420317766,
+      "loss": 0.0719,
+      "theoretical_loss": 3.538027256324182,
+      "tokens_seen": 1531969536
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005410848980901943,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5379745266687985,
+      "tokens_seen": 1532231680
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005410046541486118,
+      "loss": 0.0675,
+      "theoretical_loss": 3.537921808559449,
+      "tokens_seen": 1532493824
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005409244102070294,
+      "loss": 0.0699,
+      "theoretical_loss": 3.537869101991631,
+      "tokens_seen": 1532755968
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.000540844166265447,
+      "loss": 0.0731,
+      "theoretical_loss": 3.537816406960845,
+      "tokens_seen": 1533018112
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005407639223238645,
+      "loss": 0.0714,
+      "theoretical_loss": 3.537763723462593,
+      "tokens_seen": 1533280256
+    },
+    {
+      "epoch": 0.46,
+      "objective/train/advantage_avg": 0.00042100262362509966,
+      "objective/train/docs_used": 559697,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4558749198913574,
+      "objective/train/original_loss": 1.4558748006820679,
+      "objective/train/theoretical_loss": 3.537711051492379,
+      "objective/train/tokens_used": 1554002400,
+      "objective/train/value_avg": -0.00927734375,
+      "objective/train/value_loss": 0.0001727818016661331,
+      "objective/train/value_max": -6.109476089477539e-05,
+      "objective/train/value_min": -0.333984375,
+      "objective/train/value_reward_corr": 0.7564583381485481,
+      "objective/train/value_std": 0.0146026611328125,
+      "objective/train/weight_avg": 1.0005054473876953,
+      "objective/train/weighted_lm_loss": 1.4559611082077026,
+      "objective/train/weights_max": 1.1561311483383179,
+      "objective/train/weights_min": 0.6095831990242004,
+      "theoretical_loss": 3.537711051492379,
+      "tokens_seen": 1533542400
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005406836783822821,
+      "loss": 0.0736,
+      "theoretical_loss": 3.537711051492379,
+      "tokens_seen": 1533542400
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005406034344406997,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5376583910457127,
+      "tokens_seen": 1533804544
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005405231904991174,
+      "loss": 0.0711,
+      "theoretical_loss": 3.537605742118102,
+      "tokens_seen": 1534066688
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 0.0005404429465575349,
+      "loss": 0.0721,
+      "theoretical_loss": 3.537553104705061,
+      "tokens_seen": 1534328832
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005403627026159526,
+      "loss": 0.0747,
+      "theoretical_loss": 3.5375004788021043,
+      "tokens_seen": 1534590976
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005402824586743701,
+      "loss": 0.0724,
+      "theoretical_loss": 3.537447864404749,
+      "tokens_seen": 1534853120
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005402022147327876,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5373952615085154,
+      "tokens_seen": 1535115264
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005401219707912053,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5373426701089263,
+      "tokens_seen": 1535377408
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005400417268496228,
+      "loss": 0.0728,
+      "theoretical_loss": 3.5372900902015063,
+      "tokens_seen": 1535639552
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005399614829080405,
+      "loss": 0.0733,
+      "theoretical_loss": 3.537237521781784,
+      "tokens_seen": 1535901696
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.000539881238966458,
+      "loss": 0.072,
+      "theoretical_loss": 3.537184964845289,
+      "tokens_seen": 1536163840
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005398009950248756,
+      "loss": 0.0758,
+      "theoretical_loss": 3.5371324193875533,
+      "tokens_seen": 1536425984
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005397207510832933,
+      "loss": 0.0701,
+      "theoretical_loss": 3.537079885404113,
+      "tokens_seen": 1536688128
+    },
+    {
+      "epoch": 0.47,
+      "objective/train/advantage_avg": 0.001109464792534709,
+      "objective/train/docs_used": 561026,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.287007451057434,
+      "objective/train/original_loss": 1.2870075702667236,
+      "objective/train/theoretical_loss": 3.5370536227138585,
+      "objective/train/tokens_used": 1557279200,
+      "objective/train/value_avg": -0.005340576171875,
+      "objective/train/value_loss": 0.0006102016777731478,
+      "objective/train/value_max": -6.109476089477539e-05,
+      "objective/train/value_min": -0.71337890625,
+      "objective/train/value_reward_corr": 0.6549939678459968,
+      "objective/train/value_std": 0.0154876708984375,
+      "objective/train/weight_avg": 1.0012873411178589,
+      "objective/train/weighted_lm_loss": 1.2889665365219116,
+      "objective/train/weights_max": 1.3794907331466675,
+      "objective/train/weights_min": 0.029097316786646843,
+      "theoretical_loss": 3.5370536227138585,
+      "tokens_seen": 1536819200
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005396405071417108,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5370273628905045,
+      "tokens_seen": 1536950272
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005395602632001284,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5369748518422695,
+      "tokens_seen": 1537212416
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.000539480019258546,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5369223522549493,
+      "tokens_seen": 1537474560
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005393997753169636,
+      "loss": 0.0715,
+      "theoretical_loss": 3.53686986412409,
+      "tokens_seen": 1537736704
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005393195313753811,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5368173874452378,
+      "tokens_seen": 1537998848
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005392392874337988,
+      "loss": 0.0716,
+      "theoretical_loss": 3.536764922213944,
+      "tokens_seen": 1538260992
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005391590434922163,
+      "loss": 0.0711,
+      "theoretical_loss": 3.536712468425761,
+      "tokens_seen": 1538523136
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005390787995506339,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5366600260762433,
+      "tokens_seen": 1538785280
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005389985556090516,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5366075951609486,
+      "tokens_seen": 1539047424
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005389183116674691,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5365551756754376,
+      "tokens_seen": 1539309568
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005388380677258867,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5365027676152714,
+      "tokens_seen": 1539571712
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005387578237843043,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5364503709760164,
+      "tokens_seen": 1539833856
+    },
+    {
+      "epoch": 0.47,
+      "objective/train/advantage_avg": 0.0008831744198687375,
+      "objective/train/docs_used": 562258,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.411902904510498,
+      "objective/train/original_loss": 1.4119031429290771,
+      "objective/train/theoretical_loss": 3.536397985753239,
+      "objective/train/tokens_used": 1560556000,
+      "objective/train/value_avg": -0.006740570068359375,
+      "objective/train/value_loss": 0.00011778049520216882,
+      "objective/train/value_max": -9.459257125854492e-05,
+      "objective/train/value_min": -0.2239990234375,
+      "objective/train/value_reward_corr": 0.6256750582579078,
+      "objective/train/value_std": 0.010040283203125,
+      "objective/train/weight_avg": 1.0009410381317139,
+      "objective/train/weighted_lm_loss": 1.4133871793746948,
+      "objective/train/weights_max": 1.2288204431533813,
+      "objective/train/weights_min": 0.6369590759277344,
+      "theoretical_loss": 3.536397985753239,
+      "tokens_seen": 1540096000
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005386775798427219,
+      "loss": 0.0733,
+      "theoretical_loss": 3.536397985753239,
+      "tokens_seen": 1540096000
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005385973359011395,
+      "loss": 0.0762,
+      "theoretical_loss": 3.53634561194251,
+      "tokens_seen": 1540358144
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005385170919595571,
+      "loss": 0.0726,
+      "theoretical_loss": 3.5362932495394013,
+      "tokens_seen": 1540620288
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005384368480179746,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5362408985394875,
+      "tokens_seen": 1540882432
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005383566040763922,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5361885589383464,
+      "tokens_seen": 1541144576
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005382763601348099,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5361362307315574,
+      "tokens_seen": 1541406720
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005381961161932274,
+      "loss": 0.0728,
+      "theoretical_loss": 3.5360839139147036,
+      "tokens_seen": 1541668864
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005381158722516451,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5360316084833685,
+      "tokens_seen": 1541931008
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005380356283100626,
+      "loss": 0.0705,
+      "theoretical_loss": 3.53597931443314,
+      "tokens_seen": 1542193152
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005379553843684802,
+      "loss": 0.0724,
+      "theoretical_loss": 3.535927031759608,
+      "tokens_seen": 1542455296
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005378751404268978,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5358747604583636,
+      "tokens_seen": 1542717440
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005377948964853153,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5358225005250024,
+      "tokens_seen": 1542979584
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005377146525437329,
+      "loss": 0.0736,
+      "theoretical_loss": 3.535770251955121,
+      "tokens_seen": 1543241728
+    },
+    {
+      "epoch": 0.47,
+      "objective/train/advantage_avg": 0.0005263619241304696,
+      "objective/train/docs_used": 563510,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4112048149108887,
+      "objective/train/original_loss": 1.4112045764923096,
+      "objective/train/theoretical_loss": 3.5357441319301106,
+      "objective/train/tokens_used": 1563832800,
+      "objective/train/value_avg": -0.0087127685546875,
+      "objective/train/value_loss": 0.000459319562651217,
+      "objective/train/value_max": -3.820657730102539e-05,
+      "objective/train/value_min": -0.329833984375,
+      "objective/train/value_reward_corr": 0.6188310861340889,
+      "objective/train/value_std": 0.01538848876953125,
+      "objective/train/weight_avg": 1.0007210969924927,
+      "objective/train/weighted_lm_loss": 1.4122984409332275,
+      "objective/train/weights_max": 1.3907371759414673,
+      "objective/train/weights_min": 0.3745975196361542,
+      "theoretical_loss": 3.5357441319301106,
+      "tokens_seen": 1543372800
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005376344086021505,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5357180147443197,
+      "tokens_seen": 1543503872
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005375541646605681,
+      "loss": 0.073,
+      "theoretical_loss": 3.5356657888881986,
+      "tokens_seen": 1543766016
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005374739207189858,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5356135743823636,
+      "tokens_seen": 1544028160
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005373936767774034,
+      "loss": 0.0724,
+      "theoretical_loss": 3.5355613712224203,
+      "tokens_seen": 1544290304
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005373134328358209,
+      "loss": 0.0745,
+      "theoretical_loss": 3.53550917940398,
+      "tokens_seen": 1544552448
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005372331888942385,
+      "loss": 0.0692,
+      "theoretical_loss": 3.535456998922652,
+      "tokens_seen": 1544814592
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005371529449526561,
+      "loss": 0.0712,
+      "theoretical_loss": 3.535404829774052,
+      "tokens_seen": 1545076736
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005370727010110736,
+      "loss": 0.0698,
+      "theoretical_loss": 3.535352671953796,
+      "tokens_seen": 1545338880
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005369924570694913,
+      "loss": 0.0725,
+      "theoretical_loss": 3.5353005254575027,
+      "tokens_seen": 1545601024
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005369122131279088,
+      "loss": 0.0769,
+      "theoretical_loss": 3.5352483902807945,
+      "tokens_seen": 1545863168
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005368319691863264,
+      "loss": 0.0719,
+      "theoretical_loss": 3.535196266419295,
+      "tokens_seen": 1546125312
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005367517252447441,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5351441538686306,
+      "tokens_seen": 1546387456
+    },
+    {
+      "epoch": 0.47,
+      "objective/train/advantage_avg": 0.00033845813595689833,
+      "objective/train/docs_used": 564726,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3939182758331299,
+      "objective/train/original_loss": 1.3939182758331299,
+      "objective/train/theoretical_loss": 3.535092052624429,
+      "objective/train/tokens_used": 1567109600,
+      "objective/train/value_avg": -0.00481414794921875,
+      "objective/train/value_loss": 0.0001568766892887652,
+      "objective/train/value_max": -4.83393669128418e-05,
+      "objective/train/value_min": -0.64111328125,
+      "objective/train/value_reward_corr": 0.6628277956239046,
+      "objective/train/value_std": 0.01015472412109375,
+      "objective/train/weight_avg": 1.000410556793213,
+      "objective/train/weighted_lm_loss": 1.3948057889938354,
+      "objective/train/weights_max": 1.4327837228775024,
+      "objective/train/weights_min": 0.3745289444923401,
+      "theoretical_loss": 3.535092052624429,
+      "tokens_seen": 1546649600
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005366714813031616,
+      "loss": 0.0688,
+      "theoretical_loss": 3.535092052624429,
+      "tokens_seen": 1546649600
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005365912373615792,
+      "loss": 0.0735,
+      "theoretical_loss": 3.5350399626823226,
+      "tokens_seen": 1546911744
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005365109934199968,
+      "loss": 0.0729,
+      "theoretical_loss": 3.534987884037945,
+      "tokens_seen": 1547173888
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005364307494784144,
+      "loss": 0.072,
+      "theoretical_loss": 3.5349358166869314,
+      "tokens_seen": 1547436032
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005363505055368319,
+      "loss": 0.072,
+      "theoretical_loss": 3.534883760624921,
+      "tokens_seen": 1547698176
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005362702615952496,
+      "loss": 0.0764,
+      "theoretical_loss": 3.534831715847555,
+      "tokens_seen": 1547960320
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005361900176536671,
+      "loss": 0.0732,
+      "theoretical_loss": 3.534779682350475,
+      "tokens_seen": 1548222464
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005361097737120849,
+      "loss": 0.0747,
+      "theoretical_loss": 3.534727660129329,
+      "tokens_seen": 1548484608
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005360295297705024,
+      "loss": 0.0711,
+      "theoretical_loss": 3.534675649179764,
+      "tokens_seen": 1548746752
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005359492858289199,
+      "loss": 0.0735,
+      "theoretical_loss": 3.53462364949743,
+      "tokens_seen": 1549008896
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005358690418873376,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5345716610779814,
+      "tokens_seen": 1549271040
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005357887979457551,
+      "loss": 0.072,
+      "theoretical_loss": 3.5345196839170723,
+      "tokens_seen": 1549533184
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005357085540041727,
+      "loss": 0.0738,
+      "theoretical_loss": 3.534467718010361,
+      "tokens_seen": 1549795328
+    },
+    {
+      "epoch": 0.47,
+      "objective/train/advantage_avg": 0.0002879749226849526,
+      "objective/train/docs_used": 565981,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.505765438079834,
+      "objective/train/original_loss": 1.5057655572891235,
+      "objective/train/theoretical_loss": 3.534441739275974,
+      "objective/train/tokens_used": 1570386400,
+      "objective/train/value_avg": -0.007205963134765625,
+      "objective/train/value_loss": 0.00035114213824272156,
+      "objective/train/value_max": -6.35385513305664e-05,
+      "objective/train/value_min": -0.79736328125,
+      "objective/train/value_reward_corr": 0.6967833657591274,
+      "objective/train/value_std": 0.0176544189453125,
+      "objective/train/weight_avg": 1.0004457235336304,
+      "objective/train/weighted_lm_loss": 1.5057909488677979,
+      "objective/train/weights_max": 1.3775582313537598,
+      "objective/train/weights_min": 0.3902336657047272,
+      "theoretical_loss": 3.534441739275974,
+      "tokens_seen": 1549926400
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005356283100625903,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5344157633535085,
+      "tokens_seen": 1550057472
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005355480661210079,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5343638199421763,
+      "tokens_seen": 1550319616
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005354678221794254,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5343118877720294,
+      "tokens_seen": 1550581760
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.000535387578237843,
+      "loss": 0.0716,
+      "theoretical_loss": 3.534259966838736,
+      "tokens_seen": 1550843904
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005353073342962607,
+      "loss": 0.0697,
+      "theoretical_loss": 3.534208057137966,
+      "tokens_seen": 1551106048
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005352270903546782,
+      "loss": 0.0728,
+      "theoretical_loss": 3.5341561586653905,
+      "tokens_seen": 1551368192
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005351468464130959,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5341042714166853,
+      "tokens_seen": 1551630336
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005350666024715134,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5340523953875267,
+      "tokens_seen": 1551892480
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005349863585299311,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5340005305735946,
+      "tokens_seen": 1552154624
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005349061145883486,
+      "loss": 0.0686,
+      "theoretical_loss": 3.533948676970571,
+      "tokens_seen": 1552416768
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005348258706467661,
+      "loss": 0.0723,
+      "theoretical_loss": 3.533896834574139,
+      "tokens_seen": 1552678912
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005347456267051838,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5338450033799864,
+      "tokens_seen": 1552941056
+    },
+    {
+      "epoch": 0.47,
+      "objective/train/advantage_avg": -0.000528849835973233,
+      "objective/train/docs_used": 567156,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4279303550720215,
+      "objective/train/original_loss": 1.4279303550720215,
+      "objective/train/theoretical_loss": 3.533793183383802,
+      "objective/train/tokens_used": 1573663200,
+      "objective/train/value_avg": -0.007747650146484375,
+      "objective/train/value_loss": 0.00032320283935405314,
+      "objective/train/value_max": -5.650520324707031e-05,
+      "objective/train/value_min": -0.2115478515625,
+      "objective/train/value_reward_corr": 0.5905321513810176,
+      "objective/train/value_std": 0.012298583984375,
+      "objective/train/weight_avg": 0.999616265296936,
+      "objective/train/weighted_lm_loss": 1.4269369840621948,
+      "objective/train/weights_max": 1.101507544517517,
+      "objective/train/weights_min": 0.3709147274494171,
+      "theoretical_loss": 3.533793183383802,
+      "tokens_seen": 1553203200
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005346653827636013,
+      "loss": 0.0695,
+      "theoretical_loss": 3.533793183383802,
+      "tokens_seen": 1553203200
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.000534585138822019,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5337413745812767,
+      "tokens_seen": 1553465344
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005345048948804366,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5336895769681043,
+      "tokens_seen": 1553727488
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005344246509388542,
+      "loss": 0.0697,
+      "theoretical_loss": 3.533637790539981,
+      "tokens_seen": 1553989632
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005343444069972717,
+      "loss": 0.0672,
+      "theoretical_loss": 3.533586015292606,
+      "tokens_seen": 1554251776
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005342641630556893,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5335342512216794,
+      "tokens_seen": 1554513920
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005341839191141069,
+      "loss": 0.073,
+      "theoretical_loss": 3.5334824983229045,
+      "tokens_seen": 1554776064
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005341036751725244,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5334307565919874,
+      "tokens_seen": 1555038208
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005340234312309421,
+      "loss": 0.0695,
+      "theoretical_loss": 3.533379026024636,
+      "tokens_seen": 1555300352
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005339431872893596,
+      "loss": 0.072,
+      "theoretical_loss": 3.53332730661656,
+      "tokens_seen": 1555562496
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005338629433477772,
+      "loss": 0.0717,
+      "theoretical_loss": 3.533275598363473,
+      "tokens_seen": 1555824640
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005337826994061949,
+      "loss": 0.0677,
+      "theoretical_loss": 3.53322390126109,
+      "tokens_seen": 1556086784
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005337024554646124,
+      "loss": 0.0731,
+      "theoretical_loss": 3.533172215305129,
+      "tokens_seen": 1556348928
+    },
+    {
+      "epoch": 0.47,
+      "objective/train/advantage_avg": 0.0005209646187722683,
+      "objective/train/docs_used": 568457,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3047668933868408,
+      "objective/train/original_loss": 1.3047666549682617,
+      "objective/train/theoretical_loss": 3.5331463765057185,
+      "objective/train/tokens_used": 1576940000,
+      "objective/train/value_avg": -0.006496429443359375,
+      "objective/train/value_loss": 0.00021889428899157792,
+      "objective/train/value_max": -7.086992263793945e-05,
+      "objective/train/value_min": -0.93701171875,
+      "objective/train/value_reward_corr": 0.6957019665112554,
+      "objective/train/value_std": 0.013275146484375,
+      "objective/train/weight_avg": 1.0006130933761597,
+      "objective/train/weighted_lm_loss": 1.305483102798462,
+      "objective/train/weights_max": 1.1495888233184814,
+      "objective/train/weights_min": 0.23035363852977753,
+      "theoretical_loss": 3.5331463765057185,
+      "tokens_seen": 1556480000
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005336222115230301,
+      "loss": 0.0681,
+      "theoretical_loss": 3.533120540491309,
+      "tokens_seen": 1556611072
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005335419675814476,
+      "loss": 0.0728,
+      "theoretical_loss": 3.533068876815352,
+      "tokens_seen": 1556873216
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005334617236398652,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5330172242729834,
+      "tokens_seen": 1557135360
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005333814796982828,
+      "loss": 0.072,
+      "theoretical_loss": 3.53296558285993,
+      "tokens_seen": 1557397504
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005333012357567004,
+      "loss": 0.0671,
+      "theoretical_loss": 3.532913952571921,
+      "tokens_seen": 1557659648
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005332209918151179,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5328623334046885,
+      "tokens_seen": 1557921792
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005331407478735357,
+      "loss": 0.0698,
+      "theoretical_loss": 3.532810725353966,
+      "tokens_seen": 1558183936
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005330605039319532,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5327591284154893,
+      "tokens_seen": 1558446080
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005329802599903707,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5327075425849985,
+      "tokens_seen": 1558708224
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005329000160487884,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5326559678582337,
+      "tokens_seen": 1558970368
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005328197721072059,
+      "loss": 0.0701,
+      "theoretical_loss": 3.532604404230939,
+      "tokens_seen": 1559232512
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005327395281656235,
+      "loss": 0.0688,
+      "theoretical_loss": 3.53255285169886,
+      "tokens_seen": 1559494656
+    },
+    {
+      "epoch": 0.47,
+      "objective/train/advantage_avg": 7.476971950381994e-05,
+      "objective/train/docs_used": 569616,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4254400730133057,
+      "objective/train/original_loss": 1.4254400730133057,
+      "objective/train/theoretical_loss": 3.5325013102577443,
+      "objective/train/tokens_used": 1580216800,
+      "objective/train/value_avg": -0.008270263671875,
+      "objective/train/value_loss": 0.00038079096702858806,
+      "objective/train/value_max": -8.028745651245117e-05,
+      "objective/train/value_min": -0.393310546875,
+      "objective/train/value_reward_corr": 0.7088766731470348,
+      "objective/train/value_std": 0.0188140869140625,
+      "objective/train/weight_avg": 1.0002461671829224,
+      "objective/train/weighted_lm_loss": 1.4250798225402832,
+      "objective/train/weights_max": 1.437957763671875,
+      "objective/train/weights_min": 0.36907142400741577,
+      "theoretical_loss": 3.5325013102577443,
+      "tokens_seen": 1559756800
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005326592842240411,
+      "loss": 0.072,
+      "theoretical_loss": 3.5325013102577443,
+      "tokens_seen": 1559756800
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005325790402824587,
+      "loss": 0.0729,
+      "theoretical_loss": 3.532449779903343,
+      "tokens_seen": 1560018944
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005324987963408762,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5323982606314086,
+      "tokens_seen": 1560281088
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005324185523992938,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5323467524376966,
+      "tokens_seen": 1560543232
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005323383084577115,
+      "loss": 0.0715,
+      "theoretical_loss": 3.532295255317964,
+      "tokens_seen": 1560805376
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005322580645161291,
+      "loss": 0.0673,
+      "theoretical_loss": 3.532243769267971,
+      "tokens_seen": 1561067520
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005321778205745467,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5321922942834796,
+      "tokens_seen": 1561329664
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005320975766329642,
+      "loss": 0.0739,
+      "theoretical_loss": 3.5321408303602544,
+      "tokens_seen": 1561591808
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005320173326913819,
+      "loss": 0.0702,
+      "theoretical_loss": 3.532089377494062,
+      "tokens_seen": 1561853952
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005319370887497994,
+      "loss": 0.0754,
+      "theoretical_loss": 3.532037935680672,
+      "tokens_seen": 1562116096
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005318568448082169,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5319865049158556,
+      "tokens_seen": 1562378240
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005317766008666346,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5319350851953866,
+      "tokens_seen": 1562640384
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005316963569250521,
+      "loss": 0.0706,
+      "theoretical_loss": 3.531883676515041,
+      "tokens_seen": 1562902528
+    },
+    {
+      "epoch": 0.47,
+      "objective/train/advantage_avg": 0.00038564883288927376,
+      "objective/train/docs_used": 570864,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2100456953048706,
+      "objective/train/original_loss": 1.210045576095581,
+      "objective/train/theoretical_loss": 3.5318579763135953,
+      "objective/train/tokens_used": 1583493600,
+      "objective/train/value_avg": -0.01094818115234375,
+      "objective/train/value_loss": 0.00040198001079261303,
+      "objective/train/value_max": -0.00010389089584350586,
+      "objective/train/value_min": -0.431884765625,
+      "objective/train/value_reward_corr": 0.6607573354620048,
+      "objective/train/value_std": 0.0176239013671875,
+      "objective/train/weight_avg": 1.0005624294281006,
+      "objective/train/weighted_lm_loss": 1.2105127573013306,
+      "objective/train/weights_max": 1.5052146911621094,
+      "objective/train/weights_min": 0.3692784309387207,
+      "theoretical_loss": 3.5318579763135953,
+      "tokens_seen": 1563033600
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005316161129834697,
+      "loss": 0.073,
+      "theoretical_loss": 3.5318322788705974,
+      "tokens_seen": 1563164672
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005315358690418874,
+      "loss": 0.0679,
+      "theoretical_loss": 3.531780892257837,
+      "tokens_seen": 1563426816
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.000531455625100305,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5317295166725424,
+      "tokens_seen": 1563688960
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005313753811587225,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5316781521104996,
+      "tokens_seen": 1563951104
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005312951372171401,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5316267985674954,
+      "tokens_seen": 1564213248
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005312148932755577,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5315754560393207,
+      "tokens_seen": 1564475392
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005311346493339753,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5315241245217672,
+      "tokens_seen": 1564737536
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005310544053923929,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5314728040106305,
+      "tokens_seen": 1564999680
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005309741614508104,
+      "loss": 0.0729,
+      "theoretical_loss": 3.531421494501707,
+      "tokens_seen": 1565261824
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005308939175092282,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5313701959907955,
+      "tokens_seen": 1565523968
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005308136735676457,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5313189084736987,
+      "tokens_seen": 1565786112
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005307334296260632,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5312676319462195,
+      "tokens_seen": 1566048256
+    },
+    {
+      "epoch": 0.47,
+      "objective/train/advantage_avg": -0.000856450293213129,
+      "objective/train/docs_used": 572070,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4109565019607544,
+      "objective/train/original_loss": 1.4109562635421753,
+      "objective/train/theoretical_loss": 3.531216366404165,
+      "objective/train/tokens_used": 1586770400,
+      "objective/train/value_avg": -0.0079193115234375,
+      "objective/train/value_loss": 0.0001714797253953293,
+      "objective/train/value_max": -3.707408905029297e-05,
+      "objective/train/value_min": -0.79638671875,
+      "objective/train/value_reward_corr": 0.8089750589892425,
+      "objective/train/value_std": 0.01605224609375,
+      "objective/train/weight_avg": 0.999224066734314,
+      "objective/train/weighted_lm_loss": 1.409802794456482,
+      "objective/train/weights_max": 1.2366453409194946,
+      "objective/train/weights_min": 0.3772245943546295,
+      "theoretical_loss": 3.531216366404165,
+      "tokens_seen": 1566310400
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005306531856844809,
+      "loss": 0.0684,
+      "theoretical_loss": 3.531216366404165,
+      "tokens_seen": 1566310400
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005305729417428984,
+      "loss": 0.0682,
+      "theoretical_loss": 3.531165111843343,
+      "tokens_seen": 1566572544
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.000530492697801316,
+      "loss": 0.0709,
+      "theoretical_loss": 3.531113868259565,
+      "tokens_seen": 1566834688
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005304124538597336,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5310626356486434,
+      "tokens_seen": 1567096832
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 0.0005303322099181512,
+      "loss": 0.069,
+      "theoretical_loss": 3.5310114140063944,
+      "tokens_seen": 1567358976
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005302519659765687,
+      "loss": 0.0702,
+      "theoretical_loss": 3.530960203328635,
+      "tokens_seen": 1567621120
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005301717220349863,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5309090036111854,
+      "tokens_seen": 1567883264
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000530091478093404,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5308578148498686,
+      "tokens_seen": 1568145408
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005300112341518215,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5308066370405076,
+      "tokens_seen": 1568407552
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005299309902102392,
+      "loss": 0.0695,
+      "theoretical_loss": 3.530755470178931,
+      "tokens_seen": 1568669696
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005298507462686567,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5307043142609666,
+      "tokens_seen": 1568931840
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005297705023270744,
+      "loss": 0.0694,
+      "theoretical_loss": 3.530653169282447,
+      "tokens_seen": 1569193984
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005296902583854919,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5306020352392053,
+      "tokens_seen": 1569456128
+    },
+    {
+      "epoch": 0.48,
+      "objective/train/advantage_avg": 0.0017290321411564946,
+      "objective/train/docs_used": 573246,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3300449848175049,
+      "objective/train/original_loss": 1.3300449848175049,
+      "objective/train/theoretical_loss": 3.5305764723170125,
+      "objective/train/tokens_used": 1590047200,
+      "objective/train/value_avg": -0.00569915771484375,
+      "objective/train/value_loss": 0.00010214610665570945,
+      "objective/train/value_max": -8.153915405273438e-05,
+      "objective/train/value_min": -0.9501953125,
+      "objective/train/value_reward_corr": 0.8309916736042006,
+      "objective/train/value_std": 0.01422119140625,
+      "objective/train/weight_avg": 1.0017763376235962,
+      "objective/train/weighted_lm_loss": 1.3328149318695068,
+      "objective/train/weights_max": 1.5367772579193115,
+      "objective/train/weights_min": 0.3692037761211395,
+      "theoretical_loss": 3.5305764723170125,
+      "tokens_seen": 1569587200
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005296100144439095,
+      "loss": 0.07,
+      "theoretical_loss": 3.5305509121270777,
+      "tokens_seen": 1569718272
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005295297705023271,
+      "loss": 0.0663,
+      "theoretical_loss": 3.5304997999419028,
+      "tokens_seen": 1569980416
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005294495265607446,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5304486986795203,
+      "tokens_seen": 1570242560
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005293692826191623,
+      "loss": 0.07,
+      "theoretical_loss": 3.5303976083357735,
+      "tokens_seen": 1570504704
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005292890386775799,
+      "loss": 0.0649,
+      "theoretical_loss": 3.5303465289065077,
+      "tokens_seen": 1570766848
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005292087947359975,
+      "loss": 0.0674,
+      "theoretical_loss": 3.5302954603875696,
+      "tokens_seen": 1571028992
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000529128550794415,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5302444027748106,
+      "tokens_seen": 1571291136
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005290483068528327,
+      "loss": 0.0701,
+      "theoretical_loss": 3.530193356064081,
+      "tokens_seen": 1571553280
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005289680629112502,
+      "loss": 0.0689,
+      "theoretical_loss": 3.530142320251235,
+      "tokens_seen": 1571815424
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005288878189696677,
+      "loss": 0.0688,
+      "theoretical_loss": 3.5300912953321304,
+      "tokens_seen": 1572077568
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005288075750280854,
+      "loss": 0.0686,
+      "theoretical_loss": 3.530040281302625,
+      "tokens_seen": 1572339712
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005287273310865029,
+      "loss": 0.0736,
+      "theoretical_loss": 3.5299892781585793,
+      "tokens_seen": 1572601856
+    },
+    {
+      "epoch": 0.48,
+      "objective/train/advantage_avg": 0.0004105140978936106,
+      "objective/train/docs_used": 574423,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.474505066871643,
+      "objective/train/original_loss": 1.4745049476623535,
+      "objective/train/theoretical_loss": 3.529938285895858,
+      "objective/train/tokens_used": 1593324000,
+      "objective/train/value_avg": -0.00910186767578125,
+      "objective/train/value_loss": 0.00025253358762711287,
+      "objective/train/value_max": -6.35385513305664e-05,
+      "objective/train/value_min": -0.323486328125,
+      "objective/train/value_reward_corr": 0.7612959342784285,
+      "objective/train/value_std": 0.016571044921875,
+      "objective/train/weight_avg": 1.0005269050598145,
+      "objective/train/weighted_lm_loss": 1.4759199619293213,
+      "objective/train/weights_max": 1.180879831314087,
+      "objective/train/weights_min": 0.3688413202762604,
+      "theoretical_loss": 3.529938285895858,
+      "tokens_seen": 1572864000
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005286470871449207,
+      "loss": 0.0722,
+      "theoretical_loss": 3.529938285895858,
+      "tokens_seen": 1572864000
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005285668432033382,
+      "loss": 0.0672,
+      "theoretical_loss": 3.529887304510326,
+      "tokens_seen": 1573126144
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005284865992617558,
+      "loss": 0.0667,
+      "theoretical_loss": 3.5298363339978502,
+      "tokens_seen": 1573388288
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005284063553201734,
+      "loss": 0.0672,
+      "theoretical_loss": 3.529785374354302,
+      "tokens_seen": 1573650432
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005283261113785909,
+      "loss": 0.0688,
+      "theoretical_loss": 3.529734425575553,
+      "tokens_seen": 1573912576
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005282458674370085,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5296834876574783,
+      "tokens_seen": 1574174720
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005281656234954261,
+      "loss": 0.0702,
+      "theoretical_loss": 3.529632560595954,
+      "tokens_seen": 1574436864
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005280853795538437,
+      "loss": 0.0676,
+      "theoretical_loss": 3.5295816443868593,
+      "tokens_seen": 1574699008
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005280051356122612,
+      "loss": 0.0643,
+      "theoretical_loss": 3.529530739026076,
+      "tokens_seen": 1574961152
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000527924891670679,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5294798445094875,
+      "tokens_seen": 1575223296
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005278446477290965,
+      "loss": 0.0702,
+      "theoretical_loss": 3.529428960832979,
+      "tokens_seen": 1575485440
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000527764403787514,
+      "loss": 0.068,
+      "theoretical_loss": 3.5293780879924395,
+      "tokens_seen": 1575747584
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005276841598459317,
+      "loss": 0.0665,
+      "theoretical_loss": 3.529327225983759,
+      "tokens_seen": 1576009728
+    },
+    {
+      "epoch": 0.48,
+      "objective/train/advantage_avg": 0.00027555739507079124,
+      "objective/train/docs_used": 575737,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2302134037017822,
+      "objective/train/original_loss": 1.2302134037017822,
+      "objective/train/theoretical_loss": 3.529301799040082,
+      "objective/train/tokens_used": 1596600800,
+      "objective/train/value_avg": -0.0084075927734375,
+      "objective/train/value_loss": 0.00025561146321706474,
+      "objective/train/value_max": -5.561113357543945e-05,
+      "objective/train/value_min": -0.5439453125,
+      "objective/train/value_reward_corr": 0.6844903909527469,
+      "objective/train/value_std": 0.0140838623046875,
+      "objective/train/weight_avg": 1.0003907680511475,
+      "objective/train/weighted_lm_loss": 1.2300145626068115,
+      "objective/train/weights_max": 1.5279780626296997,
+      "objective/train/weights_min": 0.3712006211280823,
+      "theoretical_loss": 3.529301799040082,
+      "tokens_seen": 1576140800
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005276039159043492,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5292763748028295,
+      "tokens_seen": 1576271872
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005275236719627668,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5292255344455468,
+      "tokens_seen": 1576534016
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005274434280211844,
+      "loss": 0.0676,
+      "theoretical_loss": 3.529174704907807,
+      "tokens_seen": 1576796160
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000527363184079602,
+      "loss": 0.0691,
+      "theoretical_loss": 3.529123886185509,
+      "tokens_seen": 1577058304
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005272829401380196,
+      "loss": 0.067,
+      "theoretical_loss": 3.529073078274556,
+      "tokens_seen": 1577320448
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005272026961964371,
+      "loss": 0.068,
+      "theoretical_loss": 3.5290222811708505,
+      "tokens_seen": 1577582592
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005271224522548548,
+      "loss": 0.0702,
+      "theoretical_loss": 3.528971494870299,
+      "tokens_seen": 1577844736
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005270422083132724,
+      "loss": 0.0663,
+      "theoretical_loss": 3.5289207193688092,
+      "tokens_seen": 1578106880
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.00052696196437169,
+      "loss": 0.0671,
+      "theoretical_loss": 3.5288699546622913,
+      "tokens_seen": 1578369024
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005268817204301075,
+      "loss": 0.07,
+      "theoretical_loss": 3.528819200746659,
+      "tokens_seen": 1578631168
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005268014764885252,
+      "loss": 0.071,
+      "theoretical_loss": 3.5287684576178258,
+      "tokens_seen": 1578893312
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005267212325469427,
+      "loss": 0.071,
+      "theoretical_loss": 3.5287177252717097,
+      "tokens_seen": 1579155456
+    },
+    {
+      "epoch": 0.48,
+      "objective/train/advantage_avg": 0.0005158972926437855,
+      "objective/train/docs_used": 576827,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.405625343322754,
+      "objective/train/original_loss": 1.405625581741333,
+      "objective/train/theoretical_loss": 3.5286670037042303,
+      "objective/train/tokens_used": 1599877600,
+      "objective/train/value_avg": -0.0048828125,
+      "objective/train/value_loss": 0.00019700209668371826,
+      "objective/train/value_max": -6.556510925292969e-05,
+      "objective/train/value_min": -0.65673828125,
+      "objective/train/value_reward_corr": 0.6370804838393163,
+      "objective/train/value_std": 0.0123443603515625,
+      "objective/train/weight_avg": 1.0006061792373657,
+      "objective/train/weighted_lm_loss": 1.4065194129943848,
+      "objective/train/weights_max": 1.8773905038833618,
+      "objective/train/weights_min": 0.37096279859542847,
+      "theoretical_loss": 3.5286670037042303,
+      "tokens_seen": 1579417600
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005266409886053603,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5286670037042303,
+      "tokens_seen": 1579417600
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005265607446637779,
+      "loss": 0.0683,
+      "theoretical_loss": 3.528616292911309,
+      "tokens_seen": 1579679744
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005264805007221954,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5285655928888686,
+      "tokens_seen": 1579941888
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000526400256780613,
+      "loss": 0.0688,
+      "theoretical_loss": 3.5285149036328356,
+      "tokens_seen": 1580204032
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005263200128390307,
+      "loss": 0.0695,
+      "theoretical_loss": 3.528464225139139,
+      "tokens_seen": 1580466176
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005262397688974483,
+      "loss": 0.073,
+      "theoretical_loss": 3.5284135574037085,
+      "tokens_seen": 1580728320
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005261595249558658,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5283629004224766,
+      "tokens_seen": 1580990464
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005260792810142835,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5283122541913787,
+      "tokens_seen": 1581252608
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000525999037072701,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5282616187063516,
+      "tokens_seen": 1581514752
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005259187931311186,
+      "loss": 0.0687,
+      "theoretical_loss": 3.528210993963334,
+      "tokens_seen": 1581776896
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005258385491895362,
+      "loss": 0.0712,
+      "theoretical_loss": 3.528160379958268,
+      "tokens_seen": 1582039040
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005257583052479537,
+      "loss": 0.0707,
+      "theoretical_loss": 3.528109776687097,
+      "tokens_seen": 1582301184
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005256780613063715,
+      "loss": 0.0721,
+      "theoretical_loss": 3.528059184145767,
+      "tokens_seen": 1582563328
+    },
+    {
+      "epoch": 0.48,
+      "objective/train/advantage_avg": 0.001599782146513462,
+      "objective/train/docs_used": 578056,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2233972549438477,
+      "objective/train/original_loss": 1.2233970165252686,
+      "objective/train/theoretical_loss": 3.5280338918975263,
+      "objective/train/tokens_used": 1603154400,
+      "objective/train/value_avg": -0.00786590576171875,
+      "objective/train/value_loss": 0.00015577516751363873,
+      "objective/train/value_max": -6.711483001708984e-05,
+      "objective/train/value_min": -0.26220703125,
+      "objective/train/value_reward_corr": 0.7210829404756578,
+      "objective/train/value_std": 0.01317596435546875,
+      "objective/train/weight_avg": 1.001671314239502,
+      "objective/train/weighted_lm_loss": 1.2260372638702393,
+      "objective/train/weights_max": 1.1339726448059082,
+      "objective/train/weights_min": 0.3693629503250122,
+      "theoretical_loss": 3.5280338918975263,
+      "tokens_seen": 1582694400
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000525597817364789,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5280086023302264,
+      "tokens_seen": 1582825472
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005255175734232066,
+      "loss": 0.0687,
+      "theoretical_loss": 3.527958031236425,
+      "tokens_seen": 1583087616
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005254373294816242,
+      "loss": 0.0694,
+      "theoretical_loss": 3.527907470860315,
+      "tokens_seen": 1583349760
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005253570855400417,
+      "loss": 0.071,
+      "theoretical_loss": 3.527856921197852,
+      "tokens_seen": 1583611904
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005252768415984593,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5278063822449925,
+      "tokens_seen": 1583874048
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005251965976568769,
+      "loss": 0.07,
+      "theoretical_loss": 3.5277558539976956,
+      "tokens_seen": 1584136192
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005251163537152945,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5277053364519215,
+      "tokens_seen": 1584398336
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000525036109773712,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5276548296036356,
+      "tokens_seen": 1584660480
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005249558658321298,
+      "loss": 0.0685,
+      "theoretical_loss": 3.527604333448802,
+      "tokens_seen": 1584922624
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005248756218905473,
+      "loss": 0.0687,
+      "theoretical_loss": 3.527553847983389,
+      "tokens_seen": 1585184768
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005247953779489648,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5275033732033667,
+      "tokens_seen": 1585446912
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005247151340073825,
+      "loss": 0.0672,
+      "theoretical_loss": 3.5274529091047073,
+      "tokens_seen": 1585709056
+    },
+    {
+      "epoch": 0.48,
+      "objective/train/advantage_avg": 0.00040885916678234935,
+      "objective/train/docs_used": 579278,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2280428409576416,
+      "objective/train/original_loss": 1.2280428409576416,
+      "objective/train/theoretical_loss": 3.527402455683385,
+      "objective/train/tokens_used": 1606431200,
+      "objective/train/value_avg": -0.00798797607421875,
+      "objective/train/value_loss": 0.00015769110177643597,
+      "objective/train/value_max": -3.820657730102539e-05,
+      "objective/train/value_min": -0.3349609375,
+      "objective/train/value_reward_corr": 0.7601393100857727,
+      "objective/train/value_std": 0.01552581787109375,
+      "objective/train/weight_avg": 1.0004829168319702,
+      "objective/train/weighted_lm_loss": 1.2284908294677734,
+      "objective/train/weights_max": 1.20417058467865,
+      "objective/train/weights_min": 0.36865562200546265,
+      "theoretical_loss": 3.527402455683385,
+      "tokens_seen": 1585971200
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005246348900658,
+      "loss": 0.0678,
+      "theoretical_loss": 3.527402455683385,
+      "tokens_seen": 1585971200
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005245546461242177,
+      "loss": 0.0697,
+      "theoretical_loss": 3.527352012935377,
+      "tokens_seen": 1586233344
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005244744021826352,
+      "loss": 0.0695,
+      "theoretical_loss": 3.527301580856661,
+      "tokens_seen": 1586495488
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005243941582410528,
+      "loss": 0.0702,
+      "theoretical_loss": 3.527251159443219,
+      "tokens_seen": 1586757632
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005243139142994704,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5272007486910333,
+      "tokens_seen": 1587019776
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005242336703578879,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5271503485960896,
+      "tokens_seen": 1587281920
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005241534264163056,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5270999591543752,
+      "tokens_seen": 1587544064
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005240731824747232,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5270495803618793,
+      "tokens_seen": 1587806208
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005239929385331408,
+      "loss": 0.0656,
+      "theoretical_loss": 3.5269992122145952,
+      "tokens_seen": 1588068352
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005239126945915583,
+      "loss": 0.068,
+      "theoretical_loss": 3.526948854708515,
+      "tokens_seen": 1588330496
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000523832450649976,
+      "loss": 0.0688,
+      "theoretical_loss": 3.526898507839636,
+      "tokens_seen": 1588592640
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005237522067083935,
+      "loss": 0.0674,
+      "theoretical_loss": 3.526848171603956,
+      "tokens_seen": 1588854784
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005236719627668111,
+      "loss": 0.0705,
+      "theoretical_loss": 3.526797845997476,
+      "tokens_seen": 1589116928
+    },
+    {
+      "epoch": 0.48,
+      "objective/train/advantage_avg": -0.00024388304154854268,
+      "objective/train/docs_used": 580493,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3449385166168213,
+      "objective/train/original_loss": 1.3449382781982422,
+      "objective/train/theoretical_loss": 3.5267726871789367,
+      "objective/train/tokens_used": 1609708000,
+      "objective/train/value_avg": -0.00757598876953125,
+      "objective/train/value_loss": 0.0001988656003959477,
+      "objective/train/value_max": -6.252527236938477e-05,
+      "objective/train/value_min": -0.48291015625,
+      "objective/train/value_reward_corr": 0.8828137850946742,
+      "objective/train/value_std": 0.02032470703125,
+      "objective/train/weight_avg": 0.9998489618301392,
+      "objective/train/weighted_lm_loss": 1.3448597192764282,
+      "objective/train/weights_max": 1.1375212669372559,
+      "objective/train/weights_min": 0.38701966404914856,
+      "theoretical_loss": 3.5267726871789367,
+      "tokens_seen": 1589248000
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005235917188252287,
+      "loss": 0.0669,
+      "theoretical_loss": 3.5267475310161984,
+      "tokens_seen": 1589379072
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005235114748836462,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5266972266561276,
+      "tokens_seen": 1589641216
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000523431230942064,
+      "loss": 0.0673,
+      "theoretical_loss": 3.526646932913271,
+      "tokens_seen": 1589903360
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005233509870004815,
+      "loss": 0.067,
+      "theoretical_loss": 3.5265966497836376,
+      "tokens_seen": 1590165504
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005232707430588991,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5265463772632386,
+      "tokens_seen": 1590427648
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005231904991173167,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5264961153480874,
+      "tokens_seen": 1590689792
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005231102551757343,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5264458640342,
+      "tokens_seen": 1590951936
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005230300112341518,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5263956233175935,
+      "tokens_seen": 1591214080
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005229497672925694,
+      "loss": 0.067,
+      "theoretical_loss": 3.5263453931942883,
+      "tokens_seen": 1591476224
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000522869523350987,
+      "loss": 0.0673,
+      "theoretical_loss": 3.5262951736603063,
+      "tokens_seen": 1591738368
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005227892794094045,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5262449647116716,
+      "tokens_seen": 1592000512
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005227090354678223,
+      "loss": 0.0665,
+      "theoretical_loss": 3.5261947663444104,
+      "tokens_seen": 1592262656
+    },
+    {
+      "epoch": 0.48,
+      "objective/train/advantage_avg": 3.907005884684622e-05,
+      "objective/train/docs_used": 581545,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5011039972305298,
+      "objective/train/original_loss": 1.5011038780212402,
+      "objective/train/theoretical_loss": 3.5261445785545513,
+      "objective/train/tokens_used": 1612984800,
+      "objective/train/value_avg": -0.006710052490234375,
+      "objective/train/value_loss": 0.00017771663260646164,
+      "objective/train/value_max": -6.502866744995117e-05,
+      "objective/train/value_min": -0.21435546875,
+      "objective/train/value_reward_corr": 0.706507157381232,
+      "objective/train/value_std": 0.01238250732421875,
+      "objective/train/weight_avg": 1.0001215934753418,
+      "objective/train/weighted_lm_loss": 1.5013684034347534,
+      "objective/train/weights_max": 1.153806447982788,
+      "objective/train/weights_min": 0.36828702688217163,
+      "theoretical_loss": 3.5261445785545513,
+      "tokens_seen": 1592524800
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005226287915262398,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5261445785545513,
+      "tokens_seen": 1592524800
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005225485475846574,
+      "loss": 0.069,
+      "theoretical_loss": 3.526094401338125,
+      "tokens_seen": 1592786944
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000522468303643075,
+      "loss": 0.0696,
+      "theoretical_loss": 3.526044234691165,
+      "tokens_seen": 1593049088
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005223880597014925,
+      "loss": 0.0682,
+      "theoretical_loss": 3.525994078609705,
+      "tokens_seen": 1593311232
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005223078157599101,
+      "loss": 0.0696,
+      "theoretical_loss": 3.525943933089782,
+      "tokens_seen": 1593573376
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005222275718183277,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5258937981274365,
+      "tokens_seen": 1593835520
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005221473278767453,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5258436737187084,
+      "tokens_seen": 1594097664
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005220670839351629,
+      "loss": 0.0672,
+      "theoretical_loss": 3.5257935598596424,
+      "tokens_seen": 1594359808
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005219868399935806,
+      "loss": 0.0676,
+      "theoretical_loss": 3.5257434565462833,
+      "tokens_seen": 1594621952
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005219065960519981,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5256933637746792,
+      "tokens_seen": 1594884096
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005218263521104157,
+      "loss": 0.0652,
+      "theoretical_loss": 3.5256432815408796,
+      "tokens_seen": 1595146240
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005217461081688333,
+      "loss": 0.0661,
+      "theoretical_loss": 3.5255932098409364,
+      "tokens_seen": 1595408384
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005216658642272508,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5255431486709043,
+      "tokens_seen": 1595670528
+    },
+    {
+      "epoch": 0.48,
+      "objective/train/advantage_avg": 0.0021843239665031433,
+      "objective/train/docs_used": 582715,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3551676273345947,
+      "objective/train/original_loss": 1.3551673889160156,
+      "objective/train/theoretical_loss": 3.5255181220333722,
+      "objective/train/tokens_used": 1616261600,
+      "objective/train/value_avg": -0.011962890625,
+      "objective/train/value_loss": 0.0005959582631476223,
+      "objective/train/value_max": -6.556510925292969e-05,
+      "objective/train/value_min": -0.79541015625,
+      "objective/train/value_reward_corr": 0.6982835942676264,
+      "objective/train/value_std": 0.0255889892578125,
+      "objective/train/weight_avg": 1.0024514198303223,
+      "objective/train/weighted_lm_loss": 1.3577196598052979,
+      "objective/train/weights_max": 1.8855996131896973,
+      "objective/train/weights_min": 0.37003564834594727,
+      "theoretical_loss": 3.5255181220333722,
+      "tokens_seen": 1595801600
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005215856202856685,
+      "loss": 0.0651,
+      "theoretical_loss": 3.525493098026839,
+      "tokens_seen": 1595932672
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000521505376344086,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5254430579047993,
+      "tokens_seen": 1596194816
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005214251324025036,
+      "loss": 0.0659,
+      "theoretical_loss": 3.5253930283008454,
+      "tokens_seen": 1596456960
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005213448884609212,
+      "loss": 0.069,
+      "theoretical_loss": 3.5253430092110403,
+      "tokens_seen": 1596719104
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005212646445193387,
+      "loss": 0.067,
+      "theoretical_loss": 3.5252930006314482,
+      "tokens_seen": 1596981248
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005211844005777564,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5252430025581356,
+      "tokens_seen": 1597243392
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000521104156636174,
+      "loss": 0.0678,
+      "theoretical_loss": 3.5251930149871726,
+      "tokens_seen": 1597505536
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005210239126945916,
+      "loss": 0.0675,
+      "theoretical_loss": 3.5251430379146296,
+      "tokens_seen": 1597767680
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005209436687530092,
+      "loss": 0.0674,
+      "theoretical_loss": 3.5250930713365802,
+      "tokens_seen": 1598029824
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005208634248114268,
+      "loss": 0.0655,
+      "theoretical_loss": 3.525043115249099,
+      "tokens_seen": 1598291968
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005207831808698443,
+      "loss": 0.07,
+      "theoretical_loss": 3.5249931696482637,
+      "tokens_seen": 1598554112
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005207029369282619,
+      "loss": 0.0699,
+      "theoretical_loss": 3.524943234530154,
+      "tokens_seen": 1598816256
+    },
+    {
+      "epoch": 0.48,
+      "objective/train/advantage_avg": 0.00034000459709204733,
+      "objective/train/docs_used": 583998,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4760903120040894,
+      "objective/train/original_loss": 1.4760903120040894,
+      "objective/train/theoretical_loss": 3.5248933098908517,
+      "objective/train/tokens_used": 1619538400,
+      "objective/train/value_avg": -0.006893157958984375,
+      "objective/train/value_loss": 0.00016073322331067175,
+      "objective/train/value_max": -5.739927291870117e-05,
+      "objective/train/value_min": -0.345703125,
+      "objective/train/value_reward_corr": 0.6807590254059671,
+      "objective/train/value_std": 0.01171112060546875,
+      "objective/train/weight_avg": 1.0004159212112427,
+      "objective/train/weighted_lm_loss": 1.4770082235336304,
+      "objective/train/weights_max": 1.1568527221679688,
+      "objective/train/weights_min": 0.40742918848991394,
+      "theoretical_loss": 3.5248933098908517,
+      "tokens_seen": 1599078400
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005206226929866795,
+      "loss": 0.0688,
+      "theoretical_loss": 3.5248933098908517,
+      "tokens_seen": 1599078400
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.000520542449045097,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5248433957264402,
+      "tokens_seen": 1599340544
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005204622051035148,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5247934920330053,
+      "tokens_seen": 1599602688
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005203819611619323,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5247435988066353,
+      "tokens_seen": 1599864832
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005203017172203499,
+      "loss": 0.0694,
+      "theoretical_loss": 3.52469371604342,
+      "tokens_seen": 1600126976
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 0.0005202214732787675,
+      "loss": 0.0692,
+      "theoretical_loss": 3.524643843739452,
+      "tokens_seen": 1600389120
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005201412293371851,
+      "loss": 0.0727,
+      "theoretical_loss": 3.524593981890825,
+      "tokens_seen": 1600651264
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005200609853956026,
+      "loss": 0.069,
+      "theoretical_loss": 3.524544130493635,
+      "tokens_seen": 1600913408
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005199807414540202,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5244942895439815,
+      "tokens_seen": 1601175552
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005199004975124378,
+      "loss": 0.069,
+      "theoretical_loss": 3.524444459037965,
+      "tokens_seen": 1601437696
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005198202535708553,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5243946389716867,
+      "tokens_seen": 1601699840
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.000519740009629273,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5243448293412527,
+      "tokens_seen": 1601961984
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005196597656876906,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5242950301427696,
+      "tokens_seen": 1602224128
+    },
+    {
+      "epoch": 0.49,
+      "objective/train/advantage_avg": 0.0010038897162303329,
+      "objective/train/docs_used": 585126,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3920528888702393,
+      "objective/train/original_loss": 1.3920525312423706,
+      "objective/train/theoretical_loss": 3.524270134454293,
+      "objective/train/tokens_used": 1622815200,
+      "objective/train/value_avg": -0.006145477294921875,
+      "objective/train/value_loss": 0.00022318839910440147,
+      "objective/train/value_max": -3.737211227416992e-05,
+      "objective/train/value_min": -0.71484375,
+      "objective/train/value_reward_corr": 0.6453517865349159,
+      "objective/train/value_std": 0.01385498046875,
+      "objective/train/weight_avg": 1.0011008977890015,
+      "objective/train/weighted_lm_loss": 1.3941224813461304,
+      "objective/train/weights_max": 1.4599730968475342,
+      "objective/train/weights_min": 0.3827209174633026,
+      "theoretical_loss": 3.524270134454293,
+      "tokens_seen": 1602355200
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005195795217461083,
+      "loss": 0.0675,
+      "theoretical_loss": 3.5242452413723457,
+      "tokens_seen": 1602486272
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005194992778045258,
+      "loss": 0.069,
+      "theoretical_loss": 3.5241954630260923,
+      "tokens_seen": 1602748416
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005194190338629433,
+      "loss": 0.0655,
+      "theoretical_loss": 3.524145695100123,
+      "tokens_seen": 1603010560
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.000519338789921361,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5240959375905527,
+      "tokens_seen": 1603272704
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005192585459797785,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5240461904934977,
+      "tokens_seen": 1603534848
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005191783020381961,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5239964538050788,
+      "tokens_seen": 1603796992
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005190980580966137,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5239467275214165,
+      "tokens_seen": 1604059136
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005190178141550314,
+      "loss": 0.0707,
+      "theoretical_loss": 3.523897011638635,
+      "tokens_seen": 1604321280
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005189375702134489,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5238473061528586,
+      "tokens_seen": 1604583424
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005188573262718665,
+      "loss": 0.0659,
+      "theoretical_loss": 3.5237976110602163,
+      "tokens_seen": 1604845568
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005187770823302841,
+      "loss": 0.0715,
+      "theoretical_loss": 3.523747926356837,
+      "tokens_seen": 1605107712
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005186968383887016,
+      "loss": 0.0686,
+      "theoretical_loss": 3.523698252038853,
+      "tokens_seen": 1605369856
+    },
+    {
+      "epoch": 0.49,
+      "objective/train/advantage_avg": 0.00131322315428406,
+      "objective/train/docs_used": 586391,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3650020360946655,
+      "objective/train/original_loss": 1.365002155303955,
+      "objective/train/theoretical_loss": 3.5236485881023976,
+      "objective/train/tokens_used": 1626092000,
+      "objective/train/value_avg": -0.00806427001953125,
+      "objective/train/value_loss": 0.0001866552047431469,
+      "objective/train/value_max": -8.612871170043945e-05,
+      "objective/train/value_min": -0.99072265625,
+      "objective/train/value_reward_corr": 0.7294081039501898,
+      "objective/train/value_std": 0.01560211181640625,
+      "objective/train/weight_avg": 1.001397728919983,
+      "objective/train/weighted_lm_loss": 1.3666753768920898,
+      "objective/train/weights_max": 1.173264503479004,
+      "objective/train/weights_min": 0.36904817819595337,
+      "theoretical_loss": 3.5236485881023976,
+      "tokens_seen": 1605632000
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005186165944471193,
+      "loss": 0.0676,
+      "theoretical_loss": 3.5236485881023976,
+      "tokens_seen": 1605632000
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005185363505055368,
+      "loss": 0.0671,
+      "theoretical_loss": 3.5235989345436076,
+      "tokens_seen": 1605894144
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005184561065639545,
+      "loss": 0.0693,
+      "theoretical_loss": 3.52354929135862,
+      "tokens_seen": 1606156288
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.000518375862622372,
+      "loss": 0.0665,
+      "theoretical_loss": 3.5234996585435754,
+      "tokens_seen": 1606418432
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005182956186807895,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5234500360946157,
+      "tokens_seen": 1606680576
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005182153747392073,
+      "loss": 0.0663,
+      "theoretical_loss": 3.5234004240078853,
+      "tokens_seen": 1606942720
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005181351307976248,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5233508222795304,
+      "tokens_seen": 1607204864
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005180548868560424,
+      "loss": 0.0673,
+      "theoretical_loss": 3.5233012309057,
+      "tokens_seen": 1607467008
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.00051797464291446,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5232516498825426,
+      "tokens_seen": 1607729152
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005178943989728776,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5232020792062126,
+      "tokens_seen": 1607991296
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005178141550312951,
+      "loss": 0.067,
+      "theoretical_loss": 3.523152518872864,
+      "tokens_seen": 1608253440
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005177339110897127,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5231029688786526,
+      "tokens_seen": 1608515584
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005176536671481303,
+      "loss": 0.0692,
+      "theoretical_loss": 3.523053429219738,
+      "tokens_seen": 1608777728
+    },
+    {
+      "epoch": 0.49,
+      "objective/train/advantage_avg": -0.00025140075013041496,
+      "objective/train/docs_used": 587617,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.6009109020233154,
+      "objective/train/original_loss": 1.6009111404418945,
+      "objective/train/theoretical_loss": 3.5230286632648165,
+      "objective/train/tokens_used": 1629368800,
+      "objective/train/value_avg": -0.00858306884765625,
+      "objective/train/value_loss": 0.0004015047161374241,
+      "objective/train/value_max": -7.033348083496094e-05,
+      "objective/train/value_min": -0.63232421875,
+      "objective/train/value_reward_corr": 0.6128993489929545,
+      "objective/train/value_std": 0.01490020751953125,
+      "objective/train/weight_avg": 0.9999117851257324,
+      "objective/train/weighted_lm_loss": 1.6008124351501465,
+      "objective/train/weights_max": 1.3433096408843994,
+      "objective/train/weights_min": 0.2304520606994629,
+      "theoretical_loss": 3.5230286632648165,
+      "tokens_seen": 1608908800
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005175734232065478,
+      "loss": 0.0742,
+      "theoretical_loss": 3.52300389989228,
+      "tokens_seen": 1609039872
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005174931792649656,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5229543808924415,
+      "tokens_seen": 1609302016
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005174129353233831,
+      "loss": 0.0684,
+      "theoretical_loss": 3.522904872216388,
+      "tokens_seen": 1609564160
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005173326913818007,
+      "loss": 0.069,
+      "theoretical_loss": 3.5228553738602857,
+      "tokens_seen": 1609826304
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005172524474402183,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5228058858203033,
+      "tokens_seen": 1610088448
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005171722034986359,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5227564080926115,
+      "tokens_seen": 1610350592
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005170919595570535,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5227069406733844,
+      "tokens_seen": 1610612736
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.000517011715615471,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5226574835587963,
+      "tokens_seen": 1610874880
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005169314716738886,
+      "loss": 0.0676,
+      "theoretical_loss": 3.522608036745024,
+      "tokens_seen": 1611137024
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005168512277323062,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5225586002282467,
+      "tokens_seen": 1611399168
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005167709837907239,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5225091740046457,
+      "tokens_seen": 1611661312
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005166907398491414,
+      "loss": 0.072,
+      "theoretical_loss": 3.5224597580704033,
+      "tokens_seen": 1611923456
+    },
+    {
+      "epoch": 0.49,
+      "objective/train/advantage_avg": -0.00028426916105672717,
+      "objective/train/docs_used": 588885,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4351452589035034,
+      "objective/train/original_loss": 1.4351451396942139,
+      "objective/train/theoretical_loss": 3.5224103524217067,
+      "objective/train/tokens_used": 1632645600,
+      "objective/train/value_avg": -0.0109710693359375,
+      "objective/train/value_loss": 0.0004222493735142052,
+      "objective/train/value_max": -5.7816505432128906e-05,
+      "objective/train/value_min": -0.90771484375,
+      "objective/train/value_reward_corr": 0.8208013011530193,
+      "objective/train/value_std": 0.0279388427734375,
+      "objective/train/weight_avg": 0.9999049305915833,
+      "objective/train/weighted_lm_loss": 1.4349756240844727,
+      "objective/train/weights_max": 1.4129830598831177,
+      "objective/train/weights_min": 0.37088075280189514,
+      "theoretical_loss": 3.5224103524217067,
+      "tokens_seen": 1612185600
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005166104959075591,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5224103524217067,
+      "tokens_seen": 1612185600
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005165302519659766,
+      "loss": 0.0706,
+      "theoretical_loss": 3.522360957054741,
+      "tokens_seen": 1612447744
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005164500080243941,
+      "loss": 0.0668,
+      "theoretical_loss": 3.5223115719656963,
+      "tokens_seen": 1612709888
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005163697640828118,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5222621971507646,
+      "tokens_seen": 1612972032
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005162895201412293,
+      "loss": 0.0696,
+      "theoretical_loss": 3.522212832606138,
+      "tokens_seen": 1613234176
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005162092761996469,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5221634783280122,
+      "tokens_seen": 1613496320
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005161290322580645,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5221141343125852,
+      "tokens_seen": 1613758464
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005160487883164821,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5220648005560555,
+      "tokens_seen": 1614020608
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005159685443748997,
+      "loss": 0.075,
+      "theoretical_loss": 3.522015477054625,
+      "tokens_seen": 1614282752
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005158883004333173,
+      "loss": 0.0729,
+      "theoretical_loss": 3.521966163804497,
+      "tokens_seen": 1614544896
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005158080564917349,
+      "loss": 0.0707,
+      "theoretical_loss": 3.521916860801877,
+      "tokens_seen": 1614807040
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005157278125501525,
+      "loss": 0.0705,
+      "theoretical_loss": 3.521867568042973,
+      "tokens_seen": 1615069184
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005156475686085701,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5218182855239935,
+      "tokens_seen": 1615331328
+    },
+    {
+      "epoch": 0.49,
+      "objective/train/advantage_avg": 0.001808401197195053,
+      "objective/train/docs_used": 590099,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3354768753051758,
+      "objective/train/original_loss": 1.3354768753051758,
+      "objective/train/theoretical_loss": 3.5217936481032917,
+      "objective/train/tokens_used": 1635922400,
+      "objective/train/value_avg": -0.01221466064453125,
+      "objective/train/value_loss": 0.0007001345511525869,
+      "objective/train/value_max": -5.346536636352539e-05,
+      "objective/train/value_min": -0.861328125,
+      "objective/train/value_reward_corr": 0.7123812003133756,
+      "objective/train/value_std": 0.0258026123046875,
+      "objective/train/weight_avg": 1.002112865447998,
+      "objective/train/weighted_lm_loss": 1.3375283479690552,
+      "objective/train/weights_max": 1.4017969369888306,
+      "objective/train/weights_min": 0.3730887472629547,
+      "theoretical_loss": 3.5217936481032917,
+      "tokens_seen": 1615462400
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005155673246669876,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5217690132411508,
+      "tokens_seen": 1615593472
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005154870807254053,
+      "loss": 0.0709,
+      "theoretical_loss": 3.521719751190658,
+      "tokens_seen": 1615855616
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005154068367838228,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5216704993687307,
+      "tokens_seen": 1616117760
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005153265928422403,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5216212577715873,
+      "tokens_seen": 1616379904
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005152463489006581,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5215720263954458,
+      "tokens_seen": 1616642048
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005151661049590756,
+      "loss": 0.0693,
+      "theoretical_loss": 3.521522805236529,
+      "tokens_seen": 1616904192
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005150858610174932,
+      "loss": 0.0725,
+      "theoretical_loss": 3.52147359429106,
+      "tokens_seen": 1617166336
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005150056170759108,
+      "loss": 0.071,
+      "theoretical_loss": 3.5214243935552654,
+      "tokens_seen": 1617428480
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005149253731343284,
+      "loss": 0.0663,
+      "theoretical_loss": 3.5213752030253715,
+      "tokens_seen": 1617690624
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005148451291927459,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5213260226976084,
+      "tokens_seen": 1617952768
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005147648852511635,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5212768525682074,
+      "tokens_seen": 1618214912
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005146846413095811,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5212276926334027,
+      "tokens_seen": 1618477056
+    },
+    {
+      "epoch": 0.49,
+      "objective/train/advantage_avg": 0.0004180370888207108,
+      "objective/train/docs_used": 591377,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.235845923423767,
+      "objective/train/original_loss": 1.235845923423767,
+      "objective/train/theoretical_loss": 3.5211785428894293,
+      "objective/train/tokens_used": 1639199200,
+      "objective/train/value_avg": -0.0059814453125,
+      "objective/train/value_loss": 0.00012016700202366337,
+      "objective/train/value_max": -7.426738739013672e-05,
+      "objective/train/value_min": -0.394775390625,
+      "objective/train/value_reward_corr": 0.6571299259985038,
+      "objective/train/value_std": 0.00959014892578125,
+      "objective/train/weight_avg": 1.000473976135254,
+      "objective/train/weighted_lm_loss": 1.2364455461502075,
+      "objective/train/weights_max": 1.16512930393219,
+      "objective/train/weights_min": 0.39590752124786377,
+      "theoretical_loss": 3.5211785428894293,
+      "tokens_seen": 1618739200
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005146043973679986,
+      "loss": 0.0673,
+      "theoretical_loss": 3.5211785428894293,
+      "tokens_seen": 1618739200
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005145241534264164,
+      "loss": 0.072,
+      "theoretical_loss": 3.5211294033325258,
+      "tokens_seen": 1619001344
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005144439094848339,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5210802739589306,
+      "tokens_seen": 1619263488
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005143636655432516,
+      "loss": 0.0696,
+      "theoretical_loss": 3.521031154764886,
+      "tokens_seen": 1619525632
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005142834216016691,
+      "loss": 0.0708,
+      "theoretical_loss": 3.520982045746636,
+      "tokens_seen": 1619787776
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005142031776600867,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5209329469004254,
+      "tokens_seen": 1620049920
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005141229337185043,
+      "loss": 0.0715,
+      "theoretical_loss": 3.5208838582225024,
+      "tokens_seen": 1620312064
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005140426897769218,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5208347797091157,
+      "tokens_seen": 1620574208
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005139624458353394,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5207857113565177,
+      "tokens_seen": 1620836352
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.000513882201893757,
+      "loss": 0.0693,
+      "theoretical_loss": 3.520736653160962,
+      "tokens_seen": 1621098496
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005138019579521747,
+      "loss": 0.0694,
+      "theoretical_loss": 3.520687605118704,
+      "tokens_seen": 1621360640
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005137217140105922,
+      "loss": 0.0722,
+      "theoretical_loss": 3.520638567226001,
+      "tokens_seen": 1621622784
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005136414700690099,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5205895394791127,
+      "tokens_seen": 1621884928
+    },
+    {
+      "epoch": 0.49,
+      "objective/train/advantage_avg": 0.00047303320025093853,
+      "objective/train/docs_used": 592532,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3057827949523926,
+      "objective/train/original_loss": 1.3057827949523926,
+      "objective/train/theoretical_loss": 3.5205650294091804,
+      "objective/train/tokens_used": 1642476000,
+      "objective/train/value_avg": -0.0077362060546875,
+      "objective/train/value_loss": 0.0004286039329599589,
+      "objective/train/value_max": -2.9981136322021484e-05,
+      "objective/train/value_min": -0.9453125,
+      "objective/train/value_reward_corr": 0.7418295034563557,
+      "objective/train/value_std": 0.023773193359375,
+      "objective/train/weight_avg": 1.0006773471832275,
+      "objective/train/weighted_lm_loss": 1.3059204816818237,
+      "objective/train/weights_max": 2.067409038543701,
+      "objective/train/weights_min": 0.39820975065231323,
+      "theoretical_loss": 3.5205650294091804,
+      "tokens_seen": 1622016000
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005135612261274274,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5205405218743007,
+      "tokens_seen": 1622147072
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005134809821858449,
+      "loss": 0.0701,
+      "theoretical_loss": 3.520491514407828,
+      "tokens_seen": 1622409216
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005134007382442626,
+      "loss": 0.07,
+      "theoretical_loss": 3.520442517075961,
+      "tokens_seen": 1622671360
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005133204943026801,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5203935298749656,
+      "tokens_seen": 1622933504
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005132402503610978,
+      "loss": 0.0688,
+      "theoretical_loss": 3.520344552801113,
+      "tokens_seen": 1623195648
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005131600064195153,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5202955858506737,
+      "tokens_seen": 1623457792
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.000513079762477933,
+      "loss": 0.0725,
+      "theoretical_loss": 3.520246629019921,
+      "tokens_seen": 1623719936
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005129995185363506,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5201976823051306,
+      "tokens_seen": 1623982080
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005129192745947681,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5201487457025795,
+      "tokens_seen": 1624244224
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005128390306531857,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5200998192085473,
+      "tokens_seen": 1624506368
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005127587867116033,
+      "loss": 0.069,
+      "theoretical_loss": 3.5200509028193148,
+      "tokens_seen": 1624768512
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005126785427700209,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5200019965311657,
+      "tokens_seen": 1625030656
+    },
+    {
+      "epoch": 0.49,
+      "objective/train/advantage_avg": 0.0004936844925396144,
+      "objective/train/docs_used": 593743,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.600699782371521,
+      "objective/train/original_loss": 1.6006999015808105,
+      "objective/train/theoretical_loss": 3.519953100340385,
+      "objective/train/tokens_used": 1645752800,
+      "objective/train/value_avg": -0.007678985595703125,
+      "objective/train/value_loss": 0.00025552220176905394,
+      "objective/train/value_max": -6.657838821411133e-05,
+      "objective/train/value_min": -0.8681640625,
+      "objective/train/value_reward_corr": 0.7401646371099073,
+      "objective/train/value_std": 0.01708984375,
+      "objective/train/weight_avg": 1.0006078481674194,
+      "objective/train/weighted_lm_loss": 1.602151870727539,
+      "objective/train/weights_max": 1.651138186454773,
+      "objective/train/weights_min": 0.36892783641815186,
+      "theoretical_loss": 3.519953100340385,
+      "tokens_seen": 1625292800
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005125982988284384,
+      "loss": 0.0712,
+      "theoretical_loss": 3.519953100340385,
+      "tokens_seen": 1625292800
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005125180548868561,
+      "loss": 0.0729,
+      "theoretical_loss": 3.51990421424326,
+      "tokens_seen": 1625554944
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005124378109452736,
+      "loss": 0.0703,
+      "theoretical_loss": 3.519855338236079,
+      "tokens_seen": 1625817088
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005123575670036911,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5198064723151345,
+      "tokens_seen": 1626079232
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005122773230621089,
+      "loss": 0.0707,
+      "theoretical_loss": 3.519757616476719,
+      "tokens_seen": 1626341376
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005121970791205264,
+      "loss": 0.0692,
+      "theoretical_loss": 3.519708770717126,
+      "tokens_seen": 1626603520
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.000512116835178944,
+      "loss": 0.0718,
+      "theoretical_loss": 3.519659935032655,
+      "tokens_seen": 1626865664
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005120365912373616,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5196111094196034,
+      "tokens_seen": 1627127808
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005119563472957792,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5195622938742726,
+      "tokens_seen": 1627389952
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005118761033541968,
+      "loss": 0.0712,
+      "theoretical_loss": 3.519513488392965,
+      "tokens_seen": 1627652096
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005117958594126143,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5194646929719853,
+      "tokens_seen": 1627914240
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005117156154710319,
+      "loss": 0.0706,
+      "theoretical_loss": 3.519415907607641,
+      "tokens_seen": 1628176384
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005116353715294495,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5193671322962397,
+      "tokens_seen": 1628438528
+    },
+    {
+      "epoch": 0.49,
+      "objective/train/advantage_avg": 0.00031531660351902246,
+      "objective/train/docs_used": 595003,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.356420874595642,
+      "objective/train/original_loss": 1.3564209938049316,
+      "objective/train/theoretical_loss": 3.5193427484092403,
+      "objective/train/tokens_used": 1649029600,
+      "objective/train/value_avg": -0.00811004638671875,
+      "objective/train/value_loss": 0.0006286876741796732,
+      "objective/train/value_max": -0.00010639429092407227,
+      "objective/train/value_min": -0.9814453125,
+      "objective/train/value_reward_corr": 0.7337605809007626,
+      "objective/train/value_std": 0.029541015625,
+      "objective/train/weight_avg": 1.0006129741668701,
+      "objective/train/weighted_lm_loss": 1.3556339740753174,
+      "objective/train/weights_max": 2.1884546279907227,
+      "objective/train/weights_min": 0.3697238564491272,
+      "theoretical_loss": 3.5193427484092403,
+      "tokens_seen": 1628569600
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005115551275878672,
+      "loss": 0.0703,
+      "theoretical_loss": 3.519318367034093,
+      "tokens_seen": 1628700672
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005114748836462847,
+      "loss": 0.0692,
+      "theoretical_loss": 3.519269611817513,
+      "tokens_seen": 1628962816
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005113946397047024,
+      "loss": 0.0712,
+      "theoretical_loss": 3.5192208666428146,
+      "tokens_seen": 1629224960
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005113143957631199,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5191721315063136,
+      "tokens_seen": 1629487104
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005112341518215374,
+      "loss": 0.0683,
+      "theoretical_loss": 3.5191234064043293,
+      "tokens_seen": 1629749248
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005111539078799551,
+      "loss": 0.0678,
+      "theoretical_loss": 3.5190746913331816,
+      "tokens_seen": 1630011392
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005110736639383726,
+      "loss": 0.0706,
+      "theoretical_loss": 3.5190259862891926,
+      "tokens_seen": 1630273536
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005109934199967902,
+      "loss": 0.0684,
+      "theoretical_loss": 3.518977291268686,
+      "tokens_seen": 1630535680
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005109131760552078,
+      "loss": 0.071,
+      "theoretical_loss": 3.5189286062679894,
+      "tokens_seen": 1630797824
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005108329321136255,
+      "loss": 0.0697,
+      "theoretical_loss": 3.51887993128343,
+      "tokens_seen": 1631059968
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005107526881720431,
+      "loss": 0.0661,
+      "theoretical_loss": 3.518831266311339,
+      "tokens_seen": 1631322112
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005106724442304607,
+      "loss": 0.0704,
+      "theoretical_loss": 3.518782611348046,
+      "tokens_seen": 1631584256
+    },
+    {
+      "epoch": 0.49,
+      "objective/train/advantage_avg": 0.0003234239702578634,
+      "objective/train/docs_used": 596193,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4978060722351074,
+      "objective/train/original_loss": 1.4978063106536865,
+      "objective/train/theoretical_loss": 3.5187339663898873,
+      "objective/train/tokens_used": 1652306400,
+      "objective/train/value_avg": -0.005161285400390625,
+      "objective/train/value_loss": 9.56832809606567e-05,
+      "objective/train/value_max": -6.973743438720703e-05,
+      "objective/train/value_min": -0.171630859375,
+      "objective/train/value_reward_corr": 0.634942167538914,
+      "objective/train/value_std": 0.0083770751953125,
+      "objective/train/weight_avg": 1.0003665685653687,
+      "objective/train/weighted_lm_loss": 1.4985394477844238,
+      "objective/train/weights_max": 1.1682624816894531,
+      "objective/train/weights_min": 0.36887577176094055,
+      "theoretical_loss": 3.5187339663898873,
+      "tokens_seen": 1631846400
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005105922002888782,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5187339663898873,
+      "tokens_seen": 1631846400
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005105119563472958,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5186853314331974,
+      "tokens_seen": 1632108544
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005104317124057134,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5186367064743145,
+      "tokens_seen": 1632370688
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005103514684641309,
+      "loss": 0.073,
+      "theoretical_loss": 3.518588091509578,
+      "tokens_seen": 1632632832
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005102712245225486,
+      "loss": 0.0695,
+      "theoretical_loss": 3.51853948653533,
+      "tokens_seen": 1632894976
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005101909805809661,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5184908915479145,
+      "tokens_seen": 1633157120
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 0.0005101107366393837,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5184423065436756,
+      "tokens_seen": 1633419264
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005100304926978014,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5183937315189615,
+      "tokens_seen": 1633681408
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005099502487562189,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5183451664701217,
+      "tokens_seen": 1633943552
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005098700048146365,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5182966113935072,
+      "tokens_seen": 1634205696
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005097897608730541,
+      "loss": 0.0696,
+      "theoretical_loss": 3.518248066285471,
+      "tokens_seen": 1634467840
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005097095169314717,
+      "loss": 0.0676,
+      "theoretical_loss": 3.518199531142369,
+      "tokens_seen": 1634729984
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005096292729898892,
+      "loss": 0.0705,
+      "theoretical_loss": 3.518151005960557,
+      "tokens_seen": 1634992128
+    },
+    {
+      "epoch": 0.5,
+      "objective/train/advantage_avg": 0.0016112735029309988,
+      "objective/train/docs_used": 597373,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4085406064987183,
+      "objective/train/original_loss": 1.4085404872894287,
+      "objective/train/theoretical_loss": 3.518126747103997,
+      "objective/train/tokens_used": 1655583200,
+      "objective/train/value_avg": -0.00795745849609375,
+      "objective/train/value_loss": 0.0002556704275775701,
+      "objective/train/value_max": -5.692243576049805e-05,
+      "objective/train/value_min": -0.71630859375,
+      "objective/train/value_reward_corr": 0.7167812131379807,
+      "objective/train/value_std": 0.0161590576171875,
+      "objective/train/weight_avg": 1.0017259120941162,
+      "objective/train/weighted_lm_loss": 1.4109013080596924,
+      "objective/train/weights_max": 1.4989804029464722,
+      "objective/train/weights_min": 0.3708694577217102,
+      "theoretical_loss": 3.518126747103997,
+      "tokens_seen": 1635123200
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005095490290483069,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5181024907363945,
+      "tokens_seen": 1635254272
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005094687851067244,
+      "loss": 0.0692,
+      "theoretical_loss": 3.518053985466243,
+      "tokens_seen": 1635516416
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000509388541165142,
+      "loss": 0.0661,
+      "theoretical_loss": 3.518005490146464,
+      "tokens_seen": 1635778560
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005093082972235597,
+      "loss": 0.0658,
+      "theoretical_loss": 3.5179570047734225,
+      "tokens_seen": 1636040704
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005092280532819772,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5179085293434857,
+      "tokens_seen": 1636302848
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005091478093403949,
+      "loss": 0.0652,
+      "theoretical_loss": 3.517860063853022,
+      "tokens_seen": 1636564992
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005090675653988124,
+      "loss": 0.0695,
+      "theoretical_loss": 3.517811608298401,
+      "tokens_seen": 1636827136
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.00050898732145723,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5177631626759958,
+      "tokens_seen": 1637089280
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005089070775156476,
+      "loss": 0.0664,
+      "theoretical_loss": 3.5177147269821805,
+      "tokens_seen": 1637351424
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005088268335740651,
+      "loss": 0.0676,
+      "theoretical_loss": 3.517666301213331,
+      "tokens_seen": 1637613568
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005087465896324827,
+      "loss": 0.069,
+      "theoretical_loss": 3.5176178853658246,
+      "tokens_seen": 1637875712
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005086663456909003,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5175694794360424,
+      "tokens_seen": 1638137856
+    },
+    {
+      "debugging/Compilability": 1.0,
+      "debugging/distinct-1-grams": 0.7474626180525646,
+      "debugging/entropy-1-grams": 5.599013278347353,
+      "debugging/length": 499.42857142857144,
+      "debugging/num_segments": 14,
+      "debugging/raw_token_scores_avg": 0.007199673913419247,
+      "debugging/raw_token_scores_std": 0.028890423476696014,
+      "debugging/score": 0.0110880360129877,
+      "debugging/score_std": 0.01573414919444883,
+      "epoch": 0.5,
+      "objective/train/advantage_avg": 0.0007880714256316423,
+      "objective/train/docs_used": 598618,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3821892738342285,
+      "objective/train/original_loss": 1.3821892738342285,
+      "objective/train/theoretical_loss": 3.5175210834203656,
+      "objective/train/tokens_used": 1658860000,
+      "objective/train/value_avg": -0.00799560546875,
+      "objective/train/value_loss": 0.00038576649967581034,
+      "objective/train/value_max": -6.502866744995117e-05,
+      "objective/train/value_min": -0.85009765625,
+      "objective/train/value_reward_corr": 0.7350670241397752,
+      "objective/train/value_std": 0.0201873779296875,
+      "objective/train/weight_avg": 1.0009615421295166,
+      "objective/train/weighted_lm_loss": 1.383286952972412,
+      "objective/train/weights_max": 1.962690830230713,
+      "objective/train/weights_min": 0.3717447817325592,
+      "theoretical_loss": 3.5175210834203656,
+      "tokens_seen": 1638400000
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000508586101749318,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5175210834203656,
+      "tokens_seen": 1638400000
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005085058578077355,
+      "loss": 0.0736,
+      "theoretical_loss": 3.5174726973151778,
+      "tokens_seen": 1638662144
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005084256138661532,
+      "loss": 0.068,
+      "theoretical_loss": 3.517424321116865,
+      "tokens_seen": 1638924288
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005083453699245707,
+      "loss": 0.0709,
+      "theoretical_loss": 3.517375954821815,
+      "tokens_seen": 1639186432
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005082651259829882,
+      "loss": 0.0691,
+      "theoretical_loss": 3.517327598426416,
+      "tokens_seen": 1639448576
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005081848820414059,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5172792519270604,
+      "tokens_seen": 1639710720
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005081046380998234,
+      "loss": 0.0715,
+      "theoretical_loss": 3.517230915320141,
+      "tokens_seen": 1639972864
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005080243941582411,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5171825886020525,
+      "tokens_seen": 1640235008
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005079441502166586,
+      "loss": 0.0673,
+      "theoretical_loss": 3.5171342717691925,
+      "tokens_seen": 1640497152
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005078639062750763,
+      "loss": 0.0675,
+      "theoretical_loss": 3.517085964817959,
+      "tokens_seen": 1640759296
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005077836623334939,
+      "loss": 0.0694,
+      "theoretical_loss": 3.517037667744754,
+      "tokens_seen": 1641021440
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005077034183919115,
+      "loss": 0.0688,
+      "theoretical_loss": 3.516989380545979,
+      "tokens_seen": 1641283584
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000507623174450329,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5169411032180387,
+      "tokens_seen": 1641545728
+    },
+    {
+      "epoch": 0.5,
+      "objective/train/advantage_avg": -0.000734575791284442,
+      "objective/train/docs_used": 599774,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3677737712860107,
+      "objective/train/original_loss": 1.3677736520767212,
+      "objective/train/theoretical_loss": 3.5169169682545087,
+      "objective/train/tokens_used": 1662136800,
+      "objective/train/value_avg": -0.0098114013671875,
+      "objective/train/value_loss": 0.0007994266925379634,
+      "objective/train/value_max": -0.00010228157043457031,
+      "objective/train/value_min": -0.90771484375,
+      "objective/train/value_reward_corr": 0.8485356054895296,
+      "objective/train/value_std": 0.0374755859375,
+      "objective/train/weight_avg": 0.9996302127838135,
+      "objective/train/weighted_lm_loss": 1.3674414157867432,
+      "objective/train/weights_max": 1.9520905017852783,
+      "objective/train/weights_min": 0.3693559169769287,
+      "theoretical_loss": 3.5169169682545087,
+      "tokens_seen": 1641676800
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005075429305087466,
+      "loss": 0.0713,
+      "theoretical_loss": 3.51689283575734,
+      "tokens_seen": 1641807872
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005074626865671642,
+      "loss": 0.0723,
+      "theoretical_loss": 3.516844578160291,
+      "tokens_seen": 1642070016
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005073824426255817,
+      "loss": 0.0683,
+      "theoretical_loss": 3.5167963304233014,
+      "tokens_seen": 1642332160
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005073021986839994,
+      "loss": 0.0701,
+      "theoretical_loss": 3.516748092542784,
+      "tokens_seen": 1642594304
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005072219547424169,
+      "loss": 0.0718,
+      "theoretical_loss": 3.5166998645151515,
+      "tokens_seen": 1642856448
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005071417108008345,
+      "loss": 0.0722,
+      "theoretical_loss": 3.516651646336821,
+      "tokens_seen": 1643118592
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005070614668592522,
+      "loss": 0.0703,
+      "theoretical_loss": 3.5166034380042093,
+      "tokens_seen": 1643380736
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005069812229176697,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5165552395137363,
+      "tokens_seen": 1643642880
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005069009789760874,
+      "loss": 0.068,
+      "theoretical_loss": 3.516507050861823,
+      "tokens_seen": 1643905024
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005068207350345049,
+      "loss": 0.0671,
+      "theoretical_loss": 3.5164588720448937,
+      "tokens_seen": 1644167168
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005067404910929225,
+      "loss": 0.07,
+      "theoretical_loss": 3.5164107030593725,
+      "tokens_seen": 1644429312
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005066602471513401,
+      "loss": 0.0706,
+      "theoretical_loss": 3.516362543901687,
+      "tokens_seen": 1644691456
+    },
+    {
+      "epoch": 0.5,
+      "objective/train/advantage_avg": 0.0013407982187345624,
+      "objective/train/docs_used": 601010,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.416961908340454,
+      "objective/train/original_loss": 1.416961908340454,
+      "objective/train/theoretical_loss": 3.5163143945682656,
+      "objective/train/tokens_used": 1665413600,
+      "objective/train/value_avg": -0.0066070556640625,
+      "objective/train/value_loss": 9.873360977508128e-05,
+      "objective/train/value_max": -0.00011146068572998047,
+      "objective/train/value_min": -0.261474609375,
+      "objective/train/value_reward_corr": 0.7415103919665526,
+      "objective/train/value_std": 0.0123443603515625,
+      "objective/train/weight_avg": 1.0013898611068726,
+      "objective/train/weighted_lm_loss": 1.4189093112945557,
+      "objective/train/weights_max": 1.2988438606262207,
+      "objective/train/weights_min": 0.7572005987167358,
+      "theoretical_loss": 3.5163143945682656,
+      "tokens_seen": 1644953600
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005065800032097577,
+      "loss": 0.0723,
+      "theoretical_loss": 3.5163143945682656,
+      "tokens_seen": 1644953600
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005064997592681752,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5162662550555392,
+      "tokens_seen": 1645215744
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005064195153265928,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5162181253599405,
+      "tokens_seen": 1645477888
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005063392713850105,
+      "loss": 0.0704,
+      "theoretical_loss": 3.516170005477904,
+      "tokens_seen": 1645740032
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000506259027443428,
+      "loss": 0.0699,
+      "theoretical_loss": 3.516121895405866,
+      "tokens_seen": 1646002176
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005061787835018457,
+      "loss": 0.0732,
+      "theoretical_loss": 3.516073795140265,
+      "tokens_seen": 1646264320
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005060985395602632,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5160257046775407,
+      "tokens_seen": 1646526464
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005060182956186808,
+      "loss": 0.0697,
+      "theoretical_loss": 3.515977624014135,
+      "tokens_seen": 1646788608
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005059380516770984,
+      "loss": 0.0681,
+      "theoretical_loss": 3.515929553146492,
+      "tokens_seen": 1647050752
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005058578077355159,
+      "loss": 0.068,
+      "theoretical_loss": 3.515881492071057,
+      "tokens_seen": 1647312896
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005057775637939335,
+      "loss": 0.0705,
+      "theoretical_loss": 3.515833440784278,
+      "tokens_seen": 1647575040
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005056973198523511,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5157853992826036,
+      "tokens_seen": 1647837184
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005056170759107688,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5157373675624854,
+      "tokens_seen": 1648099328
+    },
+    {
+      "epoch": 0.5,
+      "objective/train/advantage_avg": 0.0007609121967107058,
+      "objective/train/docs_used": 602144,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.442240595817566,
+      "objective/train/original_loss": 1.4422404766082764,
+      "objective/train/theoretical_loss": 3.5157133553694013,
+      "objective/train/tokens_used": 1668690400,
+      "objective/train/value_avg": -0.0084991455078125,
+      "objective/train/value_loss": 0.0003012235974892974,
+      "objective/train/value_max": -0.00011771917343139648,
+      "objective/train/value_min": -0.34228515625,
+      "objective/train/value_reward_corr": 0.6924308598838854,
+      "objective/train/value_std": 0.0149688720703125,
+      "objective/train/weight_avg": 1.0008883476257324,
+      "objective/train/weighted_lm_loss": 1.4430891275405884,
+      "objective/train/weights_max": 1.3021063804626465,
+      "objective/train/weights_min": 0.22860285639762878,
+      "theoretical_loss": 3.5157133553694013,
+      "tokens_seen": 1648230400
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005055368319691864,
+      "loss": 0.0712,
+      "theoretical_loss": 3.515689345620377,
+      "tokens_seen": 1648361472
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000505456588027604,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5156413334527317,
+      "tokens_seen": 1648623616
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005053763440860215,
+      "loss": 0.0697,
+      "theoretical_loss": 3.515593331056008,
+      "tokens_seen": 1648885760
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005052961001444391,
+      "loss": 0.068,
+      "theoretical_loss": 3.5155453384266635,
+      "tokens_seen": 1649147904
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005052158562028567,
+      "loss": 0.0679,
+      "theoretical_loss": 3.515497355561159,
+      "tokens_seen": 1649410048
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005051356122612742,
+      "loss": 0.0712,
+      "theoretical_loss": 3.515449382455957,
+      "tokens_seen": 1649672192
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005050553683196919,
+      "loss": 0.07,
+      "theoretical_loss": 3.515401419107521,
+      "tokens_seen": 1649934336
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005049751243781094,
+      "loss": 0.0682,
+      "theoretical_loss": 3.515353465512317,
+      "tokens_seen": 1650196480
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000504894880436527,
+      "loss": 0.0721,
+      "theoretical_loss": 3.5153055216668134,
+      "tokens_seen": 1650458624
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005048146364949447,
+      "loss": 0.0703,
+      "theoretical_loss": 3.51525758756748,
+      "tokens_seen": 1650720768
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005047343925533623,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5152096632107876,
+      "tokens_seen": 1650982912
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005046541486117798,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5151617485932096,
+      "tokens_seen": 1651245056
+    },
+    {
+      "epoch": 0.5,
+      "objective/train/advantage_avg": 0.00040685301064513624,
+      "objective/train/docs_used": 603335,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3096529245376587,
+      "objective/train/original_loss": 1.3096530437469482,
+      "objective/train/theoretical_loss": 3.5151138437112213,
+      "objective/train/tokens_used": 1671967200,
+      "objective/train/value_avg": -0.0088348388671875,
+      "objective/train/value_loss": 0.0004952080198563635,
+      "objective/train/value_max": -9.101629257202148e-05,
+      "objective/train/value_min": -0.8935546875,
+      "objective/train/value_reward_corr": 0.6091750570438847,
+      "objective/train/value_std": 0.01788330078125,
+      "objective/train/weight_avg": 1.0006325244903564,
+      "objective/train/weighted_lm_loss": 1.3091342449188232,
+      "objective/train/weights_max": 2.221797466278076,
+      "objective/train/weights_min": 0.3713989555835724,
+      "theoretical_loss": 3.5151138437112213,
+      "tokens_seen": 1651507200
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005045739046701974,
+      "loss": 0.0665,
+      "theoretical_loss": 3.5151138437112213,
+      "tokens_seen": 1651507200
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000504493660728615,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5150659485613005,
+      "tokens_seen": 1651769344
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005044134167870326,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5150180631399244,
+      "tokens_seen": 1652031488
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005043331728454502,
+      "loss": 0.069,
+      "theoretical_loss": 3.5149701874435753,
+      "tokens_seen": 1652293632
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005042529289038677,
+      "loss": 0.074,
+      "theoretical_loss": 3.514922321468734,
+      "tokens_seen": 1652555776
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005041726849622855,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5148744652118866,
+      "tokens_seen": 1652817920
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000504092441020703,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5148266186695185,
+      "tokens_seen": 1653080064
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005040121970791205,
+      "loss": 0.069,
+      "theoretical_loss": 3.5147787818381175,
+      "tokens_seen": 1653342208
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005039319531375382,
+      "loss": 0.0696,
+      "theoretical_loss": 3.514730954714173,
+      "tokens_seen": 1653604352
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005038517091959557,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5146831372941776,
+      "tokens_seen": 1653866496
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005037714652543733,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5146353295746247,
+      "tokens_seen": 1654128640
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005036912213127909,
+      "loss": 0.0687,
+      "theoretical_loss": 3.514587531552009,
+      "tokens_seen": 1654390784
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005036109773712085,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5145397432228274,
+      "tokens_seen": 1654652928
+    },
+    {
+      "epoch": 0.5,
+      "objective/train/advantage_avg": 0.0005927335005253553,
+      "objective/train/docs_used": 604403,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4796189069747925,
+      "objective/train/original_loss": 1.479619026184082,
+      "objective/train/theoretical_loss": 3.5145158526921807,
+      "objective/train/tokens_used": 1675244000,
+      "objective/train/value_avg": -0.009033203125,
+      "objective/train/value_loss": 0.00033400507527403533,
+      "objective/train/value_max": -6.300210952758789e-05,
+      "objective/train/value_min": -0.697265625,
+      "objective/train/value_reward_corr": 0.7117968222736691,
+      "objective/train/value_std": 0.0211029052734375,
+      "objective/train/weight_avg": 1.0007562637329102,
+      "objective/train/weighted_lm_loss": 1.4803884029388428,
+      "objective/train/weights_max": 1.8430651426315308,
+      "objective/train/weights_min": 0.5630186200141907,
+      "theoretical_loss": 3.5145158526921807,
+      "tokens_seen": 1654784000
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000503530733429626,
+      "loss": 0.0722,
+      "theoretical_loss": 3.5144919645835797,
+      "tokens_seen": 1654915072
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005034504894880436,
+      "loss": 0.068,
+      "theoretical_loss": 3.514444195630766,
+      "tokens_seen": 1655177216
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005033702455464613,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5143964363608893,
+      "tokens_seen": 1655439360
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005032900016048788,
+      "loss": 0.0719,
+      "theoretical_loss": 3.514348686770454,
+      "tokens_seen": 1655701504
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005032097576632965,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5143009468559656,
+      "tokens_seen": 1655963648
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000503129513721714,
+      "loss": 0.0686,
+      "theoretical_loss": 3.514253216613932,
+      "tokens_seen": 1656225792
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005030492697801317,
+      "loss": 0.0712,
+      "theoretical_loss": 3.514205496040865,
+      "tokens_seen": 1656487936
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005029690258385492,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5141577851332735,
+      "tokens_seen": 1656750080
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005028887818969667,
+      "loss": 0.0683,
+      "theoretical_loss": 3.5141100838876724,
+      "tokens_seen": 1657012224
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005028085379553844,
+      "loss": 0.0674,
+      "theoretical_loss": 3.5140623923005774,
+      "tokens_seen": 1657274368
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005027282940138019,
+      "loss": 0.0676,
+      "theoretical_loss": 3.514014710368505,
+      "tokens_seen": 1657536512
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005026480500722196,
+      "loss": 0.0675,
+      "theoretical_loss": 3.513967038087973,
+      "tokens_seen": 1657798656
+    },
+    {
+      "epoch": 0.5,
+      "objective/train/advantage_avg": 0.0010964076500386,
+      "objective/train/docs_used": 605625,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3653086423873901,
+      "objective/train/original_loss": 1.3653085231781006,
+      "objective/train/theoretical_loss": 3.5139193754555036,
+      "objective/train/tokens_used": 1678520800,
+      "objective/train/value_avg": -0.007221221923828125,
+      "objective/train/value_loss": 0.00029157879180274904,
+      "objective/train/value_max": -7.545948028564453e-05,
+      "objective/train/value_min": -0.355712890625,
+      "objective/train/value_reward_corr": 0.6027662181567327,
+      "objective/train/value_std": 0.012420654296875,
+      "objective/train/weight_avg": 1.001227617263794,
+      "objective/train/weighted_lm_loss": 1.367215633392334,
+      "objective/train/weights_max": 1.1224560737609863,
+      "objective/train/weights_min": 0.3705088496208191,
+      "theoretical_loss": 3.5139193754555036,
+      "tokens_seen": 1658060800
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005025678061306372,
+      "loss": 0.069,
+      "theoretical_loss": 3.5139193754555036,
+      "tokens_seen": 1658060800
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005024875621890548,
+      "loss": 0.0674,
+      "theoretical_loss": 3.513871722467619,
+      "tokens_seen": 1658322944
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005024073182474723,
+      "loss": 0.0701,
+      "theoretical_loss": 3.513824079120843,
+      "tokens_seen": 1658585088
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005023270743058899,
+      "loss": 0.0698,
+      "theoretical_loss": 3.513776445411702,
+      "tokens_seen": 1658847232
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005022468303643075,
+      "loss": 0.067,
+      "theoretical_loss": 3.5137288213367235,
+      "tokens_seen": 1659109376
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000502166586422725,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5136812068924375,
+      "tokens_seen": 1659371520
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005020863424811427,
+      "loss": 0.0689,
+      "theoretical_loss": 3.513633602075376,
+      "tokens_seen": 1659633664
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005020060985395602,
+      "loss": 0.0685,
+      "theoretical_loss": 3.513586006882071,
+      "tokens_seen": 1659895808
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000501925854597978,
+      "loss": 0.0693,
+      "theoretical_loss": 3.513538421309059,
+      "tokens_seen": 1660157952
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005018456106563955,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5134908453528757,
+      "tokens_seen": 1660420096
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000501765366714813,
+      "loss": 0.0683,
+      "theoretical_loss": 3.51344327901006,
+      "tokens_seen": 1660682240
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005016851227732307,
+      "loss": 0.0697,
+      "theoretical_loss": 3.513395722277153,
+      "tokens_seen": 1660944384
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005016048788316482,
+      "loss": 0.0706,
+      "theoretical_loss": 3.513348175150696,
+      "tokens_seen": 1661206528
+    },
+    {
+      "epoch": 0.5,
+      "objective/train/advantage_avg": -1.5146979421842843e-05,
+      "objective/train/docs_used": 606795,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4264380931854248,
+      "objective/train/original_loss": 1.4264380931854248,
+      "objective/train/theoretical_loss": 3.513324405188806,
+      "objective/train/tokens_used": 1681797600,
+      "objective/train/value_avg": -0.007694244384765625,
+      "objective/train/value_loss": 0.0006180181517265737,
+      "objective/train/value_max": -4.100799560546875e-05,
+      "objective/train/value_min": -0.75,
+      "objective/train/value_reward_corr": 0.7114442243862558,
+      "objective/train/value_std": 0.0225982666015625,
+      "objective/train/weight_avg": 1.0002539157867432,
+      "objective/train/weighted_lm_loss": 1.42630934715271,
+      "objective/train/weights_max": 1.878307580947876,
+      "objective/train/weights_min": 0.36932775378227234,
+      "theoretical_loss": 3.513324405188806,
+      "tokens_seen": 1661337600
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005015246348900658,
+      "loss": 0.0659,
+      "theoretical_loss": 3.5133006376272338,
+      "tokens_seen": 1661468672
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005014443909484834,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5132531097033115,
+      "tokens_seen": 1661730816
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000501364147006901,
+      "loss": 0.0655,
+      "theoretical_loss": 3.5132055913754776,
+      "tokens_seen": 1661992960
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005012839030653185,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5131580826402806,
+      "tokens_seen": 1662255104
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005012036591237363,
+      "loss": 0.0704,
+      "theoretical_loss": 3.5131105834942726,
+      "tokens_seen": 1662517248
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005011234151821538,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5130630939340053,
+      "tokens_seen": 1662779392
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005010431712405713,
+      "loss": 0.069,
+      "theoretical_loss": 3.5130156139560347,
+      "tokens_seen": 1663041536
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.000500962927298989,
+      "loss": 0.0705,
+      "theoretical_loss": 3.512968143556917,
+      "tokens_seen": 1663303680
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005008826833574065,
+      "loss": 0.0691,
+      "theoretical_loss": 3.51292068273321,
+      "tokens_seen": 1663565824
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005008024394158241,
+      "loss": 0.0681,
+      "theoretical_loss": 3.512873231481474,
+      "tokens_seen": 1663827968
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005007221954742417,
+      "loss": 0.0682,
+      "theoretical_loss": 3.512825789798271,
+      "tokens_seen": 1664090112
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005006419515326593,
+      "loss": 0.0683,
+      "theoretical_loss": 3.5127783576801646,
+      "tokens_seen": 1664352256
+    },
+    {
+      "epoch": 0.5,
+      "objective/train/advantage_avg": 0.0006472986424341798,
+      "objective/train/docs_used": 608106,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2632955312728882,
+      "objective/train/original_loss": 1.2632955312728882,
+      "objective/train/theoretical_loss": 3.51273093512372,
+      "objective/train/tokens_used": 1685074400,
+      "objective/train/value_avg": -0.01052093505859375,
+      "objective/train/value_loss": 0.0002703302598092705,
+      "objective/train/value_max": -4.649162292480469e-05,
+      "objective/train/value_min": -0.70849609375,
+      "objective/train/value_reward_corr": 0.7499424196920753,
+      "objective/train/value_std": 0.0195159912109375,
+      "objective/train/weight_avg": 1.000773549079895,
+      "objective/train/weighted_lm_loss": 1.263318657875061,
+      "objective/train/weights_max": 1.3154696226119995,
+      "objective/train/weights_min": 0.4099985957145691,
+      "theoretical_loss": 3.51273093512372,
+      "tokens_seen": 1664614400
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005005617075910769,
+      "loss": 0.0707,
+      "theoretical_loss": 3.51273093512372,
+      "tokens_seen": 1664614400
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005004814636494944,
+      "loss": 0.0682,
+      "theoretical_loss": 3.512683522125505,
+      "tokens_seen": 1664876544
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005004012197079121,
+      "loss": 0.0682,
+      "theoretical_loss": 3.512636118682088,
+      "tokens_seen": 1665138688
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005003209757663297,
+      "loss": 0.0688,
+      "theoretical_loss": 3.5125887247900396,
+      "tokens_seen": 1665400832
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005002407318247473,
+      "loss": 0.071,
+      "theoretical_loss": 3.512541340445933,
+      "tokens_seen": 1665662976
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005001604878831648,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5124939656463416,
+      "tokens_seen": 1665925120
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 0.0005000802439415825,
+      "loss": 0.0678,
+      "theoretical_loss": 3.5124466003878423,
+      "tokens_seen": 1666187264
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0005,
+      "loss": 0.0694,
+      "theoretical_loss": 3.512399244667012,
+      "tokens_seen": 1666449408
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004999197560584176,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5123518984804303,
+      "tokens_seen": 1666711552
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004998395121168352,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5123045618246795,
+      "tokens_seen": 1666973696
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004997592681752528,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5122572346963423,
+      "tokens_seen": 1667235840
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004996790242336704,
+      "loss": 0.0682,
+      "theoretical_loss": 3.512209917092003,
+      "tokens_seen": 1667497984
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000499598780292088,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5121626090082487,
+      "tokens_seen": 1667760128
+    },
+    {
+      "epoch": 0.51,
+      "objective/train/advantage_avg": 0.0011781208449974656,
+      "objective/train/docs_used": 609331,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5010648965835571,
+      "objective/train/original_loss": 1.5010650157928467,
+      "objective/train/theoretical_loss": 3.512138958535525,
+      "objective/train/tokens_used": 1688351200,
+      "objective/train/value_avg": -0.004985809326171875,
+      "objective/train/value_loss": 0.000247854070039466,
+      "objective/train/value_max": -4.166364669799805e-05,
+      "objective/train/value_min": -0.64453125,
+      "objective/train/value_reward_corr": 0.6039150359454911,
+      "objective/train/value_std": 0.00983428955078125,
+      "objective/train/weight_avg": 1.0012835264205933,
+      "objective/train/weighted_lm_loss": 1.5030659437179565,
+      "objective/train/weights_max": 1.1962590217590332,
+      "objective/train/weights_min": 0.3713054358959198,
+      "theoretical_loss": 3.512138958535525,
+      "tokens_seen": 1667891200
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004995185363505056,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5121153104416676,
+      "tokens_seen": 1668022272
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004994382924089231,
+      "loss": 0.0654,
+      "theoretical_loss": 3.5120680213888504,
+      "tokens_seen": 1668284416
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004993580484673407,
+      "loss": 0.0689,
+      "theoretical_loss": 3.512020741846388,
+      "tokens_seen": 1668546560
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004992778045257583,
+      "loss": 0.0675,
+      "theoretical_loss": 3.511973471810875,
+      "tokens_seen": 1668808704
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004991975605841759,
+      "loss": 0.069,
+      "theoretical_loss": 3.5119262112789063,
+      "tokens_seen": 1669070848
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004991173166425935,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5118789602470786,
+      "tokens_seen": 1669332992
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000499037072701011,
+      "loss": 0.0727,
+      "theoretical_loss": 3.5118317187119916,
+      "tokens_seen": 1669595136
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004989568287594286,
+      "loss": 0.07,
+      "theoretical_loss": 3.511784486670246,
+      "tokens_seen": 1669857280
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004988765848178463,
+      "loss": 0.0673,
+      "theoretical_loss": 3.5117372641184432,
+      "tokens_seen": 1670119424
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004987963408762639,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5116900510531885,
+      "tokens_seen": 1670381568
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004987160969346815,
+      "loss": 0.069,
+      "theoretical_loss": 3.5116428474710872,
+      "tokens_seen": 1670643712
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004986358529930991,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5115956533687473,
+      "tokens_seen": 1670905856
+    },
+    {
+      "epoch": 0.51,
+      "objective/train/advantage_avg": 0.00025810531224124134,
+      "objective/train/docs_used": 610761,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.323618769645691,
+      "objective/train/original_loss": 1.3236186504364014,
+      "objective/train/theoretical_loss": 3.5115484687427783,
+      "objective/train/tokens_used": 1691628000,
+      "objective/train/value_avg": -0.005733489990234375,
+      "objective/train/value_loss": 0.00011536670353962108,
+      "objective/train/value_max": -9.387731552124023e-05,
+      "objective/train/value_min": -0.31689453125,
+      "objective/train/value_reward_corr": 0.7299094363574254,
+      "objective/train/value_std": 0.0103912353515625,
+      "objective/train/weight_avg": 1.0003145933151245,
+      "objective/train/weighted_lm_loss": 1.323650598526001,
+      "objective/train/weights_max": 1.1653170585632324,
+      "objective/train/weights_min": 0.723071813583374,
+      "theoretical_loss": 3.5115484687427783,
+      "tokens_seen": 1671168000
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004985556090515166,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5115484687427783,
+      "tokens_seen": 1671168000
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004984753651099342,
+      "loss": 0.0652,
+      "theoretical_loss": 3.5115012935897907,
+      "tokens_seen": 1671430144
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004983951211683518,
+      "loss": 0.0676,
+      "theoretical_loss": 3.5114541279063975,
+      "tokens_seen": 1671692288
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004983148772267693,
+      "loss": 0.0691,
+      "theoretical_loss": 3.511406971689214,
+      "tokens_seen": 1671954432
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004982346332851869,
+      "loss": 0.0672,
+      "theoretical_loss": 3.511359824934856,
+      "tokens_seen": 1672216576
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004981543893436046,
+      "loss": 0.0665,
+      "theoretical_loss": 3.5113126876399416,
+      "tokens_seen": 1672478720
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004980741454020222,
+      "loss": 0.0676,
+      "theoretical_loss": 3.511265559801091,
+      "tokens_seen": 1672740864
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004979939014604398,
+      "loss": 0.069,
+      "theoretical_loss": 3.5112184414149255,
+      "tokens_seen": 1673003008
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004979136575188573,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5111713324780687,
+      "tokens_seen": 1673265152
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004978334135772749,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5111242329871457,
+      "tokens_seen": 1673527296
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004977531696356925,
+      "loss": 0.0688,
+      "theoretical_loss": 3.5110771429387824,
+      "tokens_seen": 1673789440
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004976729256941101,
+      "loss": 0.0688,
+      "theoretical_loss": 3.511030062329608,
+      "tokens_seen": 1674051584
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004975926817525277,
+      "loss": 0.069,
+      "theoretical_loss": 3.5109829911562533,
+      "tokens_seen": 1674313728
+    },
+    {
+      "epoch": 0.51,
+      "objective/train/advantage_avg": 0.00035731392563320696,
+      "objective/train/docs_used": 611357,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3454358577728271,
+      "objective/train/original_loss": 1.3454358577728271,
+      "objective/train/theoretical_loss": 3.5109594591069553,
+      "objective/train/tokens_used": 1694904800,
+      "objective/train/value_avg": -0.005359649658203125,
+      "objective/train/value_loss": 0.00013092935841996223,
+      "objective/train/value_max": -5.519390106201172e-05,
+      "objective/train/value_min": -0.72705078125,
+      "objective/train/value_reward_corr": 0.6681693229191178,
+      "objective/train/value_std": 0.01116180419921875,
+      "objective/train/weight_avg": 1.0004169940948486,
+      "objective/train/weighted_lm_loss": 1.3458484411239624,
+      "objective/train/weights_max": 1.2602665424346924,
+      "objective/train/weights_min": 0.37362140417099,
+      "theoretical_loss": 3.5109594591069553,
+      "tokens_seen": 1674444800
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004975124378109452,
+      "loss": 0.0662,
+      "theoretical_loss": 3.5109359294153495,
+      "tokens_seen": 1674575872
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004974321938693629,
+      "loss": 0.0677,
+      "theoretical_loss": 3.5108888771035307,
+      "tokens_seen": 1674838016
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004973519499277805,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5108418342174317,
+      "tokens_seen": 1675100160
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004972717059861981,
+      "loss": 0.0676,
+      "theoretical_loss": 3.5107948007536907,
+      "tokens_seen": 1675362304
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004971914620446156,
+      "loss": 0.0673,
+      "theoretical_loss": 3.5107477767089454,
+      "tokens_seen": 1675624448
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004971112181030332,
+      "loss": 0.0662,
+      "theoretical_loss": 3.5107007620798374,
+      "tokens_seen": 1675886592
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004970309741614508,
+      "loss": 0.0681,
+      "theoretical_loss": 3.510653756863009,
+      "tokens_seen": 1676148736
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004969507302198684,
+      "loss": 0.0675,
+      "theoretical_loss": 3.5106067610551035,
+      "tokens_seen": 1676410880
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000496870486278286,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5105597746527675,
+      "tokens_seen": 1676673024
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004967902423367036,
+      "loss": 0.0657,
+      "theoretical_loss": 3.5105127976526482,
+      "tokens_seen": 1676935168
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004967099983951212,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5104658300513942,
+      "tokens_seen": 1677197312
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004966297544535388,
+      "loss": 0.0655,
+      "theoretical_loss": 3.5104188718456575,
+      "tokens_seen": 1677459456
+    },
+    {
+      "epoch": 0.51,
+      "objective/train/advantage_avg": 0.00032650245702825487,
+      "objective/train/docs_used": 612648,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2856462001800537,
+      "objective/train/original_loss": 1.2856462001800537,
+      "objective/train/theoretical_loss": 3.5103719230320904,
+      "objective/train/tokens_used": 1698181600,
+      "objective/train/value_avg": -0.00855255126953125,
+      "objective/train/value_loss": 0.0005081266863271594,
+      "objective/train/value_max": -4.988908767700195e-05,
+      "objective/train/value_min": -0.74755859375,
+      "objective/train/value_reward_corr": 0.6733552587749285,
+      "objective/train/value_std": 0.016265869140625,
+      "objective/train/weight_avg": 1.0005440711975098,
+      "objective/train/weighted_lm_loss": 1.2856409549713135,
+      "objective/train/weights_max": 1.4916789531707764,
+      "objective/train/weights_min": 0.3753184378147125,
+      "theoretical_loss": 3.5103719230320904,
+      "tokens_seen": 1677721600
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004965495105119564,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5103719230320904,
+      "tokens_seen": 1677721600
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000496469266570374,
+      "loss": 0.0691,
+      "theoretical_loss": 3.510324983607347,
+      "tokens_seen": 1677983744
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004963890226287915,
+      "loss": 0.0683,
+      "theoretical_loss": 3.5102780535680838,
+      "tokens_seen": 1678245888
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004963087786872091,
+      "loss": 0.07,
+      "theoretical_loss": 3.510231132910958,
+      "tokens_seen": 1678508032
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004962285347456267,
+      "loss": 0.068,
+      "theoretical_loss": 3.5101842216326293,
+      "tokens_seen": 1678770176
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004961482908040443,
+      "loss": 0.0674,
+      "theoretical_loss": 3.5101373197297594,
+      "tokens_seen": 1679032320
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004960680468624618,
+      "loss": 0.0671,
+      "theoretical_loss": 3.510090427199011,
+      "tokens_seen": 1679294464
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004959878029208794,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5100435440370483,
+      "tokens_seen": 1679556608
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004959075589792971,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5099966702405383,
+      "tokens_seen": 1679818752
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004958273150377147,
+      "loss": 0.0713,
+      "theoretical_loss": 3.5099498058061487,
+      "tokens_seen": 1680080896
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004957470710961323,
+      "loss": 0.068,
+      "theoretical_loss": 3.5099029507305497,
+      "tokens_seen": 1680343040
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004956668271545499,
+      "loss": 0.0704,
+      "theoretical_loss": 3.509856105010412,
+      "tokens_seen": 1680605184
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004955865832129674,
+      "loss": 0.0729,
+      "theoretical_loss": 3.509809268642409,
+      "tokens_seen": 1680867328
+    },
+    {
+      "epoch": 0.51,
+      "objective/train/advantage_avg": -0.0005266556981950998,
+      "objective/train/docs_used": 613922,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3320040702819824,
+      "objective/train/original_loss": 1.3320040702819824,
+      "objective/train/theoretical_loss": 3.509785853964419,
+      "objective/train/tokens_used": 1701458400,
+      "objective/train/value_avg": -0.01157379150390625,
+      "objective/train/value_loss": 0.00023857805354055017,
+      "objective/train/value_max": -3.4809112548828125e-05,
+      "objective/train/value_min": -0.982421875,
+      "objective/train/value_reward_corr": 0.9455989378013424,
+      "objective/train/value_std": 0.04180908203125,
+      "objective/train/weight_avg": 0.9995850920677185,
+      "objective/train/weighted_lm_loss": 1.331617832183838,
+      "objective/train/weights_max": 1.1823036670684814,
+      "objective/train/weights_min": 0.3787992298603058,
+      "theoretical_loss": 3.509785853964419,
+      "tokens_seen": 1680998400
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000495506339271385,
+      "loss": 0.071,
+      "theoretical_loss": 3.5097624416232156,
+      "tokens_seen": 1681129472
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004954260953298026,
+      "loss": 0.0702,
+      "theoretical_loss": 3.509715623949509,
+      "tokens_seen": 1681391616
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004953458513882202,
+      "loss": 0.0711,
+      "theoretical_loss": 3.509668815617967,
+      "tokens_seen": 1681653760
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004952656074466377,
+      "loss": 0.0692,
+      "theoretical_loss": 3.5096220166252694,
+      "tokens_seen": 1681915904
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004951853635050554,
+      "loss": 0.0683,
+      "theoretical_loss": 3.509575226968098,
+      "tokens_seen": 1682178048
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000495105119563473,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5095284466431362,
+      "tokens_seen": 1682440192
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004950248756218906,
+      "loss": 0.0681,
+      "theoretical_loss": 3.509481675647069,
+      "tokens_seen": 1682702336
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004949446316803081,
+      "loss": 0.0684,
+      "theoretical_loss": 3.509434913976583,
+      "tokens_seen": 1682964480
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004948643877387257,
+      "loss": 0.0694,
+      "theoretical_loss": 3.509388161628367,
+      "tokens_seen": 1683226624
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004947841437971433,
+      "loss": 0.0699,
+      "theoretical_loss": 3.509341418599111,
+      "tokens_seen": 1683488768
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004947038998555609,
+      "loss": 0.0695,
+      "theoretical_loss": 3.509294684885506,
+      "tokens_seen": 1683750912
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004946236559139785,
+      "loss": 0.0712,
+      "theoretical_loss": 3.509247960484247,
+      "tokens_seen": 1684013056
+    },
+    {
+      "epoch": 0.51,
+      "objective/train/advantage_avg": 0.00041788024827837944,
+      "objective/train/docs_used": 615159,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2733129262924194,
+      "objective/train/original_loss": 1.273313045501709,
+      "objective/train/theoretical_loss": 3.5092012453920285,
+      "objective/train/tokens_used": 1704735200,
+      "objective/train/value_avg": -0.00682830810546875,
+      "objective/train/value_loss": 0.00011341172648826614,
+      "objective/train/value_max": -5.143880844116211e-05,
+      "objective/train/value_min": -0.26318359375,
+      "objective/train/value_reward_corr": 0.6867672303390299,
+      "objective/train/value_std": 0.0099029541015625,
+      "objective/train/weight_avg": 1.0004736185073853,
+      "objective/train/weighted_lm_loss": 1.2733094692230225,
+      "objective/train/weights_max": 1.2000617980957031,
+      "objective/train/weights_min": 0.7181033492088318,
+      "theoretical_loss": 3.5092012453920285,
+      "tokens_seen": 1684275200
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004945434119723962,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5092012453920285,
+      "tokens_seen": 1684275200
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004944631680308137,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5091545396055475,
+      "tokens_seen": 1684537344
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004943829240892313,
+      "loss": 0.0676,
+      "theoretical_loss": 3.5091078431215017,
+      "tokens_seen": 1684799488
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004943026801476489,
+      "loss": 0.0674,
+      "theoretical_loss": 3.5090611559365925,
+      "tokens_seen": 1685061632
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004942224362060665,
+      "loss": 0.0686,
+      "theoretical_loss": 3.509014478047522,
+      "tokens_seen": 1685323776
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000494142192264484,
+      "loss": 0.071,
+      "theoretical_loss": 3.5089678094509926,
+      "tokens_seen": 1685585920
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004940619483229016,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5089211501437103,
+      "tokens_seen": 1685848064
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004939817043813192,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5088745001223822,
+      "tokens_seen": 1686110208
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004939014604397368,
+      "loss": 0.0699,
+      "theoretical_loss": 3.508827859383717,
+      "tokens_seen": 1686372352
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004938212164981544,
+      "loss": 0.0675,
+      "theoretical_loss": 3.5087812279244246,
+      "tokens_seen": 1686634496
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000493740972556572,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5087346057412176,
+      "tokens_seen": 1686896640
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004936607286149896,
+      "loss": 0.0734,
+      "theoretical_loss": 3.508687992830809,
+      "tokens_seen": 1687158784
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004935804846734072,
+      "loss": 0.0732,
+      "theoretical_loss": 3.5086413891899144,
+      "tokens_seen": 1687420928
+    },
+    {
+      "epoch": 0.51,
+      "objective/train/advantage_avg": 0.00013062905054539442,
+      "objective/train/docs_used": 616363,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4347667694091797,
+      "objective/train/original_loss": 1.4347667694091797,
+      "objective/train/theoretical_loss": 3.508618090844509,
+      "objective/train/tokens_used": 1708012000,
+      "objective/train/value_avg": -0.007480621337890625,
+      "objective/train/value_loss": 0.00023279149900190532,
+      "objective/train/value_max": -5.435943603515625e-05,
+      "objective/train/value_min": -0.2467041015625,
+      "objective/train/value_reward_corr": 0.7050742886378683,
+      "objective/train/value_std": 0.0143280029296875,
+      "objective/train/weight_avg": 1.0002341270446777,
+      "objective/train/weighted_lm_loss": 1.4352725744247437,
+      "objective/train/weights_max": 1.1323878765106201,
+      "objective/train/weights_min": 0.37623587250709534,
+      "theoretical_loss": 3.508618090844509,
+      "tokens_seen": 1687552000
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004935002407318248,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5085947948152514,
+      "tokens_seen": 1687683072
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004934199967902424,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5085482097035383,
+      "tokens_seen": 1687945216
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004933397528486599,
+      "loss": 0.0694,
+      "theoretical_loss": 3.508501633851495,
+      "tokens_seen": 1688207360
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004932595089070775,
+      "loss": 0.0672,
+      "theoretical_loss": 3.5084550672558446,
+      "tokens_seen": 1688469504
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004931792649654951,
+      "loss": 0.0694,
+      "theoretical_loss": 3.50840850991331,
+      "tokens_seen": 1688731648
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004930990210239126,
+      "loss": 0.0656,
+      "theoretical_loss": 3.5083619618206168,
+      "tokens_seen": 1688993792
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004930187770823302,
+      "loss": 0.0663,
+      "theoretical_loss": 3.5083154229744924,
+      "tokens_seen": 1689255936
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004929385331407479,
+      "loss": 0.0734,
+      "theoretical_loss": 3.5082688933716653,
+      "tokens_seen": 1689518080
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004928582891991655,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5082223730088655,
+      "tokens_seen": 1689780224
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004927780452575831,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5081758618828256,
+      "tokens_seen": 1690042368
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004926978013160007,
+      "loss": 0.0683,
+      "theoretical_loss": 3.5081293599902788,
+      "tokens_seen": 1690304512
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004926175573744183,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5080828673279614,
+      "tokens_seen": 1690566656
+    },
+    {
+      "epoch": 0.51,
+      "objective/train/advantage_avg": 0.0007474888698197901,
+      "objective/train/docs_used": 617433,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3971322774887085,
+      "objective/train/original_loss": 1.397132396697998,
+      "objective/train/theoretical_loss": 3.5080363838926094,
+      "objective/train/tokens_used": 1711288800,
+      "objective/train/value_avg": -0.007564544677734375,
+      "objective/train/value_loss": 0.0004391712718643248,
+      "objective/train/value_max": -5.692243576049805e-05,
+      "objective/train/value_min": -0.73828125,
+      "objective/train/value_reward_corr": 0.6730230280800624,
+      "objective/train/value_std": 0.021270751953125,
+      "objective/train/weight_avg": 1.0009394884109497,
+      "objective/train/weighted_lm_loss": 1.3970917463302612,
+      "objective/train/weights_max": 1.8592804670333862,
+      "objective/train/weights_min": 0.3687421381473541,
+      "theoretical_loss": 3.5080363838926094,
+      "tokens_seen": 1690828800
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004925373134328358,
+      "loss": 0.0688,
+      "theoretical_loss": 3.5080363838926094,
+      "tokens_seen": 1690828800
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004924570694912534,
+      "loss": 0.069,
+      "theoretical_loss": 3.5079899096809624,
+      "tokens_seen": 1691090944
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000492376825549671,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5079434446897597,
+      "tokens_seen": 1691353088
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004922965816080887,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5078969889157445,
+      "tokens_seen": 1691615232
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004922163376665062,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5078505423556594,
+      "tokens_seen": 1691877376
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004921360937249238,
+      "loss": 0.0707,
+      "theoretical_loss": 3.50780410500625,
+      "tokens_seen": 1692139520
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004920558497833414,
+      "loss": 0.0707,
+      "theoretical_loss": 3.507757676864264,
+      "tokens_seen": 1692401664
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004919756058417589,
+      "loss": 0.067,
+      "theoretical_loss": 3.5077112579264496,
+      "tokens_seen": 1692663808
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004918953619001765,
+      "loss": 0.0693,
+      "theoretical_loss": 3.507664848189557,
+      "tokens_seen": 1692925952
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004918151179585941,
+      "loss": 0.0716,
+      "theoretical_loss": 3.507618447650337,
+      "tokens_seen": 1693188096
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004917348740170117,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5075720563055457,
+      "tokens_seen": 1693450240
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004916546300754293,
+      "loss": 0.0668,
+      "theoretical_loss": 3.5075256741519363,
+      "tokens_seen": 1693712384
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000491574386133847,
+      "loss": 0.0667,
+      "theoretical_loss": 3.507479301186266,
+      "tokens_seen": 1693974528
+    },
+    {
+      "epoch": 0.51,
+      "objective/train/advantage_avg": -0.00035341508919373155,
+      "objective/train/docs_used": 618718,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4741839170455933,
+      "objective/train/original_loss": 1.4741837978363037,
+      "objective/train/theoretical_loss": 3.507456118147896,
+      "objective/train/tokens_used": 1714565600,
+      "objective/train/value_avg": -0.00838470458984375,
+      "objective/train/value_loss": 0.00018926823395304382,
+      "objective/train/value_max": -5.143880844116211e-05,
+      "objective/train/value_min": -0.35107421875,
+      "objective/train/value_reward_corr": 0.6988737523826818,
+      "objective/train/value_std": 0.01287078857421875,
+      "objective/train/weight_avg": 0.9997391104698181,
+      "objective/train/weighted_lm_loss": 1.4718536138534546,
+      "objective/train/weights_max": 1.1573328971862793,
+      "objective/train/weights_min": 0.6129269003868103,
+      "theoretical_loss": 3.507456118147896,
+      "tokens_seen": 1694105600
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004914941421922646,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5074329374052944,
+      "tokens_seen": 1694236672
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004914138982506821,
+      "loss": 0.0663,
+      "theoretical_loss": 3.507386582805781,
+      "tokens_seen": 1694498816
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004913336543090997,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5073402373844864,
+      "tokens_seen": 1694760960
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004912534103675173,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5072939011381763,
+      "tokens_seen": 1695023104
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004911731664259348,
+      "loss": 0.0672,
+      "theoretical_loss": 3.507247574063614,
+      "tokens_seen": 1695285248
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004910929224843524,
+      "loss": 0.0664,
+      "theoretical_loss": 3.5072012561575674,
+      "tokens_seen": 1695547392
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.00049101267854277,
+      "loss": 0.0674,
+      "theoretical_loss": 3.5071549474168036,
+      "tokens_seen": 1695809536
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004909324346011876,
+      "loss": 0.0698,
+      "theoretical_loss": 3.507108647838094,
+      "tokens_seen": 1696071680
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004908521906596052,
+      "loss": 0.0671,
+      "theoretical_loss": 3.507062357418209,
+      "tokens_seen": 1696333824
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004907719467180228,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5070160761539233,
+      "tokens_seen": 1696595968
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004906917027764404,
+      "loss": 0.0682,
+      "theoretical_loss": 3.506969804042011,
+      "tokens_seen": 1696858112
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000490611458834858,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5069235410792485,
+      "tokens_seen": 1697120256
+    },
+    {
+      "epoch": 0.51,
+      "objective/train/advantage_avg": 0.0010268003679811954,
+      "objective/train/docs_used": 619912,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3225995302200317,
+      "objective/train/original_loss": 1.3225996494293213,
+      "objective/train/theoretical_loss": 3.506877287262414,
+      "objective/train/tokens_used": 1717842400,
+      "objective/train/value_avg": -0.00870513916015625,
+      "objective/train/value_loss": 0.0001602449337951839,
+      "objective/train/value_max": -4.756450653076172e-05,
+      "objective/train/value_min": -0.217529296875,
+      "objective/train/value_reward_corr": 0.7187718829305348,
+      "objective/train/value_std": 0.01302337646484375,
+      "objective/train/weight_avg": 1.0011017322540283,
+      "objective/train/weighted_lm_loss": 1.3236242532730103,
+      "objective/train/weights_max": 1.0969452857971191,
+      "objective/train/weights_min": 0.36948272585868835,
+      "theoretical_loss": 3.506877287262414,
+      "tokens_seen": 1697382400
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004905312148932756,
+      "loss": 0.0691,
+      "theoretical_loss": 3.506877287262414,
+      "tokens_seen": 1697382400
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004904509709516932,
+      "loss": 0.0671,
+      "theoretical_loss": 3.5068310425882876,
+      "tokens_seen": 1697644544
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004903707270101108,
+      "loss": 0.0697,
+      "theoretical_loss": 3.506784807053651,
+      "tokens_seen": 1697906688
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004902904830685283,
+      "loss": 0.0703,
+      "theoretical_loss": 3.506738580655287,
+      "tokens_seen": 1698168832
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004902102391269459,
+      "loss": 0.0672,
+      "theoretical_loss": 3.5066923633899796,
+      "tokens_seen": 1698430976
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004901299951853635,
+      "loss": 0.0683,
+      "theoretical_loss": 3.506646155254516,
+      "tokens_seen": 1698693120
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.000490049751243781,
+      "loss": 0.0687,
+      "theoretical_loss": 3.506599956245684,
+      "tokens_seen": 1698955264
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 0.0004899695073021987,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5065537663602737,
+      "tokens_seen": 1699217408
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004898892633606163,
+      "loss": 0.0679,
+      "theoretical_loss": 3.506507585595075,
+      "tokens_seen": 1699479552
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004898090194190339,
+      "loss": 0.0682,
+      "theoretical_loss": 3.506461413946882,
+      "tokens_seen": 1699741696
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004897287754774515,
+      "loss": 0.0653,
+      "theoretical_loss": 3.5064152514124887,
+      "tokens_seen": 1700003840
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004896485315358691,
+      "loss": 0.0702,
+      "theoretical_loss": 3.506369097988691,
+      "tokens_seen": 1700265984
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004895682875942866,
+      "loss": 0.068,
+      "theoretical_loss": 3.5063229536722864,
+      "tokens_seen": 1700528128
+    },
+    {
+      "epoch": 0.52,
+      "objective/train/advantage_avg": 0.001018267241306603,
+      "objective/train/docs_used": 621118,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4754199981689453,
+      "objective/train/original_loss": 1.4754199981689453,
+      "objective/train/theoretical_loss": 3.5062998849283566,
+      "objective/train/tokens_used": 1721119200,
+      "objective/train/value_avg": -0.006374359130859375,
+      "objective/train/value_loss": 0.0001637029490666464,
+      "objective/train/value_max": -5.519390106201172e-05,
+      "objective/train/value_min": -0.374755859375,
+      "objective/train/value_reward_corr": 0.7205456742670221,
+      "objective/train/value_std": 0.0139312744140625,
+      "objective/train/weight_avg": 1.0010955333709717,
+      "objective/train/weighted_lm_loss": 1.4766151905059814,
+      "objective/train/weights_max": 1.219853162765503,
+      "objective/train/weights_min": 0.388961523771286,
+      "theoretical_loss": 3.5062998849283566,
+      "tokens_seen": 1700659200
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004894880436527042,
+      "loss": 0.0688,
+      "theoretical_loss": 3.5062768184600754,
+      "tokens_seen": 1700790272
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004894077997111218,
+      "loss": 0.0672,
+      "theoretical_loss": 3.5062306923488573,
+      "tokens_seen": 1701052416
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004893275557695395,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5061845753354355,
+      "tokens_seen": 1701314560
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.000489247311827957,
+      "loss": 0.0674,
+      "theoretical_loss": 3.5061384674166147,
+      "tokens_seen": 1701576704
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004891670678863746,
+      "loss": 0.0665,
+      "theoretical_loss": 3.5060923685892,
+      "tokens_seen": 1701838848
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004890868239447922,
+      "loss": 0.0714,
+      "theoretical_loss": 3.5060462788499986,
+      "tokens_seen": 1702100992
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004890065800032098,
+      "loss": 0.0685,
+      "theoretical_loss": 3.50600019819582,
+      "tokens_seen": 1702363136
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004889263360616273,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5059541266234744,
+      "tokens_seen": 1702625280
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004888460921200449,
+      "loss": 0.0689,
+      "theoretical_loss": 3.505908064129775,
+      "tokens_seen": 1702887424
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004887658481784625,
+      "loss": 0.0654,
+      "theoretical_loss": 3.5058620107115344,
+      "tokens_seen": 1703149568
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004886856042368801,
+      "loss": 0.068,
+      "theoretical_loss": 3.505815966365568,
+      "tokens_seen": 1703411712
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004886053602952977,
+      "loss": 0.07,
+      "theoretical_loss": 3.5057699310886945,
+      "tokens_seen": 1703673856
+    },
+    {
+      "epoch": 0.52,
+      "objective/train/advantage_avg": 0.0014890459133312106,
+      "objective/train/docs_used": 622235,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3723922967910767,
+      "objective/train/original_loss": 1.372392177581787,
+      "objective/train/theoretical_loss": 3.505723904877731,
+      "objective/train/tokens_used": 1724396000,
+      "objective/train/value_avg": -0.00720977783203125,
+      "objective/train/value_loss": 0.00032538699451833963,
+      "objective/train/value_max": -9.763240814208984e-05,
+      "objective/train/value_min": -0.9609375,
+      "objective/train/value_reward_corr": 0.6740766317905353,
+      "objective/train/value_std": 0.0184173583984375,
+      "objective/train/weight_avg": 1.0016391277313232,
+      "objective/train/weighted_lm_loss": 1.374948501586914,
+      "objective/train/weights_max": 2.2131869792938232,
+      "objective/train/weights_min": 0.3746318221092224,
+      "theoretical_loss": 3.505723904877731,
+      "tokens_seen": 1703936000
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004885251163537154,
+      "loss": 0.0688,
+      "theoretical_loss": 3.505723904877731,
+      "tokens_seen": 1703936000
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004884448724121329,
+      "loss": 0.0669,
+      "theoretical_loss": 3.5056778877294983,
+      "tokens_seen": 1704198144
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004883646284705505,
+      "loss": 0.068,
+      "theoretical_loss": 3.5056318796408186,
+      "tokens_seen": 1704460288
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004882843845289681,
+      "loss": 0.0684,
+      "theoretical_loss": 3.505585880608515,
+      "tokens_seen": 1704722432
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048820414058738565,
+      "loss": 0.0682,
+      "theoretical_loss": 3.505539890629412,
+      "tokens_seen": 1704984576
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048812389664580326,
+      "loss": 0.0683,
+      "theoretical_loss": 3.5054939097003377,
+      "tokens_seen": 1705246720
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004880436527042209,
+      "loss": 0.0694,
+      "theoretical_loss": 3.505447937818119,
+      "tokens_seen": 1705508864
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048796340876263843,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5054019749795864,
+      "tokens_seen": 1705771008
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048788316482105605,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5053560211815715,
+      "tokens_seen": 1706033152
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004878029208794736,
+      "loss": 0.0683,
+      "theoretical_loss": 3.505310076420907,
+      "tokens_seen": 1706295296
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048772267693789116,
+      "loss": 0.0674,
+      "theoretical_loss": 3.505264140694428,
+      "tokens_seen": 1706557440
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004876424329963088,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5052182139989707,
+      "tokens_seen": 1706819584
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004875621890547264,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5051722963313723,
+      "tokens_seen": 1707081728
+    },
+    {
+      "epoch": 0.52,
+      "objective/train/advantage_avg": 0.0002717103052418679,
+      "objective/train/docs_used": 623332,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5116854906082153,
+      "objective/train/original_loss": 1.5116853713989258,
+      "objective/train/theoretical_loss": 3.5051493408820327,
+      "objective/train/tokens_used": 1727672800,
+      "objective/train/value_avg": -0.0070037841796875,
+      "objective/train/value_loss": 0.0002370732690906152,
+      "objective/train/value_max": -4.00543212890625e-05,
+      "objective/train/value_min": -0.56884765625,
+      "objective/train/value_reward_corr": 0.7653480297142256,
+      "objective/train/value_std": 0.017669677734375,
+      "objective/train/weight_avg": 1.0003830194473267,
+      "objective/train/weighted_lm_loss": 1.5120874643325806,
+      "objective/train/weights_max": 1.4940978288650513,
+      "objective/train/weights_min": 0.37812936305999756,
+      "theoretical_loss": 3.5051493408820327,
+      "tokens_seen": 1707212800
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.000487481945113144,
+      "loss": 0.0671,
+      "theoretical_loss": 3.505126387688473,
+      "tokens_seen": 1707343872
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048740170117156156,
+      "loss": 0.0655,
+      "theoretical_loss": 3.5050804880671134,
+      "tokens_seen": 1707606016
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048732145722997917,
+      "loss": 0.0678,
+      "theoretical_loss": 3.505034597464137,
+      "tokens_seen": 1707868160
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048724121328839673,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5049887158763866,
+      "tokens_seen": 1708130304
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004871609693468143,
+      "loss": 0.0659,
+      "theoretical_loss": 3.504942843300709,
+      "tokens_seen": 1708392448
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004870807254052319,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5048969797339513,
+      "tokens_seen": 1708654592
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004870004814636495,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5048511251729626,
+      "tokens_seen": 1708916736
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004869202375220671,
+      "loss": 0.068,
+      "theoretical_loss": 3.504805279614594,
+      "tokens_seen": 1709178880
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004868399935804847,
+      "loss": 0.0704,
+      "theoretical_loss": 3.504759443055696,
+      "tokens_seen": 1709441024
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004867597496389023,
+      "loss": 0.0678,
+      "theoretical_loss": 3.5047136154931238,
+      "tokens_seen": 1709703168
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048667950569731986,
+      "loss": 0.0695,
+      "theoretical_loss": 3.5046677969237328,
+      "tokens_seen": 1709965312
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004865992617557374,
+      "loss": 0.0678,
+      "theoretical_loss": 3.504621987344379,
+      "tokens_seen": 1710227456
+    },
+    {
+      "epoch": 0.52,
+      "objective/train/advantage_avg": 0.0009433527011424303,
+      "objective/train/docs_used": 624643,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4271879196166992,
+      "objective/train/original_loss": 1.4271876811981201,
+      "objective/train/theoretical_loss": 3.504576186751921,
+      "objective/train/tokens_used": 1730949600,
+      "objective/train/value_avg": -0.007293701171875,
+      "objective/train/value_loss": 0.00016305405006278306,
+      "objective/train/value_max": -3.7610530853271484e-05,
+      "objective/train/value_min": -0.338134765625,
+      "objective/train/value_reward_corr": 0.7030048069232187,
+      "objective/train/value_std": 0.01326751708984375,
+      "objective/train/weight_avg": 1.0010199546813965,
+      "objective/train/weighted_lm_loss": 1.428601861000061,
+      "objective/train/weights_max": 1.1191952228546143,
+      "objective/train/weights_min": 0.3719745874404907,
+      "theoretical_loss": 3.504576186751921,
+      "tokens_seen": 1710489600
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.000486519017814155,
+      "loss": 0.0694,
+      "theoretical_loss": 3.504576186751921,
+      "tokens_seen": 1710489600
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048643877387257264,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5045303951432194,
+      "tokens_seen": 1710751744
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004863585299309902,
+      "loss": 0.0681,
+      "theoretical_loss": 3.504484612515136,
+      "tokens_seen": 1711013888
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004862782859894078,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5044388388645333,
+      "tokens_seen": 1711276032
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004861980420478254,
+      "loss": 0.0651,
+      "theoretical_loss": 3.5043930741882763,
+      "tokens_seen": 1711538176
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048611779810624304,
+      "loss": 0.0695,
+      "theoretical_loss": 3.504347318483232,
+      "tokens_seen": 1711800320
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048603755416466054,
+      "loss": 0.0709,
+      "theoretical_loss": 3.504301571746267,
+      "tokens_seen": 1712062464
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048595731022307815,
+      "loss": 0.0693,
+      "theoretical_loss": 3.504255833974252,
+      "tokens_seen": 1712324608
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048587706628149576,
+      "loss": 0.0683,
+      "theoretical_loss": 3.5042101051640575,
+      "tokens_seen": 1712586752
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004857968223399133,
+      "loss": 0.0707,
+      "theoretical_loss": 3.5041643853125564,
+      "tokens_seen": 1712848896
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048571657839833094,
+      "loss": 0.0697,
+      "theoretical_loss": 3.504118674416623,
+      "tokens_seen": 1713111040
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048563633445674855,
+      "loss": 0.0706,
+      "theoretical_loss": 3.504072972473133,
+      "tokens_seen": 1713373184
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048555609051516616,
+      "loss": 0.0671,
+      "theoretical_loss": 3.504027279478964,
+      "tokens_seen": 1713635328
+    },
+    {
+      "epoch": 0.52,
+      "objective/train/advantage_avg": -0.0007473386940546334,
+      "objective/train/docs_used": 625897,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2894264459609985,
+      "objective/train/original_loss": 1.289426326751709,
+      "objective/train/theoretical_loss": 3.504004436336899,
+      "objective/train/tokens_used": 1734226400,
+      "objective/train/value_avg": -0.00562286376953125,
+      "objective/train/value_loss": 0.0001757982827257365,
+      "objective/train/value_max": -2.4497509002685547e-05,
+      "objective/train/value_min": -0.226318359375,
+      "objective/train/value_reward_corr": 0.6555531158316058,
+      "objective/train/value_std": 0.01038360595703125,
+      "objective/train/weight_avg": 0.9993323683738708,
+      "objective/train/weighted_lm_loss": 1.288981556892395,
+      "objective/train/weights_max": 1.1413358449935913,
+      "objective/train/weights_min": 0.3692460358142853,
+      "theoretical_loss": 3.504004436336899,
+      "tokens_seen": 1713766400
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048547584657358366,
+      "loss": 0.069,
+      "theoretical_loss": 3.5039815954309943,
+      "tokens_seen": 1713897472
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004853956026320013,
+      "loss": 0.0679,
+      "theoretical_loss": 3.5039359203261045,
+      "tokens_seen": 1714159616
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004853153586904189,
+      "loss": 0.0678,
+      "theoretical_loss": 3.5038902541611776,
+      "tokens_seen": 1714421760
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048523511474883645,
+      "loss": 0.0686,
+      "theoretical_loss": 3.503844596933096,
+      "tokens_seen": 1714683904
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048515487080725406,
+      "loss": 0.0709,
+      "theoretical_loss": 3.503798948638746,
+      "tokens_seen": 1714946048
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004850746268656717,
+      "loss": 0.0714,
+      "theoretical_loss": 3.503753309275013,
+      "tokens_seen": 1715208192
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004849943829240893,
+      "loss": 0.0719,
+      "theoretical_loss": 3.5037076788387864,
+      "tokens_seen": 1715470336
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048491413898250684,
+      "loss": 0.0689,
+      "theoretical_loss": 3.503662057326956,
+      "tokens_seen": 1715732480
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004848338950409244,
+      "loss": 0.0689,
+      "theoretical_loss": 3.503616444736412,
+      "tokens_seen": 1715994624
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.000484753651099342,
+      "loss": 0.0673,
+      "theoretical_loss": 3.5035708410640494,
+      "tokens_seen": 1716256768
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004846734071577596,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5035252463067614,
+      "tokens_seen": 1716518912
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004845931632161772,
+      "loss": 0.0685,
+      "theoretical_loss": 3.503479660461444,
+      "tokens_seen": 1716781056
+    },
+    {
+      "epoch": 0.52,
+      "objective/train/advantage_avg": 0.0008675092831254005,
+      "objective/train/docs_used": 627098,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3586034774780273,
+      "objective/train/original_loss": 1.3586034774780273,
+      "objective/train/theoretical_loss": 3.5034340835249953,
+      "objective/train/tokens_used": 1737503200,
+      "objective/train/value_avg": -0.00962066650390625,
+      "objective/train/value_loss": 0.00028026950894854963,
+      "objective/train/value_max": -8.821487426757812e-05,
+      "objective/train/value_min": -0.76806640625,
+      "objective/train/value_reward_corr": 0.7356101623856774,
+      "objective/train/value_std": 0.0185699462890625,
+      "objective/train/weight_avg": 1.0009956359863281,
+      "objective/train/weighted_lm_loss": 1.3600226640701294,
+      "objective/train/weights_max": 1.2380048036575317,
+      "objective/train/weights_min": 0.38121721148490906,
+      "theoretical_loss": 3.5034340835249953,
+      "tokens_seen": 1717043200
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004845129192745948,
+      "loss": 0.0663,
+      "theoretical_loss": 3.5034340835249953,
+      "tokens_seen": 1717043200
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048443267533301236,
+      "loss": 0.0684,
+      "theoretical_loss": 3.503388515494315,
+      "tokens_seen": 1717305344
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048435243139142997,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5033429563663026,
+      "tokens_seen": 1717567488
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048427218744984753,
+      "loss": 0.0709,
+      "theoretical_loss": 3.503297406137861,
+      "tokens_seen": 1717829632
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048419194350826514,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5032518648058946,
+      "tokens_seen": 1718091776
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004841116995666827,
+      "loss": 0.0664,
+      "theoretical_loss": 3.503206332367308,
+      "tokens_seen": 1718353920
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004840314556251003,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5031608088190085,
+      "tokens_seen": 1718616064
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004839512116835179,
+      "loss": 0.0674,
+      "theoretical_loss": 3.5031152941579045,
+      "tokens_seen": 1718878208
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004838709677419355,
+      "loss": 0.0674,
+      "theoretical_loss": 3.5030697883809063,
+      "tokens_seen": 1719140352
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004837907238003531,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5030242914849254,
+      "tokens_seen": 1719402496
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048371047985877065,
+      "loss": 0.0709,
+      "theoretical_loss": 3.502978803466875,
+      "tokens_seen": 1719664640
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048363023591718827,
+      "loss": 0.0705,
+      "theoretical_loss": 3.502933324323669,
+      "tokens_seen": 1719926784
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004835499919756058,
+      "loss": 0.0707,
+      "theoretical_loss": 3.502887854052225,
+      "tokens_seen": 1720188928
+    },
+    {
+      "epoch": 0.52,
+      "objective/train/advantage_avg": 0.0003047508653253317,
+      "objective/train/docs_used": 628251,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5686558485031128,
+      "objective/train/original_loss": 1.5686558485031128,
+      "objective/train/theoretical_loss": 3.50286512224245,
+      "objective/train/tokens_used": 1740780000,
+      "objective/train/value_avg": -0.00920867919921875,
+      "objective/train/value_loss": 0.0004943870007991791,
+      "objective/train/value_max": -4.57763671875e-05,
+      "objective/train/value_min": -0.671875,
+      "objective/train/value_reward_corr": 0.6343885890994714,
+      "objective/train/value_std": 0.01812744140625,
+      "objective/train/weight_avg": 1.0005214214324951,
+      "objective/train/weighted_lm_loss": 1.569042682647705,
+      "objective/train/weights_max": 1.6559826135635376,
+      "objective/train/weights_min": 0.3744032382965088,
+      "theoretical_loss": 3.50286512224245,
+      "tokens_seen": 1720320000
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048346974803402344,
+      "loss": 0.0689,
+      "theoretical_loss": 3.5028423926494594,
+      "tokens_seen": 1720451072
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048338950409244105,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5027969401122925,
+      "tokens_seen": 1720713216
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004833092601508586,
+      "loss": 0.0688,
+      "theoretical_loss": 3.5027514964376447,
+      "tokens_seen": 1720975360
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004832290162092762,
+      "loss": 0.0703,
+      "theoretical_loss": 3.502706061622438,
+      "tokens_seen": 1721237504
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048314877226769383,
+      "loss": 0.0687,
+      "theoretical_loss": 3.502660635663597,
+      "tokens_seen": 1721499648
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004830685283261114,
+      "loss": 0.0699,
+      "theoretical_loss": 3.5026152185580472,
+      "tokens_seen": 1721761792
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048298828438452895,
+      "loss": 0.0686,
+      "theoretical_loss": 3.502569810302715,
+      "tokens_seen": 1722023936
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048290804044294656,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5025244108945293,
+      "tokens_seen": 1722286080
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004828277965013642,
+      "loss": 0.0683,
+      "theoretical_loss": 3.50247902033042,
+      "tokens_seen": 1722548224
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048274755255978173,
+      "loss": 0.0704,
+      "theoretical_loss": 3.502433638607319,
+      "tokens_seen": 1722810368
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048266730861819935,
+      "loss": 0.0694,
+      "theoretical_loss": 3.502388265722159,
+      "tokens_seen": 1723072512
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048258706467661696,
+      "loss": 0.0698,
+      "theoretical_loss": 3.5023429016718755,
+      "tokens_seen": 1723334656
+    },
+    {
+      "epoch": 0.52,
+      "objective/train/advantage_avg": 0.00030591912218369544,
+      "objective/train/docs_used": 629447,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4868828058242798,
+      "objective/train/original_loss": 1.4868826866149902,
+      "objective/train/theoretical_loss": 3.502297546453403,
+      "objective/train/tokens_used": 1744056800,
+      "objective/train/value_avg": -0.00592803955078125,
+      "objective/train/value_loss": 0.00018052756786346436,
+      "objective/train/value_max": -3.427267074584961e-05,
+      "objective/train/value_min": -0.2147216796875,
+      "objective/train/value_reward_corr": 0.647845852948392,
+      "objective/train/value_std": 0.01120758056640625,
+      "objective/train/weight_avg": 1.0003873109817505,
+      "objective/train/weighted_lm_loss": 1.4872801303863525,
+      "objective/train/weights_max": 1.1227530241012573,
+      "objective/train/weights_min": 0.3682018518447876,
+      "theoretical_loss": 3.502297546453403,
+      "tokens_seen": 1723596800
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048250682073503446,
+      "loss": 0.0704,
+      "theoretical_loss": 3.502297546453403,
+      "tokens_seen": 1723596800
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004824265767934521,
+      "loss": 0.0682,
+      "theoretical_loss": 3.5022522000636807,
+      "tokens_seen": 1723858944
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004823463328518697,
+      "loss": 0.0702,
+      "theoretical_loss": 3.502206862499647,
+      "tokens_seen": 1724121088
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004822660889102873,
+      "loss": 0.0684,
+      "theoretical_loss": 3.5021615337582435,
+      "tokens_seen": 1724383232
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048218584496870486,
+      "loss": 0.0686,
+      "theoretical_loss": 3.5021162138364117,
+      "tokens_seen": 1724645376
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004821056010271225,
+      "loss": 0.0692,
+      "theoretical_loss": 3.502070902731096,
+      "tokens_seen": 1724907520
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004820253570855401,
+      "loss": 0.0714,
+      "theoretical_loss": 3.502025600439241,
+      "tokens_seen": 1725169664
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004819451131439576,
+      "loss": 0.0717,
+      "theoretical_loss": 3.5019803069577944,
+      "tokens_seen": 1725431808
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004818648692023752,
+      "loss": 0.0691,
+      "theoretical_loss": 3.5019350222837042,
+      "tokens_seen": 1725693952
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004817846252607928,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5018897464139203,
+      "tokens_seen": 1725956096
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048170438131921043,
+      "loss": 0.0719,
+      "theoretical_loss": 3.501844479345394,
+      "tokens_seen": 1726218240
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.000481624137377628,
+      "loss": 0.069,
+      "theoretical_loss": 3.5017992210750783,
+      "tokens_seen": 1726480384
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004815438934360456,
+      "loss": 0.0716,
+      "theoretical_loss": 3.5017539715999275,
+      "tokens_seen": 1726742528
+    },
+    {
+      "epoch": 0.52,
+      "objective/train/advantage_avg": 0.0002910619368776679,
+      "objective/train/docs_used": 630619,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4500676393508911,
+      "objective/train/original_loss": 1.4500675201416016,
+      "objective/train/theoretical_loss": 3.501731350159588,
+      "objective/train/tokens_used": 1747333600,
+      "objective/train/value_avg": -0.01136016845703125,
+      "objective/train/value_loss": 0.0006346271256916225,
+      "objective/train/value_max": -3.0219554901123047e-05,
+      "objective/train/value_min": -0.8798828125,
+      "objective/train/value_reward_corr": 0.7903729671685932,
+      "objective/train/value_std": 0.0281829833984375,
+      "objective/train/weight_avg": 1.0005794763565063,
+      "objective/train/weighted_lm_loss": 1.4503711462020874,
+      "objective/train/weights_max": 2.2934980392456055,
+      "objective/train/weights_min": 0.3900670111179352,
+      "theoretical_loss": 3.501731350159588,
+      "tokens_seen": 1726873600
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004814636494944632,
+      "loss": 0.0696,
+      "theoretical_loss": 3.5017087309168975,
+      "tokens_seen": 1727004672
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048138340555288077,
+      "loss": 0.0707,
+      "theoretical_loss": 3.501663499022947,
+      "tokens_seen": 1727266816
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048130316161129833,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5016182759150336,
+      "tokens_seen": 1727528960
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048122291766971594,
+      "loss": 0.0688,
+      "theoretical_loss": 3.501573061590118,
+      "tokens_seen": 1727791104
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048114267372813355,
+      "loss": 0.0678,
+      "theoretical_loss": 3.5015278560451626,
+      "tokens_seen": 1728053248
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004810624297865511,
+      "loss": 0.069,
+      "theoretical_loss": 3.501482659277131,
+      "tokens_seen": 1728315392
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004809821858449687,
+      "loss": 0.0697,
+      "theoretical_loss": 3.501437471282988,
+      "tokens_seen": 1728577536
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048090194190338634,
+      "loss": 0.0669,
+      "theoretical_loss": 3.5013922920597,
+      "tokens_seen": 1728839680
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004808216979618039,
+      "loss": 0.0687,
+      "theoretical_loss": 3.5013471216042356,
+      "tokens_seen": 1729101824
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048074145402022145,
+      "loss": 0.0665,
+      "theoretical_loss": 3.5013019599135635,
+      "tokens_seen": 1729363968
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048066121007863907,
+      "loss": 0.0698,
+      "theoretical_loss": 3.501256806984656,
+      "tokens_seen": 1729626112
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004805809661370567,
+      "loss": 0.0697,
+      "theoretical_loss": 3.5012116628144847,
+      "tokens_seen": 1729888256
+    },
+    {
+      "epoch": 0.52,
+      "objective/train/advantage_avg": -0.0012186147505417466,
+      "objective/train/docs_used": 631803,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4452341794967651,
+      "objective/train/original_loss": 1.4452341794967651,
+      "objective/train/theoretical_loss": 3.5011665274000245,
+      "objective/train/tokens_used": 1750610400,
+      "objective/train/value_avg": -0.00994110107421875,
+      "objective/train/value_loss": 0.0008707749075256288,
+      "objective/train/value_max": -5.692243576049805e-05,
+      "objective/train/value_min": -0.7236328125,
+      "objective/train/value_reward_corr": 0.7037228321696455,
+      "objective/train/value_std": 0.0241546630859375,
+      "objective/train/weight_avg": 0.9991577863693237,
+      "objective/train/weighted_lm_loss": 1.4447067975997925,
+      "objective/train/weights_max": 1.752554178237915,
+      "objective/train/weights_min": 0.371359258890152,
+      "theoretical_loss": 3.5011665274000245,
+      "tokens_seen": 1730150400
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048050072219547424,
+      "loss": 0.0666,
+      "theoretical_loss": 3.5011665274000245,
+      "tokens_seen": 1730150400
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048042047825389185,
+      "loss": 0.072,
+      "theoretical_loss": 3.50112140073825,
+      "tokens_seen": 1730412544
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048034023431230946,
+      "loss": 0.0693,
+      "theoretical_loss": 3.501076282826139,
+      "tokens_seen": 1730674688
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.000480259990370727,
+      "loss": 0.0694,
+      "theoretical_loss": 3.5010311736606696,
+      "tokens_seen": 1730936832
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048017974642914463,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5009860732388223,
+      "tokens_seen": 1731198976
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.0004800995024875622,
+      "loss": 0.0705,
+      "theoretical_loss": 3.5009409815575787,
+      "tokens_seen": 1731461120
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00048001925854597975,
+      "loss": 0.0678,
+      "theoretical_loss": 3.500895898613922,
+      "tokens_seen": 1731723264
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.00047993901460439736,
+      "loss": 0.0709,
+      "theoretical_loss": 3.5008508244048366,
+      "tokens_seen": 1731985408
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 0.000479858770662815,
+      "loss": 0.0688,
+      "theoretical_loss": 3.5008057589273083,
+      "tokens_seen": 1732247552
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004797785267212326,
+      "loss": 0.0674,
+      "theoretical_loss": 3.500760702178325,
+      "tokens_seen": 1732509696
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047969828277965015,
+      "loss": 0.0655,
+      "theoretical_loss": 3.5007156541548756,
+      "tokens_seen": 1732771840
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047961803883806776,
+      "loss": 0.0708,
+      "theoretical_loss": 3.5006706148539513,
+      "tokens_seen": 1733033984
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004795377948964853,
+      "loss": 0.0716,
+      "theoretical_loss": 3.500625584272543,
+      "tokens_seen": 1733296128
+    },
+    {
+      "epoch": 0.53,
+      "objective/train/advantage_avg": 0.0006020920118317008,
+      "objective/train/docs_used": 632954,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3608434200286865,
+      "objective/train/original_loss": 1.360843539237976,
+      "objective/train/theoretical_loss": 3.500603072250718,
+      "objective/train/tokens_used": 1753887200,
+      "objective/train/value_avg": -0.007480621337890625,
+      "objective/train/value_loss": 0.0004119122459087521,
+      "objective/train/value_max": -6.556510925292969e-05,
+      "objective/train/value_min": -0.68798828125,
+      "objective/train/value_reward_corr": 0.6353150839836534,
+      "objective/train/value_std": 0.0165557861328125,
+      "objective/train/weight_avg": 1.0007774829864502,
+      "objective/train/weighted_lm_loss": 1.3612209558486938,
+      "objective/train/weights_max": 1.4956903457641602,
+      "objective/train/weights_min": 0.36957859992980957,
+      "theoretical_loss": 3.500603072250718,
+      "tokens_seen": 1733427200
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004794575509549029,
+      "loss": 0.0681,
+      "theoretical_loss": 3.5005805624076456,
+      "tokens_seen": 1733558272
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004793773070133205,
+      "loss": 0.0688,
+      "theoretical_loss": 3.500535549256253,
+      "tokens_seen": 1733820416
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004792970630717381,
+      "loss": 0.0672,
+      "theoretical_loss": 3.5004905448153627,
+      "tokens_seen": 1734082560
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004792168191301557,
+      "loss": 0.0705,
+      "theoretical_loss": 3.500445549081972,
+      "tokens_seen": 1734344704
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047913657518857327,
+      "loss": 0.069,
+      "theoretical_loss": 3.500400562053081,
+      "tokens_seen": 1734606848
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004790563312469909,
+      "loss": 0.0698,
+      "theoretical_loss": 3.50035558372569,
+      "tokens_seen": 1734868992
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047897608730540844,
+      "loss": 0.0711,
+      "theoretical_loss": 3.5003106140968026,
+      "tokens_seen": 1735131136
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.000478895843363826,
+      "loss": 0.0701,
+      "theoretical_loss": 3.5002656531634213,
+      "tokens_seen": 1735393280
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004788155994222436,
+      "loss": 0.0693,
+      "theoretical_loss": 3.5002207009225526,
+      "tokens_seen": 1735655424
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004787353554806612,
+      "loss": 0.0685,
+      "theoretical_loss": 3.5001757573712036,
+      "tokens_seen": 1735917568
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047865511153907884,
+      "loss": 0.0702,
+      "theoretical_loss": 3.5001308225063816,
+      "tokens_seen": 1736179712
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004785748675974964,
+      "loss": 0.0689,
+      "theoretical_loss": 3.500085896325098,
+      "tokens_seen": 1736441856
+    },
+    {
+      "epoch": 0.53,
+      "objective/train/advantage_avg": -0.00023456213239114732,
+      "objective/train/docs_used": 634251,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4015569686889648,
+      "objective/train/original_loss": 1.401557207107544,
+      "objective/train/theoretical_loss": 3.5000409788243623,
+      "objective/train/tokens_used": 1757164000,
+      "objective/train/value_avg": -0.006313323974609375,
+      "objective/train/value_loss": 0.00014978312538005412,
+      "objective/train/value_max": -6.812810897827148e-05,
+      "objective/train/value_min": -0.3037109375,
+      "objective/train/value_reward_corr": 0.7739671798839213,
+      "objective/train/value_std": 0.01079559326171875,
+      "objective/train/weight_avg": 0.9998375773429871,
+      "objective/train/weighted_lm_loss": 1.4020917415618896,
+      "objective/train/weights_max": 1.1075072288513184,
+      "objective/train/weights_min": 0.6098971962928772,
+      "theoretical_loss": 3.5000409788243623,
+      "tokens_seen": 1736704000
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.000478494623655914,
+      "loss": 0.069,
+      "theoretical_loss": 3.5000409788243623,
+      "tokens_seen": 1736704000
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004784143797143316,
+      "loss": 0.0706,
+      "theoretical_loss": 3.4999960700011896,
+      "tokens_seen": 1736966144
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047833413577274913,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4999511698525927,
+      "tokens_seen": 1737228288
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047825389183116674,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4999062783755877,
+      "tokens_seen": 1737490432
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047817364788958435,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4998613955671924,
+      "tokens_seen": 1737752576
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004780934039480019,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4998165214244246,
+      "tokens_seen": 1738014720
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004780131600064195,
+      "loss": 0.0695,
+      "theoretical_loss": 3.499771655944306,
+      "tokens_seen": 1738276864
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047793291606483714,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4997267991238568,
+      "tokens_seen": 1738539008
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047785267212325475,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4996819509601016,
+      "tokens_seen": 1738801152
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047777242818167225,
+      "loss": 0.0672,
+      "theoretical_loss": 3.499637111450064,
+      "tokens_seen": 1739063296
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047769218424008987,
+      "loss": 0.0708,
+      "theoretical_loss": 3.499592280590771,
+      "tokens_seen": 1739325440
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004776119402985075,
+      "loss": 0.0711,
+      "theoretical_loss": 3.499547458379249,
+      "tokens_seen": 1739587584
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047753169635692504,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4995026448125284,
+      "tokens_seen": 1739849728
+    },
+    {
+      "epoch": 0.53,
+      "objective/train/advantage_avg": 0.000589077128097415,
+      "objective/train/docs_used": 635535,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3270041942596436,
+      "objective/train/original_loss": 1.327004313468933,
+      "objective/train/theoretical_loss": 3.49948024127004,
+      "objective/train/tokens_used": 1760440800,
+      "objective/train/value_avg": -0.005573272705078125,
+      "objective/train/value_loss": 0.00010660909174475819,
+      "objective/train/value_max": -6.657838821411133e-05,
+      "objective/train/value_min": -0.2166748046875,
+      "objective/train/value_reward_corr": 0.6834865549309408,
+      "objective/train/value_std": 0.00957489013671875,
+      "objective/train/weight_avg": 1.000641107559204,
+      "objective/train/weighted_lm_loss": 1.3277722597122192,
+      "objective/train/weights_max": 1.1486068964004517,
+      "objective/train/weights_min": 0.7876083850860596,
+      "theoretical_loss": 3.49948024127004,
+      "tokens_seen": 1739980800
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047745145241534265,
+      "loss": 0.067,
+      "theoretical_loss": 3.499457839887639,
+      "tokens_seen": 1740111872
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047737120847376026,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4994130436016126,
+      "tokens_seen": 1740374016
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004772909645321779,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4993682559514836,
+      "tokens_seen": 1740636160
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004772107205905954,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4993234769342862,
+      "tokens_seen": 1740898304
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.000477130476649013,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4992787065470567,
+      "tokens_seen": 1741160448
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004770502327074306,
+      "loss": 0.0676,
+      "theoretical_loss": 3.499233944786834,
+      "tokens_seen": 1741422592
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047696998876584816,
+      "loss": 0.0679,
+      "theoretical_loss": 3.499189191650656,
+      "tokens_seen": 1741684736
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004768897448242658,
+      "loss": 0.0683,
+      "theoretical_loss": 3.499144447135565,
+      "tokens_seen": 1741946880
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004768095008826834,
+      "loss": 0.0715,
+      "theoretical_loss": 3.499099711238602,
+      "tokens_seen": 1742209024
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.000476729256941101,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4990549839568112,
+      "tokens_seen": 1742471168
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047664901299951856,
+      "loss": 0.0704,
+      "theoretical_loss": 3.4990102652872377,
+      "tokens_seen": 1742733312
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004765687690579361,
+      "loss": 0.0712,
+      "theoretical_loss": 3.4989655552269285,
+      "tokens_seen": 1742995456
+    },
+    {
+      "epoch": 0.53,
+      "objective/train/advantage_avg": 0.0008427195716649294,
+      "objective/train/docs_used": 636817,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3889329433441162,
+      "objective/train/original_loss": 1.3889329433441162,
+      "objective/train/theoretical_loss": 3.4989208537729315,
+      "objective/train/tokens_used": 1763717600,
+      "objective/train/value_avg": -0.004062652587890625,
+      "objective/train/value_loss": 0.00011055525101255625,
+      "objective/train/value_max": -4.5418739318847656e-05,
+      "objective/train/value_min": -0.203125,
+      "objective/train/value_reward_corr": 0.4945603011219797,
+      "objective/train/value_std": 0.006389617919921875,
+      "objective/train/weight_avg": 1.0008906126022339,
+      "objective/train/weighted_lm_loss": 1.3909051418304443,
+      "objective/train/weights_max": 1.1333805322647095,
+      "objective/train/weights_min": 0.3810659945011139,
+      "theoretical_loss": 3.4989208537729315,
+      "tokens_seen": 1743257600
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047648852511635373,
+      "loss": 0.0711,
+      "theoretical_loss": 3.4989208537729315,
+      "tokens_seen": 1743257600
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004764082811747713,
+      "loss": 0.0704,
+      "theoretical_loss": 3.498876160922296,
+      "tokens_seen": 1743519744
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004763280372331889,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4988314766720734,
+      "tokens_seen": 1743781888
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004762477932916065,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4987868010193157,
+      "tokens_seen": 1744044032
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047616754935002407,
+      "loss": 0.0646,
+      "theoretical_loss": 3.4987421339610774,
+      "tokens_seen": 1744306176
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004760873054084417,
+      "loss": 0.0683,
+      "theoretical_loss": 3.498697475494413,
+      "tokens_seen": 1744568320
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047600706146685924,
+      "loss": 0.0693,
+      "theoretical_loss": 3.49865282561638,
+      "tokens_seen": 1744830464
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047592681752527686,
+      "loss": 0.0702,
+      "theoretical_loss": 3.498608184324037,
+      "tokens_seen": 1745092608
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004758465735836944,
+      "loss": 0.0698,
+      "theoretical_loss": 3.498563551614443,
+      "tokens_seen": 1745354752
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.000475766329642112,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4985189274846586,
+      "tokens_seen": 1745616896
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047568608570052964,
+      "loss": 0.0712,
+      "theoretical_loss": 3.498474311931748,
+      "tokens_seen": 1745879040
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004756058417589472,
+      "loss": 0.0711,
+      "theoretical_loss": 3.4984297049527737,
+      "tokens_seen": 1746141184
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004755255978173648,
+      "loss": 0.0695,
+      "theoretical_loss": 3.498385106544802,
+      "tokens_seen": 1746403328
+    },
+    {
+      "epoch": 0.53,
+      "objective/train/advantage_avg": 0.0013079342897981405,
+      "objective/train/docs_used": 637932,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4305325746536255,
+      "objective/train/original_loss": 1.4305325746536255,
+      "objective/train/theoretical_loss": 3.498362810554026,
+      "objective/train/tokens_used": 1766994400,
+      "objective/train/value_avg": -0.00585174560546875,
+      "objective/train/value_loss": 0.0001362788025289774,
+      "objective/train/value_max": -3.349781036376953e-05,
+      "objective/train/value_min": -0.398681640625,
+      "objective/train/value_reward_corr": 0.6356918638700473,
+      "objective/train/value_std": 0.01058197021484375,
+      "objective/train/weight_avg": 1.0013703107833862,
+      "objective/train/weighted_lm_loss": 1.4319835901260376,
+      "objective/train/weights_max": 1.1282485723495483,
+      "objective/train/weights_min": 0.36817076802253723,
+      "theoretical_loss": 3.498362810554026,
+      "tokens_seen": 1746534400
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004754453538757824,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4983405167049004,
+      "tokens_seen": 1746665472
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004753651099342,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4982959354301366,
+      "tokens_seen": 1746927616
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047528486599261754,
+      "loss": 0.0651,
+      "theoretical_loss": 3.49825136271758,
+      "tokens_seen": 1747189760
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047520462205103515,
+      "loss": 0.0704,
+      "theoretical_loss": 3.498206798564303,
+      "tokens_seen": 1747451904
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047512437810945276,
+      "loss": 0.0687,
+      "theoretical_loss": 3.498162242967377,
+      "tokens_seen": 1747714048
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004750441341678703,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4981176959238773,
+      "tokens_seen": 1747976192
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047496389022628794,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4980731574308788,
+      "tokens_seen": 1748238336
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047488364628470555,
+      "loss": 0.0657,
+      "theoretical_loss": 3.498028627485459,
+      "tokens_seen": 1748500480
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004748034023431231,
+      "loss": 0.0686,
+      "theoretical_loss": 3.497984106084696,
+      "tokens_seen": 1748762624
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047472315840154066,
+      "loss": 0.0715,
+      "theoretical_loss": 3.4979395932256696,
+      "tokens_seen": 1749024768
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004746429144599583,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4978950889054614,
+      "tokens_seen": 1749286912
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004745626705183759,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4978505931211545,
+      "tokens_seen": 1749549056
+    },
+    {
+      "epoch": 0.53,
+      "objective/train/advantage_avg": -0.0006571173435077071,
+      "objective/train/docs_used": 639149,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3665317296981812,
+      "objective/train/original_loss": 1.3665317296981812,
+      "objective/train/theoretical_loss": 3.4978061058698326,
+      "objective/train/tokens_used": 1770271200,
+      "objective/train/value_avg": -0.007808685302734375,
+      "objective/train/value_loss": 0.00021528959041461349,
+      "objective/train/value_max": -5.7816505432128906e-05,
+      "objective/train/value_min": -0.264404296875,
+      "objective/train/value_reward_corr": 0.8101872704556123,
+      "objective/train/value_std": 0.016448974609375,
+      "objective/train/weight_avg": 0.999443769454956,
+      "objective/train/weighted_lm_loss": 1.3659167289733887,
+      "objective/train/weights_max": 1.2142077684402466,
+      "objective/train/weights_min": 0.3840899169445038,
+      "theoretical_loss": 3.4978061058698326,
+      "tokens_seen": 1749811200
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047448242657679345,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4978061058698326,
+      "tokens_seen": 1749811200
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047440218263521106,
+      "loss": 0.0721,
+      "theoretical_loss": 3.4977616271485816,
+      "tokens_seen": 1750073344
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004743219386936287,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4977171569544883,
+      "tokens_seen": 1750335488
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004742416947520462,
+      "loss": 0.0671,
+      "theoretical_loss": 3.497672695284641,
+      "tokens_seen": 1750597632
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004741614508104638,
+      "loss": 0.0683,
+      "theoretical_loss": 3.497628242136131,
+      "tokens_seen": 1750859776
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004740812068688814,
+      "loss": 0.0675,
+      "theoretical_loss": 3.497583797506048,
+      "tokens_seen": 1751121920
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.000474000962927299,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4975393613914854,
+      "tokens_seen": 1751384064
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004739207189857166,
+      "loss": 0.0691,
+      "theoretical_loss": 3.497494933789538,
+      "tokens_seen": 1751646208
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004738404750441342,
+      "loss": 0.071,
+      "theoretical_loss": 3.4974505146973005,
+      "tokens_seen": 1751908352
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004737602311025518,
+      "loss": 0.072,
+      "theoretical_loss": 3.497406104111871,
+      "tokens_seen": 1752170496
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047367998716096936,
+      "loss": 0.071,
+      "theoretical_loss": 3.4973617020303465,
+      "tokens_seen": 1752432640
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004735997432193869,
+      "loss": 0.0704,
+      "theoretical_loss": 3.497317308449828,
+      "tokens_seen": 1752694784
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047351949927780453,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4972729233674174,
+      "tokens_seen": 1752956928
+    },
+    {
+      "epoch": 0.53,
+      "objective/train/advantage_avg": 0.0008459224482066929,
+      "objective/train/docs_used": 640338,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3503456115722656,
+      "objective/train/original_loss": 1.3503456115722656,
+      "objective/train/theoretical_loss": 3.4972507340120966,
+      "objective/train/tokens_used": 1773548000,
+      "objective/train/value_avg": -0.00661468505859375,
+      "objective/train/value_loss": 0.0001745822373777628,
+      "objective/train/value_max": -3.9458274841308594e-05,
+      "objective/train/value_min": -0.2017822265625,
+      "objective/train/value_reward_corr": 0.8029245560216447,
+      "objective/train/value_std": 0.01456451416015625,
+      "objective/train/weight_avg": 1.0009305477142334,
+      "objective/train/weighted_lm_loss": 1.351764440536499,
+      "objective/train/weights_max": 1.1526963710784912,
+      "objective/train/weights_min": 0.719582736492157,
+      "theoretical_loss": 3.4972507340120966,
+      "tokens_seen": 1753088000
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047343925533622214,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4972285467802164,
+      "tokens_seen": 1753219072
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004733590113946397,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4971841786853295,
+      "tokens_seen": 1753481216
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004732787674530573,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4971398190798615,
+      "tokens_seen": 1753743360
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004731985235114749,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4970954679609214,
+      "tokens_seen": 1754005504
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004731182795698925,
+      "loss": 0.0725,
+      "theoretical_loss": 3.4970511253256156,
+      "tokens_seen": 1754267648
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047303803562831004,
+      "loss": 0.0719,
+      "theoretical_loss": 3.4970067911710556,
+      "tokens_seen": 1754529792
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047295779168672765,
+      "loss": 0.0727,
+      "theoretical_loss": 3.4969624654943514,
+      "tokens_seen": 1754791936
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047287754774514527,
+      "loss": 0.0695,
+      "theoretical_loss": 3.496918148292616,
+      "tokens_seen": 1755054080
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004727973038035628,
+      "loss": 0.07,
+      "theoretical_loss": 3.496873839562964,
+      "tokens_seen": 1755316224
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047271705986198044,
+      "loss": 0.0693,
+      "theoretical_loss": 3.49682953930251,
+      "tokens_seen": 1755578368
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047263681592039805,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4967852475083725,
+      "tokens_seen": 1755840512
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004725565719788156,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4967409641776683,
+      "tokens_seen": 1756102656
+    },
+    {
+      "epoch": 0.53,
+      "objective/train/advantage_avg": 0.0014478386146947742,
+      "objective/train/docs_used": 641471,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3184295892715454,
+      "objective/train/original_loss": 1.3184294700622559,
+      "objective/train/theoretical_loss": 3.4966966893075178,
+      "objective/train/tokens_used": 1776824800,
+      "objective/train/value_avg": -0.0074462890625,
+      "objective/train/value_loss": 0.0001307493366766721,
+      "objective/train/value_max": -3.594160079956055e-05,
+      "objective/train/value_min": -0.240234375,
+      "objective/train/value_reward_corr": 0.7224784018460088,
+      "objective/train/value_std": 0.01288604736328125,
+      "objective/train/weight_avg": 1.001512050628662,
+      "objective/train/weighted_lm_loss": 1.3203070163726807,
+      "objective/train/weights_max": 1.2715470790863037,
+      "objective/train/weights_min": 0.6124921441078186,
+      "theoretical_loss": 3.4966966893075178,
+      "tokens_seen": 1756364800
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047247632803723317,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4966966893075178,
+      "tokens_seen": 1756364800
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004723960840956508,
+      "loss": 0.0704,
+      "theoretical_loss": 3.496652422895042,
+      "tokens_seen": 1756626944
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004723158401540684,
+      "loss": 0.069,
+      "theoretical_loss": 3.496608164937364,
+      "tokens_seen": 1756889088
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047223559621248595,
+      "loss": 0.0705,
+      "theoretical_loss": 3.496563915431607,
+      "tokens_seen": 1757151232
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047215535227090356,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4965196743748965,
+      "tokens_seen": 1757413376
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004720751083293212,
+      "loss": 0.0709,
+      "theoretical_loss": 3.49647544176436,
+      "tokens_seen": 1757675520
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047199486438773873,
+      "loss": 0.0729,
+      "theoretical_loss": 3.4964312175971246,
+      "tokens_seen": 1757937664
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047191462044615635,
+      "loss": 0.072,
+      "theoretical_loss": 3.496387001870321,
+      "tokens_seen": 1758199808
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004718343765045739,
+      "loss": 0.0706,
+      "theoretical_loss": 3.49634279458108,
+      "tokens_seen": 1758461952
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047175413256299146,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4962985957265333,
+      "tokens_seen": 1758724096
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004716738886214091,
+      "loss": 0.0681,
+      "theoretical_loss": 3.496254405303815,
+      "tokens_seen": 1758986240
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004715936446798267,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4962102233100607,
+      "tokens_seen": 1759248384
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004715134007382443,
+      "loss": 0.0728,
+      "theoretical_loss": 3.4961660497424063,
+      "tokens_seen": 1759510528
+    },
+    {
+      "epoch": 0.53,
+      "objective/train/advantage_avg": 0.0015444692689925432,
+      "objective/train/docs_used": 642679,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3258466720581055,
+      "objective/train/original_loss": 1.3258464336395264,
+      "objective/train/theoretical_loss": 3.4961439661174727,
+      "objective/train/tokens_used": 1780101600,
+      "objective/train/value_avg": -0.007198333740234375,
+      "objective/train/value_loss": 0.0004468359693419188,
+      "objective/train/value_max": -6.604194641113281e-05,
+      "objective/train/value_min": -0.8720703125,
+      "objective/train/value_reward_corr": 0.6289400413877285,
+      "objective/train/value_std": 0.0161285400390625,
+      "objective/train/weight_avg": 1.0017292499542236,
+      "objective/train/weighted_lm_loss": 1.3280916213989258,
+      "objective/train/weights_max": 1.299414873123169,
+      "objective/train/weights_min": 0.36901721358299255,
+      "theoretical_loss": 3.4961439661174727,
+      "tokens_seen": 1759641600
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047143315679666186,
+      "loss": 0.0701,
+      "theoretical_loss": 3.496121884597991,
+      "tokens_seen": 1759772672
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004713529128550795,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4960777278739528,
+      "tokens_seen": 1760034816
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047127266891349703,
+      "loss": 0.0721,
+      "theoretical_loss": 3.4960335795674338,
+      "tokens_seen": 1760296960
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004711924249719146,
+      "loss": 0.0703,
+      "theoretical_loss": 3.495989439675575,
+      "tokens_seen": 1760559104
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004711121810303322,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4959453081955205,
+      "tokens_seen": 1760821248
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004710319370887498,
+      "loss": 0.0688,
+      "theoretical_loss": 3.495901185124416,
+      "tokens_seen": 1761083392
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047095169314716743,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4958570704594067,
+      "tokens_seen": 1761345536
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.000470871449205585,
+      "loss": 0.0685,
+      "theoretical_loss": 3.495812964197641,
+      "tokens_seen": 1761607680
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004707912052640026,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4957688663362685,
+      "tokens_seen": 1761869824
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004707109613224202,
+      "loss": 0.0716,
+      "theoretical_loss": 3.495724776872439,
+      "tokens_seen": 1762131968
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004706307173808377,
+      "loss": 0.0709,
+      "theoretical_loss": 3.4956806958033044,
+      "tokens_seen": 1762394112
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047055047343925533,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4956366231260185,
+      "tokens_seen": 1762656256
+    },
+    {
+      "epoch": 0.53,
+      "objective/train/advantage_avg": 0.000993677182123065,
+      "objective/train/docs_used": 643830,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.203426718711853,
+      "objective/train/original_loss": 1.2034268379211426,
+      "objective/train/theoretical_loss": 3.4955925588377363,
+      "objective/train/tokens_used": 1783378400,
+      "objective/train/value_avg": -0.004913330078125,
+      "objective/train/value_loss": 0.00017356479656882584,
+      "objective/train/value_max": -5.1021575927734375e-05,
+      "objective/train/value_min": -0.369384765625,
+      "objective/train/value_reward_corr": 0.572270645143692,
+      "objective/train/value_std": 0.009613037109375,
+      "objective/train/weight_avg": 1.0010701417922974,
+      "objective/train/weighted_lm_loss": 1.20392644405365,
+      "objective/train/weights_max": 1.2256743907928467,
+      "objective/train/weights_min": 0.3710845410823822,
+      "theoretical_loss": 3.4955925588377363,
+      "tokens_seen": 1762918400
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047047022949767294,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4955925588377363,
+      "tokens_seen": 1762918400
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047038998555609055,
+      "loss": 0.0729,
+      "theoretical_loss": 3.4955485029356135,
+      "tokens_seen": 1763180544
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004703097416145081,
+      "loss": 0.0693,
+      "theoretical_loss": 3.495504455416807,
+      "tokens_seen": 1763442688
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004702294976729257,
+      "loss": 0.0674,
+      "theoretical_loss": 3.495460416278477,
+      "tokens_seen": 1763704832
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047014925373134334,
+      "loss": 0.069,
+      "theoretical_loss": 3.4954163855177827,
+      "tokens_seen": 1763966976
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00047006900978976084,
+      "loss": 0.0701,
+      "theoretical_loss": 3.495372363131886,
+      "tokens_seen": 1764229120
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00046998876584817845,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4953283491179503,
+      "tokens_seen": 1764491264
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00046990852190659607,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4952843434731395,
+      "tokens_seen": 1764753408
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.0004698282779650136,
+      "loss": 0.0698,
+      "theoretical_loss": 3.49524034619462,
+      "tokens_seen": 1765015552
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 0.00046974803402343124,
+      "loss": 0.0717,
+      "theoretical_loss": 3.495196357279559,
+      "tokens_seen": 1765277696
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046966779008184885,
+      "loss": 0.0715,
+      "theoretical_loss": 3.495152376725124,
+      "tokens_seen": 1765539840
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046958754614026646,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4951084045284864,
+      "tokens_seen": 1765801984
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046950730219868397,
+      "loss": 0.0698,
+      "theoretical_loss": 3.495064440686816,
+      "tokens_seen": 1766064128
+    },
+    {
+      "epoch": 0.54,
+      "objective/train/advantage_avg": 0.000716030306648463,
+      "objective/train/docs_used": 644977,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5920703411102295,
+      "objective/train/original_loss": 1.5920703411102295,
+      "objective/train/theoretical_loss": 3.495042461898211,
+      "objective/train/tokens_used": 1786655200,
+      "objective/train/value_avg": -0.00447845458984375,
+      "objective/train/value_loss": 0.00016367589705623686,
+      "objective/train/value_max": -5.3048133850097656e-05,
+      "objective/train/value_min": -0.255615234375,
+      "objective/train/value_reward_corr": 0.5714298891637101,
+      "objective/train/value_std": 0.00833892822265625,
+      "objective/train/weight_avg": 1.0007867813110352,
+      "objective/train/weighted_lm_loss": 1.5934929847717285,
+      "objective/train/weights_max": 1.171118140220642,
+      "objective/train/weights_min": 0.36817076802253723,
+      "theoretical_loss": 3.495042461898211,
+      "tokens_seen": 1766195200
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004694270582571016,
+      "loss": 0.0699,
+      "theoretical_loss": 3.495020485197287,
+      "tokens_seen": 1766326272
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004693468143155192,
+      "loss": 0.0701,
+      "theoretical_loss": 3.494976538057073,
+      "tokens_seen": 1766588416
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046926657037393675,
+      "loss": 0.071,
+      "theoretical_loss": 3.4949325992633486,
+      "tokens_seen": 1766850560
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046918632643235436,
+      "loss": 0.0706,
+      "theoretical_loss": 3.4948886688132923,
+      "tokens_seen": 1767112704
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.000469106082490772,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4948447467040804,
+      "tokens_seen": 1767374848
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004690258385491896,
+      "loss": 0.0695,
+      "theoretical_loss": 3.494800832932894,
+      "tokens_seen": 1767636992
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046894559460760715,
+      "loss": 0.0663,
+      "theoretical_loss": 3.494756927496913,
+      "tokens_seen": 1767899136
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004688653506660247,
+      "loss": 0.0684,
+      "theoretical_loss": 3.49471303039332,
+      "tokens_seen": 1768161280
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004687851067244423,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4946691416192985,
+      "tokens_seen": 1768423424
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004687048627828599,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4946252611720348,
+      "tokens_seen": 1768685568
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004686246188412775,
+      "loss": 0.0706,
+      "theoretical_loss": 3.494581389048714,
+      "tokens_seen": 1768947712
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004685443748996951,
+      "loss": 0.0694,
+      "theoretical_loss": 3.494537525246524,
+      "tokens_seen": 1769209856
+    },
+    {
+      "epoch": 0.54,
+      "objective/train/advantage_avg": 0.00046761715202592313,
+      "objective/train/docs_used": 646001,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.280922293663025,
+      "objective/train/original_loss": 1.2809221744537354,
+      "objective/train/theoretical_loss": 3.4944936697626545,
+      "objective/train/tokens_used": 1789932000,
+      "objective/train/value_avg": -0.00762939453125,
+      "objective/train/value_loss": 0.0009191989665850997,
+      "objective/train/value_max": -5.02467155456543e-05,
+      "objective/train/value_min": -0.57470703125,
+      "objective/train/value_reward_corr": 0.6030385380283607,
+      "objective/train/value_std": 0.0171356201171875,
+      "objective/train/weight_avg": 1.0007696151733398,
+      "objective/train/weighted_lm_loss": 1.2816685438156128,
+      "objective/train/weights_max": 1.776610016822815,
+      "objective/train/weights_min": 0.0282779261469841,
+      "theoretical_loss": 3.4944936697626545,
+      "tokens_seen": 1769472000
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004684641309581127,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4944936697626545,
+      "tokens_seen": 1769472000
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046838388701653027,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4944498225942953,
+      "tokens_seen": 1769734144
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046830364307494783,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4944059837386394,
+      "tokens_seen": 1769996288
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046822339913336544,
+      "loss": 0.0712,
+      "theoretical_loss": 3.494362153192879,
+      "tokens_seen": 1770258432
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.000468143155191783,
+      "loss": 0.0698,
+      "theoretical_loss": 3.494318330954209,
+      "tokens_seen": 1770520576
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004680629112502006,
+      "loss": 0.0691,
+      "theoretical_loss": 3.494274517019826,
+      "tokens_seen": 1770782720
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004679826673086182,
+      "loss": 0.0701,
+      "theoretical_loss": 3.494230711386926,
+      "tokens_seen": 1771044864
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004679024233670358,
+      "loss": 0.068,
+      "theoretical_loss": 3.4941869140527095,
+      "tokens_seen": 1771307008
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004678221794254534,
+      "loss": 0.0687,
+      "theoretical_loss": 3.494143125014375,
+      "tokens_seen": 1771569152
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046774193548387096,
+      "loss": 0.0716,
+      "theoretical_loss": 3.4940993442691246,
+      "tokens_seen": 1771831296
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046766169154228857,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4940555718141613,
+      "tokens_seen": 1772093440
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046758144760070613,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4940118076466886,
+      "tokens_seen": 1772355584
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046750120365912374,
+      "loss": 0.069,
+      "theoretical_loss": 3.493968051763912,
+      "tokens_seen": 1772617728
+    },
+    {
+      "epoch": 0.54,
+      "objective/train/advantage_avg": -0.0010038224281743169,
+      "objective/train/docs_used": 647211,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5441070795059204,
+      "objective/train/original_loss": 1.54410719871521,
+      "objective/train/theoretical_loss": 3.493946176928412,
+      "objective/train/tokens_used": 1793208800,
+      "objective/train/value_avg": -0.00521087646484375,
+      "objective/train/value_loss": 0.0002107528707711026,
+      "objective/train/value_max": -8.094310760498047e-05,
+      "objective/train/value_min": -0.254638671875,
+      "objective/train/value_reward_corr": 0.7200556121099299,
+      "objective/train/value_std": 0.01129150390625,
+      "objective/train/weight_avg": 0.9990896582603455,
+      "objective/train/weighted_lm_loss": 1.5431265830993652,
+      "objective/train/weights_max": 1.1554977893829346,
+      "objective/train/weights_min": 0.38069406151771545,
+      "theoretical_loss": 3.493946176928412,
+      "tokens_seen": 1772748800
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046742095971754135,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4939243041630395,
+      "tokens_seen": 1772879872
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004673407157759589,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4938805648412776,
+      "tokens_seen": 1773142016
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004672604718343765,
+      "loss": 0.07,
+      "theoretical_loss": 3.4938368337958368,
+      "tokens_seen": 1773404160
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046718022789279414,
+      "loss": 0.0677,
+      "theoretical_loss": 3.493793111023928,
+      "tokens_seen": 1773666304
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004670999839512117,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4937493965227633,
+      "tokens_seen": 1773928448
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046701974000962925,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4937056902895565,
+      "tokens_seen": 1774190592
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046693949606804687,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4936619923215226,
+      "tokens_seen": 1774452736
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004668592521264645,
+      "loss": 0.0662,
+      "theoretical_loss": 3.493618302615878,
+      "tokens_seen": 1774714880
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046677900818488204,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4935746211698393,
+      "tokens_seen": 1774977024
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046669876424329965,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4935309479806262,
+      "tokens_seen": 1775239168
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046661852030171726,
+      "loss": 0.0717,
+      "theoretical_loss": 3.49348728304546,
+      "tokens_seen": 1775501312
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004665382763601348,
+      "loss": 0.0662,
+      "theoretical_loss": 3.493443626361561,
+      "tokens_seen": 1775763456
+    },
+    {
+      "epoch": 0.54,
+      "objective/train/advantage_avg": 0.0011500284308567643,
+      "objective/train/docs_used": 648417,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3718308210372925,
+      "objective/train/original_loss": 1.371830701828003,
+      "objective/train/theoretical_loss": 3.4933999779261526,
+      "objective/train/tokens_used": 1796485600,
+      "objective/train/value_avg": -0.006534576416015625,
+      "objective/train/value_loss": 0.00019606102432589978,
+      "objective/train/value_max": -8.094310760498047e-05,
+      "objective/train/value_min": -0.62353515625,
+      "objective/train/value_reward_corr": 0.6778037876062109,
+      "objective/train/value_std": 0.0126953125,
+      "objective/train/weight_avg": 1.0012376308441162,
+      "objective/train/weighted_lm_loss": 1.37326979637146,
+      "objective/train/weights_max": 1.2040235996246338,
+      "objective/train/weights_min": 0.3703491687774658,
+      "theoretical_loss": 3.4933999779261526,
+      "tokens_seen": 1776025600
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004664580324185524,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4933999779261526,
+      "tokens_seen": 1776025600
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046637778847697,
+      "loss": 0.069,
+      "theoretical_loss": 3.4933563377364596,
+      "tokens_seen": 1776287744
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004662975445353876,
+      "loss": 0.0681,
+      "theoretical_loss": 3.493312705789708,
+      "tokens_seen": 1776549888
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046621730059380516,
+      "loss": 0.07,
+      "theoretical_loss": 3.493269082083123,
+      "tokens_seen": 1776812032
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004661370566522228,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4932254666139357,
+      "tokens_seen": 1777074176
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004660568127106404,
+      "loss": 0.0649,
+      "theoretical_loss": 3.493181859379374,
+      "tokens_seen": 1777336320
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.000465976568769058,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4931382603766696,
+      "tokens_seen": 1777598464
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004658963248274755,
+      "loss": 0.0662,
+      "theoretical_loss": 3.493094669603055,
+      "tokens_seen": 1777860608
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004658160808858931,
+      "loss": 0.0685,
+      "theoretical_loss": 3.493051087055764,
+      "tokens_seen": 1778122752
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046573583694431073,
+      "loss": 0.071,
+      "theoretical_loss": 3.493007512732031,
+      "tokens_seen": 1778384896
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004656555930027283,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4929639466290934,
+      "tokens_seen": 1778647040
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004655753490611459,
+      "loss": 0.0681,
+      "theoretical_loss": 3.492920388744188,
+      "tokens_seen": 1778909184
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004654951051195635,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4928768390745555,
+      "tokens_seen": 1779171328
+    },
+    {
+      "epoch": 0.54,
+      "objective/train/advantage_avg": 0.0007943502278067172,
+      "objective/train/docs_used": 649543,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4501819610595703,
+      "objective/train/original_loss": 1.4501819610595703,
+      "objective/train/theoretical_loss": 3.4928550673196033,
+      "objective/train/tokens_used": 1799762400,
+      "objective/train/value_avg": -0.00791168212890625,
+      "objective/train/value_loss": 0.0001940899674082175,
+      "objective/train/value_max": -5.1856040954589844e-05,
+      "objective/train/value_min": -0.533203125,
+      "objective/train/value_reward_corr": 0.7600946184960429,
+      "objective/train/value_std": 0.0164642333984375,
+      "objective/train/weight_avg": 1.0008825063705444,
+      "objective/train/weighted_lm_loss": 1.4511456489562988,
+      "objective/train/weights_max": 1.1041755676269531,
+      "objective/train/weights_min": 0.36851781606674194,
+      "theoretical_loss": 3.4928550673196033,
+      "tokens_seen": 1779302400
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046541486117798107,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4928332976174348,
+      "tokens_seen": 1779433472
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046533461723639863,
+      "loss": 0.0679,
+      "theoretical_loss": 3.492789764370068,
+      "tokens_seen": 1779695616
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046525437329481624,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4927462393296986,
+      "tokens_seen": 1779957760
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046517412935323386,
+      "loss": 0.0721,
+      "theoretical_loss": 3.492702722493571,
+      "tokens_seen": 1780219904
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004650938854116514,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4926592138589307,
+      "tokens_seen": 1780482048
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.000465013641470069,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4926157134230253,
+      "tokens_seen": 1780744192
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046493339752848664,
+      "loss": 0.07,
+      "theoretical_loss": 3.4925722211831025,
+      "tokens_seen": 1781006336
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004648531535869042,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4925287371364124,
+      "tokens_seen": 1781268480
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046477290964532176,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4924852612802066,
+      "tokens_seen": 1781530624
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046469266570373937,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4924417936117376,
+      "tokens_seen": 1781792768
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.000464612421762157,
+      "loss": 0.0705,
+      "theoretical_loss": 3.492398334128258,
+      "tokens_seen": 1782054912
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046453217782057454,
+      "loss": 0.0663,
+      "theoretical_loss": 3.492354882827023,
+      "tokens_seen": 1782317056
+    },
+    {
+      "epoch": 0.54,
+      "objective/train/advantage_avg": 0.0007882177596911788,
+      "objective/train/docs_used": 650724,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5506643056869507,
+      "objective/train/original_loss": 1.5506644248962402,
+      "objective/train/theoretical_loss": 3.4923114397052903,
+      "objective/train/tokens_used": 1803039200,
+      "objective/train/value_avg": -0.00815582275390625,
+      "objective/train/value_loss": 0.00025313167134299874,
+      "objective/train/value_max": -0.0001323223114013672,
+      "objective/train/value_min": -0.2152099609375,
+      "objective/train/value_reward_corr": 0.6268371834840105,
+      "objective/train/value_std": 0.0126953125,
+      "objective/train/weight_avg": 1.000898003578186,
+      "objective/train/weighted_lm_loss": 1.551283597946167,
+      "objective/train/weights_max": 1.1620044708251953,
+      "objective/train/weights_min": 0.3706134557723999,
+      "theoretical_loss": 3.4923114397052903,
+      "tokens_seen": 1782579200
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046445193387899215,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4923114397052903,
+      "tokens_seen": 1782579200
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046437168993740976,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4922680047603167,
+      "tokens_seen": 1782841344
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004642914459958273,
+      "loss": 0.0717,
+      "theoretical_loss": 3.4922245779893615,
+      "tokens_seen": 1783103488
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046421120205424494,
+      "loss": 0.0726,
+      "theoretical_loss": 3.492181159389685,
+      "tokens_seen": 1783365632
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004641309581126625,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4921377489585486,
+      "tokens_seen": 1783627776
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046405071417108005,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4920943466932153,
+      "tokens_seen": 1783889920
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046397047022949767,
+      "loss": 0.0681,
+      "theoretical_loss": 3.49205095259095,
+      "tokens_seen": 1784152064
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004638902262879153,
+      "loss": 0.0685,
+      "theoretical_loss": 3.492007566649018,
+      "tokens_seen": 1784414208
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004638099823463329,
+      "loss": 0.0681,
+      "theoretical_loss": 3.491964188864686,
+      "tokens_seen": 1784676352
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046372973840475045,
+      "loss": 0.0728,
+      "theoretical_loss": 3.491920819235223,
+      "tokens_seen": 1784938496
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046364949446316806,
+      "loss": 0.0719,
+      "theoretical_loss": 3.491877457757898,
+      "tokens_seen": 1785200640
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004635692505215856,
+      "loss": 0.0682,
+      "theoretical_loss": 3.491834104429982,
+      "tokens_seen": 1785462784
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004634890065800032,
+      "loss": 0.0672,
+      "theoretical_loss": 3.491790759248747,
+      "tokens_seen": 1785724928
+    },
+    {
+      "epoch": 0.54,
+      "objective/train/advantage_avg": 0.001205379725433886,
+      "objective/train/docs_used": 652000,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.570763111114502,
+      "objective/train/original_loss": 1.570763111114502,
+      "objective/train/theoretical_loss": 3.4917690897122826,
+      "objective/train/tokens_used": 1806316000,
+      "objective/train/value_avg": -0.00769805908203125,
+      "objective/train/value_loss": 0.0003391410573385656,
+      "objective/train/value_max": -4.988908767700195e-05,
+      "objective/train/value_min": -0.9736328125,
+      "objective/train/value_reward_corr": 0.8042611814725403,
+      "objective/train/value_std": 0.0225830078125,
+      "objective/train/weight_avg": 1.0013576745986938,
+      "objective/train/weighted_lm_loss": 1.5740770101547241,
+      "objective/train/weights_max": 1.6434879302978516,
+      "objective/train/weights_min": 0.38063594698905945,
+      "theoretical_loss": 3.4917690897122826,
+      "tokens_seen": 1785856000
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004634087626384208,
+      "loss": 0.0706,
+      "theoretical_loss": 3.491747422211467,
+      "tokens_seen": 1785987072
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004633285186968384,
+      "loss": 0.0681,
+      "theoretical_loss": 3.491704093315416,
+      "tokens_seen": 1786249216
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.000463248274755256,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4916607725578714,
+      "tokens_seen": 1786511360
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004631680308136736,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4916174599361103,
+      "tokens_seen": 1786773504
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004630877868720912,
+      "loss": 0.0723,
+      "theoretical_loss": 3.491574155447411,
+      "tokens_seen": 1787035648
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046300754293050875,
+      "loss": 0.0684,
+      "theoretical_loss": 3.491530859089054,
+      "tokens_seen": 1787297792
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004629272989889263,
+      "loss": 0.0669,
+      "theoretical_loss": 3.49148757085832,
+      "tokens_seen": 1787559936
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004628470550473439,
+      "loss": 0.0723,
+      "theoretical_loss": 3.4914442907524927,
+      "tokens_seen": 1787822080
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046276681110576153,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4914010187688556,
+      "tokens_seen": 1788084224
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046268656716417914,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4913577549046937,
+      "tokens_seen": 1788346368
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004626063232225967,
+      "loss": 0.0673,
+      "theoretical_loss": 3.491314499157294,
+      "tokens_seen": 1788608512
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004625260792810143,
+      "loss": 0.0682,
+      "theoretical_loss": 3.491271251523945,
+      "tokens_seen": 1788870656
+    },
+    {
+      "epoch": 0.54,
+      "objective/train/advantage_avg": 0.0005922395503148437,
+      "objective/train/docs_used": 653179,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4670391082763672,
+      "objective/train/original_loss": 1.4670391082763672,
+      "objective/train/theoretical_loss": 3.491228012001935,
+      "objective/train/tokens_used": 1809592800,
+      "objective/train/value_avg": -0.00855255126953125,
+      "objective/train/value_loss": 0.00044217510730959475,
+      "objective/train/value_max": -7.784366607666016e-05,
+      "objective/train/value_min": -0.64892578125,
+      "objective/train/value_reward_corr": 0.5852997173970007,
+      "objective/train/value_std": 0.0131988525390625,
+      "objective/train/weight_avg": 1.000777006149292,
+      "objective/train/weighted_lm_loss": 1.4677743911743164,
+      "objective/train/weights_max": 1.29947829246521,
+      "objective/train/weights_min": 0.36863452196121216,
+      "theoretical_loss": 3.491228012001935,
+      "tokens_seen": 1789132800
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004624458353394319,
+      "loss": 0.0687,
+      "theoretical_loss": 3.491228012001935,
+      "tokens_seen": 1789132800
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046236559139784943,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4911847805885547,
+      "tokens_seen": 1789394944
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046228534745626704,
+      "loss": 0.0685,
+      "theoretical_loss": 3.491141557281096,
+      "tokens_seen": 1789657088
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046220510351468465,
+      "loss": 0.0728,
+      "theoretical_loss": 3.491098342076852,
+      "tokens_seen": 1789919232
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046212485957310227,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4910551349731183,
+      "tokens_seen": 1790181376
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004620446156315198,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4910119359671885,
+      "tokens_seen": 1790443520
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046196437168993744,
+      "loss": 0.0654,
+      "theoretical_loss": 3.490968745056361,
+      "tokens_seen": 1790705664
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046188412774835505,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4909255622379343,
+      "tokens_seen": 1790967808
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046180388380677256,
+      "loss": 0.0687,
+      "theoretical_loss": 3.490882387509207,
+      "tokens_seen": 1791229952
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046172363986519017,
+      "loss": 0.07,
+      "theoretical_loss": 3.490839220867481,
+      "tokens_seen": 1791492096
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004616433959236078,
+      "loss": 0.0658,
+      "theoretical_loss": 3.490796062310058,
+      "tokens_seen": 1791754240
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046156315198202534,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4907529118342415,
+      "tokens_seen": 1792016384
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046148290804044295,
+      "loss": 0.0677,
+      "theoretical_loss": 3.490709769437337,
+      "tokens_seen": 1792278528
+    },
+    {
+      "epoch": 0.54,
+      "objective/train/advantage_avg": 0.0001855674636317417,
+      "objective/train/docs_used": 654198,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2086480855941772,
+      "objective/train/original_loss": 1.2086482048034668,
+      "objective/train/theoretical_loss": 3.4906882012676346,
+      "objective/train/tokens_used": 1812869600,
+      "objective/train/value_avg": -0.007686614990234375,
+      "objective/train/value_loss": 0.00019353711104486138,
+      "objective/train/value_max": -5.066394805908203e-05,
+      "objective/train/value_min": -0.783203125,
+      "objective/train/value_reward_corr": 0.6790004659897202,
+      "objective/train/value_std": 0.01259613037109375,
+      "objective/train/weight_avg": 1.0002752542495728,
+      "objective/train/weighted_lm_loss": 1.20932936668396,
+      "objective/train/weights_max": 1.4085056781768799,
+      "objective/train/weights_min": 0.4097484350204468,
+      "theoretical_loss": 3.4906882012676346,
+      "tokens_seen": 1792409600
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046140266409886056,
+      "loss": 0.0655,
+      "theoretical_loss": 3.49066663511665,
+      "tokens_seen": 1792540672
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004613224201572782,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4906235088694872,
+      "tokens_seen": 1792802816
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004612421762156957,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4905803906931587,
+      "tokens_seen": 1793064960
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004611619322741133,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4905372805849737,
+      "tokens_seen": 1793327104
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004610816883325309,
+      "loss": 0.0657,
+      "theoretical_loss": 3.490494178542243,
+      "tokens_seen": 1793589248
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046100144439094846,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4904510845622805,
+      "tokens_seen": 1793851392
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004609212004493661,
+      "loss": 0.069,
+      "theoretical_loss": 3.490407998642399,
+      "tokens_seen": 1794113536
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004608409565077837,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4903649207799137,
+      "tokens_seen": 1794375680
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004607607125662013,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4903218509721414,
+      "tokens_seen": 1794637824
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046068046862461886,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4902787892163993,
+      "tokens_seen": 1794899968
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004606002246830364,
+      "loss": 0.0684,
+      "theoretical_loss": 3.490235735510007,
+      "tokens_seen": 1795162112
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046051998074145403,
+      "loss": 0.0702,
+      "theoretical_loss": 3.490192689850284,
+      "tokens_seen": 1795424256
+    },
+    {
+      "epoch": 0.54,
+      "objective/train/advantage_avg": 6.26305063633481e-06,
+      "objective/train/docs_used": 655419,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3910480737686157,
+      "objective/train/original_loss": 1.3910481929779053,
+      "objective/train/theoretical_loss": 3.4901496522345528,
+      "objective/train/tokens_used": 1816146400,
+      "objective/train/value_avg": -0.00922393798828125,
+      "objective/train/value_loss": 0.00022351609368342906,
+      "objective/train/value_max": -4.756450653076172e-05,
+      "objective/train/value_min": -0.712890625,
+      "objective/train/value_reward_corr": 0.7772234877013644,
+      "objective/train/value_std": 0.01708984375,
+      "objective/train/weight_avg": 1.0001089572906494,
+      "objective/train/weighted_lm_loss": 1.390507698059082,
+      "objective/train/weights_max": 1.151555061340332,
+      "objective/train/weights_min": 0.38723233342170715,
+      "theoretical_loss": 3.4901496522345528,
+      "tokens_seen": 1795686400
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004604397367998716,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4901496522345528,
+      "tokens_seen": 1795686400
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004603594928582892,
+      "loss": 0.0686,
+      "theoretical_loss": 3.490106622660136,
+      "tokens_seen": 1795948544
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004602792489167068,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4900636011243567,
+      "tokens_seen": 1796210688
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046019900497512443,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4900205876245414,
+      "tokens_seen": 1796472832
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.000460118761033542,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4899775821580166,
+      "tokens_seen": 1796734976
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00046003851709195954,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4899345847221097,
+      "tokens_seen": 1796997120
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00045995827315037716,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4898915953141505,
+      "tokens_seen": 1797259264
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004598780292087947,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4898486139314695,
+      "tokens_seen": 1797521408
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00045979778526721233,
+      "loss": 0.07,
+      "theoretical_loss": 3.489805640571398,
+      "tokens_seen": 1797783552
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.00045971754132562994,
+      "loss": 0.0664,
+      "theoretical_loss": 3.48976267523127,
+      "tokens_seen": 1798045696
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 0.0004596372973840475,
+      "loss": 0.068,
+      "theoretical_loss": 3.4897197179084185,
+      "tokens_seen": 1798307840
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004595570534424651,
+      "loss": 0.0701,
+      "theoretical_loss": 3.48967676860018,
+      "tokens_seen": 1798569984
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004594768095008827,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4896338273038916,
+      "tokens_seen": 1798832128
+    },
+    {
+      "epoch": 0.55,
+      "objective/train/advantage_avg": 0.0007114228792488575,
+      "objective/train/docs_used": 656735,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3487783670425415,
+      "objective/train/original_loss": 1.348778247833252,
+      "objective/train/theoretical_loss": 3.4896123596593966,
+      "objective/train/tokens_used": 1819423200,
+      "objective/train/value_avg": -0.0080108642578125,
+      "objective/train/value_loss": 0.00022283625730779022,
+      "objective/train/value_max": -7.545948028564453e-05,
+      "objective/train/value_min": -0.61279296875,
+      "objective/train/value_reward_corr": 0.7115171592647844,
+      "objective/train/value_std": 0.014739990234375,
+      "objective/train/weight_avg": 1.0008171796798706,
+      "objective/train/weighted_lm_loss": 1.3492664098739624,
+      "objective/train/weights_max": 1.3917560577392578,
+      "objective/train/weights_min": 0.5280845165252686,
+      "theoretical_loss": 3.4896123596593966,
+      "tokens_seen": 1798963200
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004593965655593003,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4895908940168905,
+      "tokens_seen": 1799094272
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045931632161771784,
+      "loss": 0.0699,
+      "theoretical_loss": 3.489547968736517,
+      "tokens_seen": 1799356416
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045923607767613545,
+      "loss": 0.07,
+      "theoretical_loss": 3.4895050514601116,
+      "tokens_seen": 1799618560
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045915583373455307,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4894621421850163,
+      "tokens_seen": 1799880704
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004590755897929706,
+      "loss": 0.0689,
+      "theoretical_loss": 3.489419240908574,
+      "tokens_seen": 1800142848
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045899534585138824,
+      "loss": 0.0706,
+      "theoretical_loss": 3.48937634762813,
+      "tokens_seen": 1800404992
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045891510190980585,
+      "loss": 0.0674,
+      "theoretical_loss": 3.489333462341029,
+      "tokens_seen": 1800667136
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004588348579682234,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4892905850446185,
+      "tokens_seen": 1800929280
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045875461402664097,
+      "loss": 0.0675,
+      "theoretical_loss": 3.489247715736247,
+      "tokens_seen": 1801191424
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004586743700850586,
+      "loss": 0.0695,
+      "theoretical_loss": 3.489204854413264,
+      "tokens_seen": 1801453568
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004585941261434762,
+      "loss": 0.066,
+      "theoretical_loss": 3.48916200107302,
+      "tokens_seen": 1801715712
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045851388220189375,
+      "loss": 0.0666,
+      "theoretical_loss": 3.489119155712868,
+      "tokens_seen": 1801977856
+    },
+    {
+      "epoch": 0.55,
+      "objective/train/advantage_avg": -0.0004979989607818425,
+      "objective/train/docs_used": 657910,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4576196670532227,
+      "objective/train/original_loss": 1.4576196670532227,
+      "objective/train/theoretical_loss": 3.4890763183301607,
+      "objective/train/tokens_used": 1822700000,
+      "objective/train/value_avg": -0.01110076904296875,
+      "objective/train/value_loss": 0.000849848089274019,
+      "objective/train/value_max": -4.756450653076172e-05,
+      "objective/train/value_min": -0.74853515625,
+      "objective/train/value_reward_corr": 0.656073592097922,
+      "objective/train/value_std": 0.0252685546875,
+      "objective/train/weight_avg": 0.9998723268508911,
+      "objective/train/weighted_lm_loss": 1.4578678607940674,
+      "objective/train/weights_max": 2.0341408252716064,
+      "objective/train/weights_min": 0.37089207768440247,
+      "theoretical_loss": 3.4890763183301607,
+      "tokens_seen": 1802240000
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045843363826031136,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4890763183301607,
+      "tokens_seen": 1802240000
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.000458353394318729,
+      "loss": 0.0693,
+      "theoretical_loss": 3.489033488922253,
+      "tokens_seen": 1802502144
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045827315037714653,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4889906674865,
+      "tokens_seen": 1802764288
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004581929064355641,
+      "loss": 0.0719,
+      "theoretical_loss": 3.48894785402026,
+      "tokens_seen": 1803026432
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004581126624939817,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4889050485208912,
+      "tokens_seen": 1803288576
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004580324185523993,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4888622509857523,
+      "tokens_seen": 1803550720
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004579521746108169,
+      "loss": 0.0653,
+      "theoretical_loss": 3.488819461412205,
+      "tokens_seen": 1803812864
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004578719306692345,
+      "loss": 0.0706,
+      "theoretical_loss": 3.4887766797976116,
+      "tokens_seen": 1804075008
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004577916867276521,
+      "loss": 0.0704,
+      "theoretical_loss": 3.488733906139336,
+      "tokens_seen": 1804337152
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004577114427860697,
+      "loss": 0.0709,
+      "theoretical_loss": 3.4886911404347414,
+      "tokens_seen": 1804599296
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004576311988444872,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4886483826811947,
+      "tokens_seen": 1804861440
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045755095490290483,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4886056328760633,
+      "tokens_seen": 1805123584
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045747071096132244,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4885628910167155,
+      "tokens_seen": 1805385728
+    },
+    {
+      "epoch": 0.55,
+      "objective/train/advantage_avg": -0.00019410904496908188,
+      "objective/train/docs_used": 658936,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.221300482749939,
+      "objective/train/original_loss": 1.2213003635406494,
+      "objective/train/theoretical_loss": 3.4885415230658885,
+      "objective/train/tokens_used": 1825976800,
+      "objective/train/value_avg": -0.008758544921875,
+      "objective/train/value_loss": 0.0004874709411524236,
+      "objective/train/value_max": -2.8431415557861328e-05,
+      "objective/train/value_min": -0.734375,
+      "objective/train/value_reward_corr": 0.7181355256849492,
+      "objective/train/value_std": 0.0221710205078125,
+      "objective/train/weight_avg": 1.0000203847885132,
+      "objective/train/weighted_lm_loss": 1.220677137374878,
+      "objective/train/weights_max": 1.7585545778274536,
+      "objective/train/weights_min": 0.37890908122062683,
+      "theoretical_loss": 3.4885415230658885,
+      "tokens_seen": 1805516800
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045739046701974,
+      "loss": 0.0703,
+      "theoretical_loss": 3.488520157100521,
+      "tokens_seen": 1805647872
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004573102230781576,
+      "loss": 0.0703,
+      "theoretical_loss": 3.4884774311248505,
+      "tokens_seen": 1805910016
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045722997913657523,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4884347130870768,
+      "tokens_seen": 1806172160
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004571497351949928,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4883920029845727,
+      "tokens_seen": 1806434304
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045706949125341034,
+      "loss": 0.0689,
+      "theoretical_loss": 3.488349300814713,
+      "tokens_seen": 1806696448
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045698924731182796,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4883066065748745,
+      "tokens_seen": 1806958592
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045690900337024557,
+      "loss": 0.0706,
+      "theoretical_loss": 3.488263920262434,
+      "tokens_seen": 1807220736
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045682875942866313,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4882212418747693,
+      "tokens_seen": 1807482880
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045674851548708074,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4881785714092617,
+      "tokens_seen": 1807745024
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045666827154549835,
+      "loss": 0.0703,
+      "theoretical_loss": 3.48813590886329,
+      "tokens_seen": 1808007168
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004565880276039159,
+      "loss": 0.0687,
+      "theoretical_loss": 3.488093254234238,
+      "tokens_seen": 1808269312
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045650778366233347,
+      "loss": 0.0683,
+      "theoretical_loss": 3.488050607519489,
+      "tokens_seen": 1808531456
+    },
+    {
+      "epoch": 0.55,
+      "objective/train/advantage_avg": 0.0005030391039326787,
+      "objective/train/docs_used": 660069,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3549247980117798,
+      "objective/train/original_loss": 1.3549246788024902,
+      "objective/train/theoretical_loss": 3.4880079687164267,
+      "objective/train/tokens_used": 1829253600,
+      "objective/train/value_avg": -0.0085906982421875,
+      "objective/train/value_loss": 0.0005426432471722364,
+      "objective/train/value_max": -3.88026237487793e-05,
+      "objective/train/value_min": -0.63037109375,
+      "objective/train/value_reward_corr": 0.6518784997929621,
+      "objective/train/value_std": 0.019317626953125,
+      "objective/train/weight_avg": 1.0007175207138062,
+      "objective/train/weighted_lm_loss": 1.355029821395874,
+      "objective/train/weights_max": 1.2896806001663208,
+      "objective/train/weights_min": 0.24797950685024261,
+      "theoretical_loss": 3.4880079687164267,
+      "tokens_seen": 1808793600
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004564275397207511,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4880079687164267,
+      "tokens_seen": 1808793600
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004563472957791687,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4879653378224384,
+      "tokens_seen": 1809055744
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045626705183758625,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4879227148349106,
+      "tokens_seen": 1809317888
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045618680789600387,
+      "loss": 0.0664,
+      "theoretical_loss": 3.487880099751232,
+      "tokens_seen": 1809580032
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004561065639544215,
+      "loss": 0.0645,
+      "theoretical_loss": 3.487837492568792,
+      "tokens_seen": 1809842176
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045602632001283904,
+      "loss": 0.0697,
+      "theoretical_loss": 3.487794893284981,
+      "tokens_seen": 1810104320
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045594607607125665,
+      "loss": 0.0685,
+      "theoretical_loss": 3.487752301897192,
+      "tokens_seen": 1810366464
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004558658321296742,
+      "loss": 0.0701,
+      "theoretical_loss": 3.487709718402818,
+      "tokens_seen": 1810628608
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045578558818809177,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4876671427992543,
+      "tokens_seen": 1810890752
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004557053442465094,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4876245750838955,
+      "tokens_seen": 1811152896
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.000455625100304927,
+      "loss": 0.0692,
+      "theoretical_loss": 3.48758201525414,
+      "tokens_seen": 1811415040
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004555448563633446,
+      "loss": 0.0668,
+      "theoretical_loss": 3.487539463307385,
+      "tokens_seen": 1811677184
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045546461242176216,
+      "loss": 0.068,
+      "theoretical_loss": 3.4874969192410306,
+      "tokens_seen": 1811939328
+    },
+    {
+      "epoch": 0.55,
+      "objective/train/advantage_avg": 0.0007302032900042832,
+      "objective/train/docs_used": 661289,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.294402837753296,
+      "objective/train/original_loss": 1.2944025993347168,
+      "objective/train/theoretical_loss": 3.4874756501621915,
+      "objective/train/tokens_used": 1832530400,
+      "objective/train/value_avg": -0.01059722900390625,
+      "objective/train/value_loss": 0.0002705962979234755,
+      "objective/train/value_max": -3.045797348022461e-05,
+      "objective/train/value_min": -0.693359375,
+      "objective/train/value_reward_corr": 0.7348594278962314,
+      "objective/train/value_std": 0.018096923828125,
+      "objective/train/weight_avg": 1.0008525848388672,
+      "objective/train/weighted_lm_loss": 1.2947546243667603,
+      "objective/train/weights_max": 1.1715208292007446,
+      "objective/train/weights_min": 0.3793371617794037,
+      "theoretical_loss": 3.4874756501621915,
+      "tokens_seen": 1812070400
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004553843684801798,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4874543830524782,
+      "tokens_seen": 1812201472
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045530412453859733,
+      "loss": 0.0693,
+      "theoretical_loss": 3.487411854739128,
+      "tokens_seen": 1812463616
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004552238805970149,
+      "loss": 0.0693,
+      "theoretical_loss": 3.487369334298386,
+      "tokens_seen": 1812725760
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004551436366554325,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4873268217276543,
+      "tokens_seen": 1812987904
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004550633927138501,
+      "loss": 0.0697,
+      "theoretical_loss": 3.48728431702434,
+      "tokens_seen": 1813250048
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045498314877226773,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4872418201858495,
+      "tokens_seen": 1813512192
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004549029048306853,
+      "loss": 0.069,
+      "theoretical_loss": 3.487199331209591,
+      "tokens_seen": 1813774336
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004548226608891029,
+      "loss": 0.0709,
+      "theoretical_loss": 3.487156850092974,
+      "tokens_seen": 1814036480
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004547424169475205,
+      "loss": 0.0717,
+      "theoretical_loss": 3.487114376833409,
+      "tokens_seen": 1814298624
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.000454662173005938,
+      "loss": 0.0678,
+      "theoretical_loss": 3.487071911428308,
+      "tokens_seen": 1814560768
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045458192906435563,
+      "loss": 0.066,
+      "theoretical_loss": 3.487029453875085,
+      "tokens_seen": 1814822912
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045450168512277324,
+      "loss": 0.0706,
+      "theoretical_loss": 3.4869870041711524,
+      "tokens_seen": 1815085056
+    },
+    {
+      "epoch": 0.55,
+      "objective/train/advantage_avg": 0.0023117726668715477,
+      "objective/train/docs_used": 662349,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2869470119476318,
+      "objective/train/original_loss": 1.2869467735290527,
+      "objective/train/theoretical_loss": 3.4869445623139272,
+      "objective/train/tokens_used": 1835807200,
+      "objective/train/value_avg": -0.006900787353515625,
+      "objective/train/value_loss": 0.00016944836534094065,
+      "objective/train/value_max": -2.7120113372802734e-05,
+      "objective/train/value_min": -0.77392578125,
+      "objective/train/value_reward_corr": 0.7259664757718066,
+      "objective/train/value_std": 0.01531982421875,
+      "objective/train/weight_avg": 1.0023906230926514,
+      "objective/train/weighted_lm_loss": 1.2890067100524902,
+      "objective/train/weights_max": 1.332244634628296,
+      "objective/train/weights_min": 0.37565645575523376,
+      "theoretical_loss": 3.4869445623139272,
+      "tokens_seen": 1815347200
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045442144118119086,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4869445623139272,
+      "tokens_seen": 1815347200
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004543411972396084,
+      "loss": 0.0716,
+      "theoretical_loss": 3.4869021283008257,
+      "tokens_seen": 1815609344
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.000454260953298026,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4868597021292658,
+      "tokens_seen": 1815871488
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045418070935644364,
+      "loss": 0.0711,
+      "theoretical_loss": 3.4868172837966673,
+      "tokens_seen": 1816133632
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045410046541486114,
+      "loss": 0.0664,
+      "theoretical_loss": 3.48677487330045,
+      "tokens_seen": 1816395776
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045402022147327876,
+      "loss": 0.0682,
+      "theoretical_loss": 3.486732470638036,
+      "tokens_seen": 1816657920
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045393997753169637,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4866900758068478,
+      "tokens_seen": 1816920064
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.000453859733590114,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4866476888043096,
+      "tokens_seen": 1817182208
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045377948964853154,
+      "loss": 0.0681,
+      "theoretical_loss": 3.486605309627847,
+      "tokens_seen": 1817444352
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045369924570694915,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4865629382748864,
+      "tokens_seen": 1817706496
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045361900176536676,
+      "loss": 0.0699,
+      "theoretical_loss": 3.486520574742855,
+      "tokens_seen": 1817968640
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045353875782378427,
+      "loss": 0.0667,
+      "theoretical_loss": 3.486478219029183,
+      "tokens_seen": 1818230784
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004534585138822019,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4864358711312997,
+      "tokens_seen": 1818492928
+    },
+    {
+      "epoch": 0.55,
+      "objective/train/advantage_avg": 0.0007954819593578577,
+      "objective/train/docs_used": 663450,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3562122583389282,
+      "objective/train/original_loss": 1.3562121391296387,
+      "objective/train/theoretical_loss": 3.486414700112476,
+      "objective/train/tokens_used": 1839084000,
+      "objective/train/value_avg": -0.00738525390625,
+      "objective/train/value_loss": 0.0002610041992738843,
+      "objective/train/value_max": -5.650520324707031e-05,
+      "objective/train/value_min": -0.3369140625,
+      "objective/train/value_reward_corr": 0.6166013932809289,
+      "objective/train/value_std": 0.0124969482421875,
+      "objective/train/weight_avg": 1.0009105205535889,
+      "objective/train/weighted_lm_loss": 1.3572161197662354,
+      "objective/train/weights_max": 1.1161935329437256,
+      "objective/train/weights_min": 0.36863452196121216,
+      "theoretical_loss": 3.486414700112476,
+      "tokens_seen": 1818624000
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004533782699406195,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4863935310466365,
+      "tokens_seen": 1818755072
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045329802599903705,
+      "loss": 0.069,
+      "theoretical_loss": 3.486351198772626,
+      "tokens_seen": 1819017216
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045321778205745467,
+      "loss": 0.068,
+      "theoretical_loss": 3.4863088743067023,
+      "tokens_seen": 1819279360
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004531375381158723,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4862665576463003,
+      "tokens_seen": 1819541504
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004530572941742899,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4862242487888566,
+      "tokens_seen": 1819803648
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045297705023270745,
+      "loss": 0.068,
+      "theoretical_loss": 3.486181947731808,
+      "tokens_seen": 1820065792
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.000452896806291125,
+      "loss": 0.0688,
+      "theoretical_loss": 3.486139654472594,
+      "tokens_seen": 1820327936
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004528165623495426,
+      "loss": 0.0681,
+      "theoretical_loss": 3.486097369008654,
+      "tokens_seen": 1820590080
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004527363184079602,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4860550913374286,
+      "tokens_seen": 1820852224
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004526560744663778,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4860128214563613,
+      "tokens_seen": 1821114368
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004525758305247954,
+      "loss": 0.071,
+      "theoretical_loss": 3.4859705593628947,
+      "tokens_seen": 1821376512
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.000452495586583213,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4859283050544736,
+      "tokens_seen": 1821638656
+    },
+    {
+      "epoch": 0.55,
+      "objective/train/advantage_avg": -0.000316399586154148,
+      "objective/train/docs_used": 664649,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.22597074508667,
+      "objective/train/original_loss": 1.22597074508667,
+      "objective/train/theoretical_loss": 3.4858860585285445,
+      "objective/train/tokens_used": 1842360800,
+      "objective/train/value_avg": -0.00605010986328125,
+      "objective/train/value_loss": 0.00014420539082493633,
+      "objective/train/value_max": -3.916025161743164e-05,
+      "objective/train/value_min": -0.37353515625,
+      "objective/train/value_reward_corr": 0.7083982900799979,
+      "objective/train/value_std": 0.012542724609375,
+      "objective/train/weight_avg": 0.9997512102127075,
+      "objective/train/weighted_lm_loss": 1.224969506263733,
+      "objective/train/weights_max": 1.3325648307800293,
+      "objective/train/weights_min": 0.37342190742492676,
+      "theoretical_loss": 3.4858860585285445,
+      "tokens_seen": 1821900800
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004524153426416306,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4858860585285445,
+      "tokens_seen": 1821900800
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045233509870004813,
+      "loss": 0.0697,
+      "theoretical_loss": 3.485843819782554,
+      "tokens_seen": 1822162944
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045225485475846575,
+      "loss": 0.0719,
+      "theoretical_loss": 3.48580158881395,
+      "tokens_seen": 1822425088
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004521746108168833,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4857593656201833,
+      "tokens_seen": 1822687232
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004520943668753009,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4857171501987034,
+      "tokens_seen": 1822949376
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045201412293371853,
+      "loss": 0.067,
+      "theoretical_loss": 3.4856749425469635,
+      "tokens_seen": 1823211520
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045193387899213614,
+      "loss": 0.0739,
+      "theoretical_loss": 3.485632742662416,
+      "tokens_seen": 1823473664
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004518536350505537,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4855905505425144,
+      "tokens_seen": 1823735808
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045177339110897126,
+      "loss": 0.0688,
+      "theoretical_loss": 3.485548366184716,
+      "tokens_seen": 1823997952
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045169314716738887,
+      "loss": 0.0711,
+      "theoretical_loss": 3.4855061895864763,
+      "tokens_seen": 1824260096
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045161290322580643,
+      "loss": 0.0692,
+      "theoretical_loss": 3.485464020745253,
+      "tokens_seen": 1824522240
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045153265928422404,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4854218596585067,
+      "tokens_seen": 1824784384
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045145241534264165,
+      "loss": 0.0676,
+      "theoretical_loss": 3.485379706323697,
+      "tokens_seen": 1825046528
+    },
+    {
+      "epoch": 0.55,
+      "objective/train/advantage_avg": 0.0007323589525185525,
+      "objective/train/docs_used": 665640,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.218908667564392,
+      "objective/train/original_loss": 1.2189085483551025,
+      "objective/train/theoretical_loss": 3.4853586325624746,
+      "objective/train/tokens_used": 1845637600,
+      "objective/train/value_avg": -0.00598907470703125,
+      "objective/train/value_loss": 0.0001442520588170737,
+      "objective/train/value_max": -4.57763671875e-05,
+      "objective/train/value_min": -0.74755859375,
+      "objective/train/value_reward_corr": 0.608078860093181,
+      "objective/train/value_std": 0.01049041748046875,
+      "objective/train/weight_avg": 1.000801682472229,
+      "objective/train/weighted_lm_loss": 1.2201964855194092,
+      "objective/train/weights_max": 2.111837863922119,
+      "objective/train/weights_min": 0.38332873582839966,
+      "theoretical_loss": 3.4853586325624746,
+      "tokens_seen": 1825177600
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004513721714010592,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4853375607382846,
+      "tokens_seen": 1825308672
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004512919274594768,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4852954228997337,
+      "tokens_seen": 1825570816
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045121168351789444,
+      "loss": 0.0689,
+      "theoretical_loss": 3.485253292805507,
+      "tokens_seen": 1825832960
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.000451131439576312,
+      "loss": 0.0703,
+      "theoretical_loss": 3.48521117045307,
+      "tokens_seen": 1826095104
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045105119563472956,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4851690558398896,
+      "tokens_seen": 1826357248
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045097095169314717,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4851269489634324,
+      "tokens_seen": 1826619392
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004508907077515648,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4850848498211677,
+      "tokens_seen": 1826881536
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045081046380998234,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4850427584105654,
+      "tokens_seen": 1827143680
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045073021986839995,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4850006747290965,
+      "tokens_seen": 1827405824
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045064997592681756,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4849585987742326,
+      "tokens_seen": 1827667968
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004505697319852351,
+      "loss": 0.0714,
+      "theoretical_loss": 3.4849165305434484,
+      "tokens_seen": 1827930112
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004504894880436527,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4848744700342174,
+      "tokens_seen": 1828192256
+    },
+    {
+      "epoch": 0.55,
+      "objective/train/advantage_avg": 0.0006137097370810807,
+      "objective/train/docs_used": 666768,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2766127586364746,
+      "objective/train/original_loss": 1.276612639427185,
+      "objective/train/theoretical_loss": 3.484832417244016,
+      "objective/train/tokens_used": 1848914400,
+      "objective/train/value_avg": -0.0087432861328125,
+      "objective/train/value_loss": 0.0001864040532382205,
+      "objective/train/value_max": -8.887052536010742e-05,
+      "objective/train/value_min": -0.80712890625,
+      "objective/train/value_reward_corr": 0.7289446558793178,
+      "objective/train/value_std": 0.01708984375,
+      "objective/train/weight_avg": 1.0007046461105347,
+      "objective/train/weighted_lm_loss": 1.2769469022750854,
+      "objective/train/weights_max": 2.075507879257202,
+      "objective/train/weights_min": 0.3701034188270569,
+      "theoretical_loss": 3.484832417244016,
+      "tokens_seen": 1828454400
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004504092441020703,
+      "loss": 0.0698,
+      "theoretical_loss": 3.484832417244016,
+      "tokens_seen": 1828454400
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004503290001604879,
+      "loss": 0.0682,
+      "theoretical_loss": 3.484790372170321,
+      "tokens_seen": 1828716544
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00045024875621890546,
+      "loss": 0.0669,
+      "theoretical_loss": 3.484748334810611,
+      "tokens_seen": 1828978688
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004501685122773231,
+      "loss": 0.0678,
+      "theoretical_loss": 3.484706305162365,
+      "tokens_seen": 1829240832
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004500882683357407,
+      "loss": 0.0672,
+      "theoretical_loss": 3.484664283223064,
+      "tokens_seen": 1829502976
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004500080243941583,
+      "loss": 0.0692,
+      "theoretical_loss": 3.484622268990189,
+      "tokens_seen": 1829765120
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004499277804525758,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4845802624612237,
+      "tokens_seen": 1830027264
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004498475365109934,
+      "loss": 0.0677,
+      "theoretical_loss": 3.484538263633652,
+      "tokens_seen": 1830289408
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.00044976729256941103,
+      "loss": 0.0694,
+      "theoretical_loss": 3.484496272504959,
+      "tokens_seen": 1830551552
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004496870486278286,
+      "loss": 0.0677,
+      "theoretical_loss": 3.484454289072631,
+      "tokens_seen": 1830813696
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004496068046862462,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4844123133341567,
+      "tokens_seen": 1831075840
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 0.0004495265607446638,
+      "loss": 0.0673,
+      "theoretical_loss": 3.484370345287024,
+      "tokens_seen": 1831337984
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044944631680308143,
+      "loss": 0.0666,
+      "theoretical_loss": 3.484328384928723,
+      "tokens_seen": 1831600128
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": 0.0005676686996594071,
+      "objective/train/docs_used": 668070,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.306203007698059,
+      "objective/train/original_loss": 1.3062028884887695,
+      "objective/train/theoretical_loss": 3.484307407632101,
+      "objective/train/tokens_used": 1852191200,
+      "objective/train/value_avg": -0.00788116455078125,
+      "objective/train/value_loss": 0.00015821759006939828,
+      "objective/train/value_max": -5.8770179748535156e-05,
+      "objective/train/value_min": -0.2489013671875,
+      "objective/train/value_reward_corr": 0.7492681530981704,
+      "objective/train/value_std": 0.0144500732421875,
+      "objective/train/weight_avg": 1.0006389617919922,
+      "objective/train/weighted_lm_loss": 1.307263731956482,
+      "objective/train/weights_max": 1.1343940496444702,
+      "objective/train/weights_min": 0.3684152066707611,
+      "theoretical_loss": 3.484307407632101,
+      "tokens_seen": 1831731200
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044936607286149893,
+      "loss": 0.0674,
+      "theoretical_loss": 3.484286432256745,
+      "tokens_seen": 1831862272
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044928582891991654,
+      "loss": 0.0686,
+      "theoretical_loss": 3.484244487268583,
+      "tokens_seen": 1832124416
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044920558497833416,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4842025499617297,
+      "tokens_seen": 1832386560
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004491253410367517,
+      "loss": 0.0716,
+      "theoretical_loss": 3.4841606203336806,
+      "tokens_seen": 1832648704
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044904509709516933,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4841186983819306,
+      "tokens_seen": 1832910848
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044896485315358694,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4840767841039777,
+      "tokens_seen": 1833172992
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004488846092120045,
+      "loss": 0.0696,
+      "theoretical_loss": 3.48403487749732,
+      "tokens_seen": 1833435136
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044880436527042206,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4839929785594563,
+      "tokens_seen": 1833697280
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044872412132883967,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4839510872878883,
+      "tokens_seen": 1833959424
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004486438773872573,
+      "loss": 0.0674,
+      "theoretical_loss": 3.483909203680117,
+      "tokens_seen": 1834221568
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044856363344567484,
+      "loss": 0.0696,
+      "theoretical_loss": 3.483867327733645,
+      "tokens_seen": 1834483712
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044848338950409245,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4838254594459777,
+      "tokens_seen": 1834745856
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": -0.00016138824867084622,
+      "objective/train/docs_used": 669245,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.327626347541809,
+      "objective/train/original_loss": 1.3276262283325195,
+      "objective/train/theoretical_loss": 3.483783598814619,
+      "objective/train/tokens_used": 1855468000,
+      "objective/train/value_avg": -0.00666046142578125,
+      "objective/train/value_loss": 0.0002919305698014796,
+      "objective/train/value_max": -4.166364669799805e-05,
+      "objective/train/value_min": -0.806640625,
+      "objective/train/value_reward_corr": 0.6681307867211738,
+      "objective/train/value_std": 0.016143798828125,
+      "objective/train/weight_avg": 0.9999749064445496,
+      "objective/train/weighted_lm_loss": 1.327237606048584,
+      "objective/train/weights_max": 2.2403690814971924,
+      "objective/train/weights_min": 0.3757309913635254,
+      "theoretical_loss": 3.483783598814619,
+      "tokens_seen": 1835008000
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044840314556251007,
+      "loss": 0.0686,
+      "theoretical_loss": 3.483783598814619,
+      "tokens_seen": 1835008000
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004483229016209276,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4837417458370767,
+      "tokens_seen": 1835270144
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044824265767934524,
+      "loss": 0.0671,
+      "theoretical_loss": 3.483699900510857,
+      "tokens_seen": 1835532288
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004481624137377628,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4836580628334697,
+      "tokens_seen": 1835794432
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004480821697961804,
+      "loss": 0.0719,
+      "theoretical_loss": 3.4836162328024245,
+      "tokens_seen": 1836056576
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044800192585459797,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4835744104152324,
+      "tokens_seen": 1836318720
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004479216819130156,
+      "loss": 0.0697,
+      "theoretical_loss": 3.483532595669406,
+      "tokens_seen": 1836580864
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004478414379714332,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4834907885624586,
+      "tokens_seen": 1836843008
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044776119402985075,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4834489890919045,
+      "tokens_seen": 1837105152
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044768095008826836,
+      "loss": 0.0675,
+      "theoretical_loss": 3.48340719725526,
+      "tokens_seen": 1837367296
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004476007061466859,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4833654130500413,
+      "tokens_seen": 1837629440
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004475204622051035,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4833236364737674,
+      "tokens_seen": 1837891584
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004474402182635211,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4832818675239574,
+      "tokens_seen": 1838153728
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": 0.00036707372055388987,
+      "objective/train/docs_used": 670495,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.408747673034668,
+      "objective/train/original_loss": 1.408747673034668,
+      "objective/train/theoretical_loss": 3.483260985908201,
+      "objective/train/tokens_used": 1858744800,
+      "objective/train/value_avg": -0.00868988037109375,
+      "objective/train/value_loss": 0.0002044513530563563,
+      "objective/train/value_max": -6.109476089477539e-05,
+      "objective/train/value_min": -0.31201171875,
+      "objective/train/value_reward_corr": 0.7900703831940288,
+      "objective/train/value_std": 0.0180511474609375,
+      "objective/train/weight_avg": 1.0004642009735107,
+      "objective/train/weighted_lm_loss": 1.4089860916137695,
+      "objective/train/weights_max": 1.1311835050582886,
+      "objective/train/weights_min": 0.37661492824554443,
+      "theoretical_loss": 3.483260985908201,
+      "tokens_seen": 1838284800
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004473599743219387,
+      "loss": 0.0687,
+      "theoretical_loss": 3.483240106198131,
+      "tokens_seen": 1838415872
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004472797303803563,
+      "loss": 0.0709,
+      "theoretical_loss": 3.483198352493811,
+      "tokens_seen": 1838678016
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004471994864387739,
+      "loss": 0.0718,
+      "theoretical_loss": 3.4831566064085187,
+      "tokens_seen": 1838940160
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004471192424971915,
+      "loss": 0.0679,
+      "theoretical_loss": 3.483114867939779,
+      "tokens_seen": 1839202304
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044703899855560905,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4830731370851167,
+      "tokens_seen": 1839464448
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004469587546140266,
+      "loss": 0.0675,
+      "theoretical_loss": 3.483031413842058,
+      "tokens_seen": 1839726592
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004468785106724442,
+      "loss": 0.07,
+      "theoretical_loss": 3.4829896982081303,
+      "tokens_seen": 1839988736
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044679826673086183,
+      "loss": 0.07,
+      "theoretical_loss": 3.4829479901808624,
+      "tokens_seen": 1840250880
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044671802278927944,
+      "loss": 0.0669,
+      "theoretical_loss": 3.482906289757784,
+      "tokens_seen": 1840513024
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.000446637778847697,
+      "loss": 0.0699,
+      "theoretical_loss": 3.482864596936425,
+      "tokens_seen": 1840775168
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004465575349061146,
+      "loss": 0.0684,
+      "theoretical_loss": 3.482822911714318,
+      "tokens_seen": 1841037312
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044647729096453223,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4827812340889963,
+      "tokens_seen": 1841299456
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": 0.00044453368172980845,
+      "objective/train/docs_used": 671743,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.554997444152832,
+      "objective/train/original_loss": 1.554997205734253,
+      "objective/train/theoretical_loss": 3.4827395640579946,
+      "objective/train/tokens_used": 1862021600,
+      "objective/train/value_avg": -0.0125579833984375,
+      "objective/train/value_loss": 0.0002647726214490831,
+      "objective/train/value_max": -6.604194641113281e-05,
+      "objective/train/value_min": -0.376953125,
+      "objective/train/value_reward_corr": 0.9078510782178396,
+      "objective/train/value_std": 0.03216552734375,
+      "objective/train/weight_avg": 1.00057053565979,
+      "objective/train/weighted_lm_loss": 1.5566171407699585,
+      "objective/train/weights_max": 1.1502026319503784,
+      "objective/train/weights_min": 0.38612306118011475,
+      "theoretical_loss": 3.4827395640579946,
+      "tokens_seen": 1841561600
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044639704702294973,
+      "loss": 0.069,
+      "theoretical_loss": 3.4827395640579946,
+      "tokens_seen": 1841561600
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044631680308136734,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4826979016188475,
+      "tokens_seen": 1841823744
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044623655913978496,
+      "loss": 0.0691,
+      "theoretical_loss": 3.482656246769092,
+      "tokens_seen": 1842085888
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044615631519820257,
+      "loss": 0.0685,
+      "theoretical_loss": 3.482614599506266,
+      "tokens_seen": 1842348032
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044607607125662013,
+      "loss": 0.0697,
+      "theoretical_loss": 3.482572959827908,
+      "tokens_seen": 1842610176
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044599582731503774,
+      "loss": 0.07,
+      "theoretical_loss": 3.482531327731558,
+      "tokens_seen": 1842872320
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044591558337345535,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4824897032147577,
+      "tokens_seen": 1843134464
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044583533943187286,
+      "loss": 0.0674,
+      "theoretical_loss": 3.482448086275049,
+      "tokens_seen": 1843396608
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044575509549029047,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4824064769099756,
+      "tokens_seen": 1843658752
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004456748515487081,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4823648751170824,
+      "tokens_seen": 1843920896
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004455946076071257,
+      "loss": 0.0665,
+      "theoretical_loss": 3.482323280893915,
+      "tokens_seen": 1844183040
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044551436366554325,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4822816942380195,
+      "tokens_seen": 1844445184
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044543411972396087,
+      "loss": 0.067,
+      "theoretical_loss": 3.4822401151469453,
+      "tokens_seen": 1844707328
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": 0.0007432449492625892,
+      "objective/train/docs_used": 672951,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3937199115753174,
+      "objective/train/original_loss": 1.393720030784607,
+      "objective/train/theoretical_loss": 3.4822193284374503,
+      "objective/train/tokens_used": 1865298400,
+      "objective/train/value_avg": -0.0093536376953125,
+      "objective/train/value_loss": 0.0005561860743910074,
+      "objective/train/value_max": -5.608797073364258e-05,
+      "objective/train/value_min": -0.68896484375,
+      "objective/train/value_reward_corr": 0.6907717640017331,
+      "objective/train/value_std": 0.0215911865234375,
+      "objective/train/weight_avg": 1.0009503364562988,
+      "objective/train/weighted_lm_loss": 1.393741488456726,
+      "objective/train/weights_max": 1.7102184295654297,
+      "objective/train/weights_min": 0.06439780443906784,
+      "theoretical_loss": 3.4822193284374503,
+      "tokens_seen": 1844838400
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004453538757823785,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4821985436182405,
+      "tokens_seen": 1844969472
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044527363184079604,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4821569796494565,
+      "tokens_seen": 1845231616
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004451933878992136,
+      "loss": 0.066,
+      "theoretical_loss": 3.482115423238144,
+      "tokens_seen": 1845493760
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004451131439576312,
+      "loss": 0.0711,
+      "theoretical_loss": 3.4820738743818556,
+      "tokens_seen": 1845755904
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044503290001604877,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4820323330781457,
+      "tokens_seen": 1846018048
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004449526560744664,
+      "loss": 0.0677,
+      "theoretical_loss": 3.481990799324568,
+      "tokens_seen": 1846280192
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.000444872412132884,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4819492731186807,
+      "tokens_seen": 1846542336
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004447921681913016,
+      "loss": 0.0668,
+      "theoretical_loss": 3.481907754458039,
+      "tokens_seen": 1846804480
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044471192424971916,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4818662433402014,
+      "tokens_seen": 1847066624
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004446316803081367,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4818247397627284,
+      "tokens_seen": 1847328768
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044455143636655433,
+      "loss": 0.0668,
+      "theoretical_loss": 3.48178324372318,
+      "tokens_seen": 1847590912
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004444711924249719,
+      "loss": 0.0648,
+      "theoretical_loss": 3.481741755219118,
+      "tokens_seen": 1847853056
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": 0.00023911338939797133,
+      "objective/train/docs_used": 673981,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3846668004989624,
+      "objective/train/original_loss": 1.3846670389175415,
+      "objective/train/theoretical_loss": 3.481700274248105,
+      "objective/train/tokens_used": 1868575200,
+      "objective/train/value_avg": -0.007419586181640625,
+      "objective/train/value_loss": 0.00018692106823436916,
+      "objective/train/value_max": -3.5643577575683594e-05,
+      "objective/train/value_min": -0.2083740234375,
+      "objective/train/value_reward_corr": 0.6955404663018411,
+      "objective/train/value_std": 0.0128326416015625,
+      "objective/train/weight_avg": 1.0003236532211304,
+      "objective/train/weighted_lm_loss": 1.3853992223739624,
+      "objective/train/weights_max": 1.1434308290481567,
+      "objective/train/weights_min": 0.36988183856010437,
+      "theoretical_loss": 3.481700274248105,
+      "tokens_seen": 1848115200
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004443909484833895,
+      "loss": 0.0708,
+      "theoretical_loss": 3.481700274248105,
+      "tokens_seen": 1848115200
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004443107045418071,
+      "loss": 0.0708,
+      "theoretical_loss": 3.481658800807706,
+      "tokens_seen": 1848377344
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044423046060022473,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4816173348954846,
+      "tokens_seen": 1848639488
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004441502166586423,
+      "loss": 0.0697,
+      "theoretical_loss": 3.481575876509008,
+      "tokens_seen": 1848901632
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044406997271705985,
+      "loss": 0.0726,
+      "theoretical_loss": 3.4815344256458434,
+      "tokens_seen": 1849163776
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044398972877547746,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4814929823035596,
+      "tokens_seen": 1849425920
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.000443909484833895,
+      "loss": 0.0709,
+      "theoretical_loss": 3.481451546479726,
+      "tokens_seen": 1849688064
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044382924089231263,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4814101181719135,
+      "tokens_seen": 1849950208
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044374899695073024,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4813686973776936,
+      "tokens_seen": 1850212352
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044366875300914786,
+      "loss": 0.0671,
+      "theoretical_loss": 3.48132728409464,
+      "tokens_seen": 1850474496
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004435885090675654,
+      "loss": 0.0713,
+      "theoretical_loss": 3.4812858783203264,
+      "tokens_seen": 1850736640
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.000443508265125983,
+      "loss": 0.0717,
+      "theoretical_loss": 3.481244480052329,
+      "tokens_seen": 1850998784
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004434280211844006,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4812030892882224,
+      "tokens_seen": 1851260928
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": -0.0012252043234184384,
+      "objective/train/docs_used": 675199,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.43165922164917,
+      "objective/train/original_loss": 1.4316593408584595,
+      "objective/train/theoretical_loss": 3.4811823967193716,
+      "objective/train/tokens_used": 1871852000,
+      "objective/train/value_avg": -0.00992584228515625,
+      "objective/train/value_loss": 0.00022598551004193723,
+      "objective/train/value_max": -2.7120113372802734e-05,
+      "objective/train/value_min": -0.332763671875,
+      "objective/train/value_reward_corr": 0.7498402724827826,
+      "objective/train/value_std": 0.0165863037109375,
+      "objective/train/weight_avg": 0.9988827109336853,
+      "objective/train/weighted_lm_loss": 1.4289666414260864,
+      "objective/train/weights_max": 1.112414002418518,
+      "objective/train/weights_min": 0.3692946434020996,
+      "theoretical_loss": 3.4811823967193716,
+      "tokens_seen": 1851392000
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044334777724281814,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4811617060255857,
+      "tokens_seen": 1851523072
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044326753330123576,
+      "loss": 0.069,
+      "theoretical_loss": 3.481120330261997,
+      "tokens_seen": 1851785216
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044318728935965337,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4810789619950366,
+      "tokens_seen": 1852047360
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004431070454180709,
+      "loss": 0.0679,
+      "theoretical_loss": 3.481037601222285,
+      "tokens_seen": 1852309504
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044302680147648854,
+      "loss": 0.0691,
+      "theoretical_loss": 3.480996247941324,
+      "tokens_seen": 1852571648
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044294655753490615,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4809549021497372,
+      "tokens_seen": 1852833792
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004428663135933237,
+      "loss": 0.069,
+      "theoretical_loss": 3.480913563845109,
+      "tokens_seen": 1853095936
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044278606965174127,
+      "loss": 0.0659,
+      "theoretical_loss": 3.480872233025024,
+      "tokens_seen": 1853358080
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004427058257101589,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4808309096870698,
+      "tokens_seen": 1853620224
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004426255817685765,
+      "loss": 0.0666,
+      "theoretical_loss": 3.480789593828834,
+      "tokens_seen": 1853882368
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044254533782699405,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4807482854479037,
+      "tokens_seen": 1854144512
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044246509388541167,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4807069845418708,
+      "tokens_seen": 1854406656
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": 0.0017240240704268217,
+      "objective/train/docs_used": 676483,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4873814582824707,
+      "objective/train/original_loss": 1.4873812198638916,
+      "objective/train/theoretical_loss": 3.4806656911083254,
+      "objective/train/tokens_used": 1875128800,
+      "objective/train/value_avg": -0.00719451904296875,
+      "objective/train/value_loss": 0.00026378282927908003,
+      "objective/train/value_max": -4.988908767700195e-05,
+      "objective/train/value_min": -0.61474609375,
+      "objective/train/value_reward_corr": 0.6747173744723667,
+      "objective/train/value_std": 0.01470184326171875,
+      "objective/train/weight_avg": 1.001841425895691,
+      "objective/train/weighted_lm_loss": 1.489801287651062,
+      "objective/train/weights_max": 1.463398814201355,
+      "objective/train/weights_min": 0.37753555178642273,
+      "theoretical_loss": 3.4806656911083254,
+      "tokens_seen": 1854668800
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004423848499438293,
+      "loss": 0.0721,
+      "theoretical_loss": 3.4806656911083254,
+      "tokens_seen": 1854668800
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044230460600224684,
+      "loss": 0.069,
+      "theoretical_loss": 3.48062440514486,
+      "tokens_seen": 1854930944
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004422243620606644,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4805831266490674,
+      "tokens_seen": 1855193088
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.000442144118119082,
+      "loss": 0.0668,
+      "theoretical_loss": 3.480541855618542,
+      "tokens_seen": 1855455232
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004420638741774996,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4805005920508796,
+      "tokens_seen": 1855717376
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004419836302359172,
+      "loss": 0.0689,
+      "theoretical_loss": 3.480459335943676,
+      "tokens_seen": 1855979520
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004419033862943348,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4804180872945305,
+      "tokens_seen": 1856241664
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004418231423527524,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4803768461010405,
+      "tokens_seen": 1856503808
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044174289841117,
+      "loss": 0.0709,
+      "theoretical_loss": 3.4803356123608062,
+      "tokens_seen": 1856765952
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004416626544695875,
+      "loss": 0.0674,
+      "theoretical_loss": 3.480294386071429,
+      "tokens_seen": 1857028096
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044158241052800513,
+      "loss": 0.068,
+      "theoretical_loss": 3.4802531672305106,
+      "tokens_seen": 1857290240
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044150216658642275,
+      "loss": 0.0702,
+      "theoretical_loss": 3.480211955835654,
+      "tokens_seen": 1857552384
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004414219226448403,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4801707518844647,
+      "tokens_seen": 1857814528
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": 0.00042786714038811624,
+      "objective/train/docs_used": 677697,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2605946063995361,
+      "objective/train/original_loss": 1.2605946063995361,
+      "objective/train/theoretical_loss": 3.4801501526994967,
+      "objective/train/tokens_used": 1878405600,
+      "objective/train/value_avg": -0.00734710693359375,
+      "objective/train/value_loss": 0.00021454686066135764,
+      "objective/train/value_max": -3.069639205932617e-05,
+      "objective/train/value_min": -0.1751708984375,
+      "objective/train/value_reward_corr": 0.6534736056748658,
+      "objective/train/value_std": 0.01160430908203125,
+      "objective/train/weight_avg": 1.0005258321762085,
+      "objective/train/weighted_lm_loss": 1.2597084045410156,
+      "objective/train/weights_max": 1.098419189453125,
+      "objective/train/weights_min": 0.3697379231452942,
+      "theoretical_loss": 3.4801501526994967,
+      "tokens_seen": 1857945600
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004413416787032579,
+      "loss": 0.0655,
+      "theoretical_loss": 3.480129555374547,
+      "tokens_seen": 1858076672
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044126143476167553,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4800883663035083,
+      "tokens_seen": 1858338816
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004411811908200931,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4800471846689556,
+      "tokens_seen": 1858600960
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044110094687851065,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4800060104684984,
+      "tokens_seen": 1858863104
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044102070293692826,
+      "loss": 0.0642,
+      "theoretical_loss": 3.4799648436997463,
+      "tokens_seen": 1859125248
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044094045899534587,
+      "loss": 0.067,
+      "theoretical_loss": 3.47992368436031,
+      "tokens_seen": 1859387392
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044086021505376343,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4798825324478018,
+      "tokens_seen": 1859649536
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044077997111218104,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4798413879598353,
+      "tokens_seen": 1859911680
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044069972717059865,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4798002508940242,
+      "tokens_seen": 1860173824
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004406194832290162,
+      "loss": 0.0697,
+      "theoretical_loss": 3.479759121247984,
+      "tokens_seen": 1860435968
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004405392392874338,
+      "loss": 0.0692,
+      "theoretical_loss": 3.479717999019332,
+      "tokens_seen": 1860698112
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004404589953458514,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4796768842056847,
+      "tokens_seen": 1860960256
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": -0.0001284509344259277,
+      "objective/train/docs_used": 678968,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.393393874168396,
+      "objective/train/original_loss": 1.3933937549591064,
+      "objective/train/theoretical_loss": 3.4796357768046615,
+      "objective/train/tokens_used": 1881682400,
+      "objective/train/value_avg": -0.00817108154296875,
+      "objective/train/value_loss": 0.0005822849925607443,
+      "objective/train/value_max": -5.227327346801758e-05,
+      "objective/train/value_min": -0.6240234375,
+      "objective/train/value_reward_corr": 0.5789475016654009,
+      "objective/train/value_std": 0.01580810546875,
+      "objective/train/weight_avg": 1.0001150369644165,
+      "objective/train/weighted_lm_loss": 1.3927175998687744,
+      "objective/train/weights_max": 1.866422414779663,
+      "objective/train/weights_min": 0.2532411217689514,
+      "theoretical_loss": 3.4796357768046615,
+      "tokens_seen": 1861222400
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.000440378751404269,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4796357768046615,
+      "tokens_seen": 1861222400
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044029850746268656,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4795946768138823,
+      "tokens_seen": 1861484544
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044021826352110417,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4795535842309677,
+      "tokens_seen": 1861746688
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004401380195795218,
+      "loss": 0.0715,
+      "theoretical_loss": 3.4795124990535395,
+      "tokens_seen": 1862008832
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00044005777563793934,
+      "loss": 0.0658,
+      "theoretical_loss": 3.479471421279222,
+      "tokens_seen": 1862270976
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00043997753169635695,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4794303509056377,
+      "tokens_seen": 1862533120
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004398972877547745,
+      "loss": 0.068,
+      "theoretical_loss": 3.479389287930413,
+      "tokens_seen": 1862795264
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004398170438131921,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4793482323511746,
+      "tokens_seen": 1863057408
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004397367998716097,
+      "loss": 0.071,
+      "theoretical_loss": 3.479307184165549,
+      "tokens_seen": 1863319552
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004396565559300273,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4792661433711656,
+      "tokens_seen": 1863581696
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004395763119884449,
+      "loss": 0.0688,
+      "theoretical_loss": 3.479225109965653,
+      "tokens_seen": 1863843840
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.00043949606804686246,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4791840839466435,
+      "tokens_seen": 1864105984
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 0.0004394158241052801,
+      "loss": 0.069,
+      "theoretical_loss": 3.479143065311768,
+      "tokens_seen": 1864368128
+    },
+    {
+      "epoch": 0.56,
+      "objective/train/advantage_avg": 0.0003821462451014668,
+      "objective/train/docs_used": 680280,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1618165969848633,
+      "objective/train/original_loss": 1.1618165969848633,
+      "objective/train/theoretical_loss": 3.479122558762641,
+      "objective/train/tokens_used": 1884959200,
+      "objective/train/value_avg": -0.005954742431640625,
+      "objective/train/value_loss": 0.0002789311984088272,
+      "objective/train/value_max": -6.014108657836914e-05,
+      "objective/train/value_min": -0.3486328125,
+      "objective/train/value_reward_corr": 0.5803413402854933,
+      "objective/train/value_std": 0.0120849609375,
+      "objective/train/weight_avg": 1.000502347946167,
+      "objective/train/weighted_lm_loss": 1.162131667137146,
+      "objective/train/weights_max": 1.3961803913116455,
+      "objective/train/weights_min": 0.39436399936676025,
+      "theoretical_loss": 3.479122558762641,
+      "tokens_seen": 1864499200
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043933558016369764,
+      "loss": 0.0667,
+      "theoretical_loss": 3.47910205405866,
+      "tokens_seen": 1864630272
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004392553362221152,
+      "loss": 0.0649,
+      "theoretical_loss": 3.479061050184953,
+      "tokens_seen": 1864892416
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004391750922805328,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4790200536882825,
+      "tokens_seen": 1865154560
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004390948483389504,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4789790645662846,
+      "tokens_seen": 1865416704
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043901460439736803,
+      "loss": 0.07,
+      "theoretical_loss": 3.478938082816597,
+      "tokens_seen": 1865678848
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004389343604557856,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4788971084368576,
+      "tokens_seen": 1865940992
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004388541165142032,
+      "loss": 0.0706,
+      "theoretical_loss": 3.478856141424706,
+      "tokens_seen": 1866203136
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004387738725726208,
+      "loss": 0.0698,
+      "theoretical_loss": 3.478815181777783,
+      "tokens_seen": 1866465280
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004386936286310383,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4787742294937303,
+      "tokens_seen": 1866727424
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043861338468945593,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4787332845701906,
+      "tokens_seen": 1866989568
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043853314074787354,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4786923470048077,
+      "tokens_seen": 1867251712
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043845289680629116,
+      "loss": 0.0699,
+      "theoretical_loss": 3.478651416795227,
+      "tokens_seen": 1867513856
+    },
+    {
+      "epoch": 0.57,
+      "objective/train/advantage_avg": 0.0006913833203725517,
+      "objective/train/docs_used": 681429,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.311995267868042,
+      "objective/train/original_loss": 1.311995267868042,
+      "objective/train/theoretical_loss": 3.478610493939094,
+      "objective/train/tokens_used": 1888236000,
+      "objective/train/value_avg": -0.00714111328125,
+      "objective/train/value_loss": 0.00026110518956556916,
+      "objective/train/value_max": -3.88026237487793e-05,
+      "objective/train/value_min": -0.63818359375,
+      "objective/train/value_reward_corr": 0.824775167023786,
+      "objective/train/value_std": 0.02099609375,
+      "objective/train/weight_avg": 1.0008121728897095,
+      "objective/train/weighted_lm_loss": 1.3127737045288086,
+      "objective/train/weights_max": 1.6890244483947754,
+      "objective/train/weights_min": 0.3925867974758148,
+      "theoretical_loss": 3.478610493939094,
+      "tokens_seen": 1867776000
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004383726528647087,
+      "loss": 0.0672,
+      "theoretical_loss": 3.478610493939094,
+      "tokens_seen": 1867776000
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043829240892312633,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4785695784340556,
+      "tokens_seen": 1868038144
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043821216498154394,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4785286702777602,
+      "tokens_seen": 1868300288
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043813192103996145,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4784877694678573,
+      "tokens_seen": 1868562432
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043805167709837906,
+      "loss": 0.0695,
+      "theoretical_loss": 3.478446876001997,
+      "tokens_seen": 1868824576
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043797143315679667,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4784059898778312,
+      "tokens_seen": 1869086720
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004378911892152143,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4783651110930123,
+      "tokens_seen": 1869348864
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043781094527363184,
+      "loss": 0.0698,
+      "theoretical_loss": 3.478324239645193,
+      "tokens_seen": 1869611008
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043773070133204945,
+      "loss": 0.0715,
+      "theoretical_loss": 3.478283375532029,
+      "tokens_seen": 1869873152
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043765045739046707,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4782425187511756,
+      "tokens_seen": 1870135296
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043757021344888457,
+      "loss": 0.0688,
+      "theoretical_loss": 3.47820166930029,
+      "tokens_seen": 1870397440
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004374899695073022,
+      "loss": 0.0675,
+      "theoretical_loss": 3.47816082717703,
+      "tokens_seen": 1870659584
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004374097255657198,
+      "loss": 0.0664,
+      "theoretical_loss": 3.478119992379054,
+      "tokens_seen": 1870921728
+    },
+    {
+      "epoch": 0.57,
+      "objective/train/advantage_avg": 0.000833529862575233,
+      "objective/train/docs_used": 682627,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3730967044830322,
+      "objective/train/original_loss": 1.3730967044830322,
+      "objective/train/theoretical_loss": 3.4780995777263164,
+      "objective/train/tokens_used": 1891512800,
+      "objective/train/value_avg": -0.00873565673828125,
+      "objective/train/value_loss": 0.0007085340330377221,
+      "objective/train/value_max": -7.426738739013672e-05,
+      "objective/train/value_min": -0.892578125,
+      "objective/train/value_reward_corr": 0.7890690165374458,
+      "objective/train/value_std": 0.02630615234375,
+      "objective/train/weight_avg": 1.0011346340179443,
+      "objective/train/weighted_lm_loss": 1.3734325170516968,
+      "objective/train/weights_max": 1.620538592338562,
+      "objective/train/weights_min": 0.3686436712741852,
+      "theoretical_loss": 3.4780995777263164,
+      "tokens_seen": 1871052800
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004373294816241374,
+      "loss": 0.0721,
+      "theoretical_loss": 3.478079164904022,
+      "tokens_seen": 1871183872
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043724923768255497,
+      "loss": 0.0711,
+      "theoretical_loss": 3.4780383447495966,
+      "tokens_seen": 1871446016
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004371689937409726,
+      "loss": 0.0669,
+      "theoretical_loss": 3.477997531913439,
+      "tokens_seen": 1871708160
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004370887497993902,
+      "loss": 0.0716,
+      "theoretical_loss": 3.477956726393212,
+      "tokens_seen": 1871970304
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043700850585780775,
+      "loss": 0.0702,
+      "theoretical_loss": 3.477915928186581,
+      "tokens_seen": 1872232448
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004369282619162253,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4778751372912105,
+      "tokens_seen": 1872494592
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004368480179746429,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4778343537047673,
+      "tokens_seen": 1872756736
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004367677740330605,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4777935774249196,
+      "tokens_seen": 1873018880
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004366875300914781,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4777528084493348,
+      "tokens_seen": 1873281024
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004366072861498957,
+      "loss": 0.0683,
+      "theoretical_loss": 3.477712046775684,
+      "tokens_seen": 1873543168
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004365270422083133,
+      "loss": 0.0694,
+      "theoretical_loss": 3.477671292401637,
+      "tokens_seen": 1873805312
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004364467982667309,
+      "loss": 0.068,
+      "theoretical_loss": 3.477630545324866,
+      "tokens_seen": 1874067456
+    },
+    {
+      "epoch": 0.57,
+      "objective/train/advantage_avg": -0.0009515096899122,
+      "objective/train/docs_used": 683764,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4821826219558716,
+      "objective/train/original_loss": 1.4821826219558716,
+      "objective/train/theoretical_loss": 3.477589805543044,
+      "objective/train/tokens_used": 1894789600,
+      "objective/train/value_avg": -0.0078277587890625,
+      "objective/train/value_loss": 0.00039342811214737594,
+      "objective/train/value_max": -5.269050598144531e-05,
+      "objective/train/value_min": -0.91845703125,
+      "objective/train/value_reward_corr": 0.6983727521608852,
+      "objective/train/value_std": 0.0157318115234375,
+      "objective/train/weight_avg": 0.9992176294326782,
+      "objective/train/weighted_lm_loss": 1.4811550378799438,
+      "objective/train/weights_max": 1.3486071825027466,
+      "objective/train/weights_min": 0.3824990689754486,
+      "theoretical_loss": 3.477589805543044,
+      "tokens_seen": 1874329600
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043636655432514843,
+      "loss": 0.0708,
+      "theoretical_loss": 3.477589805543044,
+      "tokens_seen": 1874329600
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043628631038356605,
+      "loss": 0.0719,
+      "theoretical_loss": 3.477549073053845,
+      "tokens_seen": 1874591744
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004362060664419836,
+      "loss": 0.0701,
+      "theoretical_loss": 3.477508347854944,
+      "tokens_seen": 1874853888
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004361258225004012,
+      "loss": 0.0688,
+      "theoretical_loss": 3.477467629944017,
+      "tokens_seen": 1875116032
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043604557855881883,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4774269193187406,
+      "tokens_seen": 1875378176
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043596533461723644,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4773862159767943,
+      "tokens_seen": 1875640320
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.000435885090675654,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4773455199158567,
+      "tokens_seen": 1875902464
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004358048467340716,
+      "loss": 0.0716,
+      "theoretical_loss": 3.4773048311336083,
+      "tokens_seen": 1876164608
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004357246027924892,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4772641496277306,
+      "tokens_seen": 1876426752
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043564435885090673,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4772234753959057,
+      "tokens_seen": 1876688896
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043556411490932434,
+      "loss": 0.0689,
+      "theoretical_loss": 3.477182808435818,
+      "tokens_seen": 1876951040
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043548387096774196,
+      "loss": 0.0675,
+      "theoretical_loss": 3.477142148745151,
+      "tokens_seen": 1877213184
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043540362702615957,
+      "loss": 0.0698,
+      "theoretical_loss": 3.477101496321591,
+      "tokens_seen": 1877475328
+    },
+    {
+      "epoch": 0.57,
+      "objective/train/advantage_avg": 0.00011132625513710082,
+      "objective/train/docs_used": 684931,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.389501690864563,
+      "objective/train/original_loss": 1.3895015716552734,
+      "objective/train/theoretical_loss": 3.4770811728342537,
+      "objective/train/tokens_used": 1898066400,
+      "objective/train/value_avg": -0.0081787109375,
+      "objective/train/value_loss": 0.00024089377257041633,
+      "objective/train/value_max": -4.267692565917969e-05,
+      "objective/train/value_min": -0.45849609375,
+      "objective/train/value_reward_corr": 0.6890634759535107,
+      "objective/train/value_std": 0.01363372802734375,
+      "objective/train/weight_avg": 1.0002244710922241,
+      "objective/train/weighted_lm_loss": 1.3898746967315674,
+      "objective/train/weights_max": 1.43800687789917,
+      "objective/train/weights_min": 0.4132646918296814,
+      "theoretical_loss": 3.4770811728342537,
+      "tokens_seen": 1877606400
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043532338308457713,
+      "loss": 0.0703,
+      "theoretical_loss": 3.4770608511628254,
+      "tokens_seen": 1877737472
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043524313914299474,
+      "loss": 0.0727,
+      "theoretical_loss": 3.477020213266541,
+      "tokens_seen": 1877999616
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004351628952014123,
+      "loss": 0.0717,
+      "theoretical_loss": 3.476979582630427,
+      "tokens_seen": 1878261760
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043508265125982986,
+      "loss": 0.0704,
+      "theoretical_loss": 3.4769389592521733,
+      "tokens_seen": 1878523904
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043500240731824747,
+      "loss": 0.0729,
+      "theoretical_loss": 3.476898343129471,
+      "tokens_seen": 1878786048
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004349221633766651,
+      "loss": 0.0666,
+      "theoretical_loss": 3.476857734260012,
+      "tokens_seen": 1879048192
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043484191943508264,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4768171326414894,
+      "tokens_seen": 1879310336
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043476167549350025,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4767765382715976,
+      "tokens_seen": 1879572480
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043468143155191787,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4767359511480316,
+      "tokens_seen": 1879834624
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004346011876103354,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4766953712684874,
+      "tokens_seen": 1880096768
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.000434520943668753,
+      "loss": 0.068,
+      "theoretical_loss": 3.476654798630663,
+      "tokens_seen": 1880358912
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004344406997271706,
+      "loss": 0.0687,
+      "theoretical_loss": 3.476614233232256,
+      "tokens_seen": 1880621056
+    },
+    {
+      "epoch": 0.57,
+      "objective/train/advantage_avg": 0.0005197719438001513,
+      "objective/train/docs_used": 686045,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3149397373199463,
+      "objective/train/original_loss": 1.3149397373199463,
+      "objective/train/theoretical_loss": 3.476573675070966,
+      "objective/train/tokens_used": 1901343200,
+      "objective/train/value_avg": -0.00659942626953125,
+      "objective/train/value_loss": 0.00010198719974141568,
+      "objective/train/value_max": -4.166364669799805e-05,
+      "objective/train/value_min": -0.342041015625,
+      "objective/train/value_reward_corr": 0.7580199579982837,
+      "objective/train/value_std": 0.011993408203125,
+      "objective/train/weight_avg": 1.0005700588226318,
+      "objective/train/weighted_lm_loss": 1.3154451847076416,
+      "objective/train/weights_max": 1.1764196157455444,
+      "objective/train/weights_min": 0.7539494037628174,
+      "theoretical_loss": 3.476573675070966,
+      "tokens_seen": 1880883200
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004343604557855882,
+      "loss": 0.0678,
+      "theoretical_loss": 3.476573675070966,
+      "tokens_seen": 1880883200
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043428021184400577,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4765331241444937,
+      "tokens_seen": 1881145344
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004341999679024234,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4764925804505404,
+      "tokens_seen": 1881407488
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.000434119723960841,
+      "loss": 0.0671,
+      "theoretical_loss": 3.476452043986809,
+      "tokens_seen": 1881669632
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004340394800192586,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4764115147510033,
+      "tokens_seen": 1881931776
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004339592360776761,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4763709927408266,
+      "tokens_seen": 1882193920
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004338789921360937,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4763304779539865,
+      "tokens_seen": 1882456064
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043379874819451133,
+      "loss": 0.0716,
+      "theoretical_loss": 3.476289970388188,
+      "tokens_seen": 1882718208
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004337185042529289,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4762494700411404,
+      "tokens_seen": 1882980352
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004336382603113465,
+      "loss": 0.0682,
+      "theoretical_loss": 3.476208976910552,
+      "tokens_seen": 1883242496
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004335580163697641,
+      "loss": 0.0692,
+      "theoretical_loss": 3.476168490994132,
+      "tokens_seen": 1883504640
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043347777242818173,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4761280122895926,
+      "tokens_seen": 1883766784
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043339752848659923,
+      "loss": 0.0709,
+      "theoretical_loss": 3.476087540794645,
+      "tokens_seen": 1884028928
+    },
+    {
+      "epoch": 0.57,
+      "objective/train/advantage_avg": -0.00022904000070411712,
+      "objective/train/docs_used": 687181,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.41051185131073,
+      "objective/train/original_loss": 1.4105117321014404,
+      "objective/train/theoretical_loss": 3.4760673077500535,
+      "objective/train/tokens_used": 1904620000,
+      "objective/train/value_avg": -0.005123138427734375,
+      "objective/train/value_loss": 0.00014444446424022317,
+      "objective/train/value_max": -3.6776065826416016e-05,
+      "objective/train/value_min": -0.370849609375,
+      "objective/train/value_reward_corr": 0.6669141158683306,
+      "objective/train/value_std": 0.01131439208984375,
+      "objective/train/weight_avg": 0.9998379349708557,
+      "objective/train/weighted_lm_loss": 1.410766363143921,
+      "objective/train/weights_max": 1.1633652448654175,
+      "objective/train/weights_min": 0.3695588707923889,
+      "theoretical_loss": 3.4760673077500535,
+      "tokens_seen": 1884160000
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043331728454501685,
+      "loss": 0.071,
+      "theoretical_loss": 3.476047076507002,
+      "tokens_seen": 1884291072
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043323704060343446,
+      "loss": 0.069,
+      "theoretical_loss": 3.4760066194243784,
+      "tokens_seen": 1884553216
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.000433156796661852,
+      "loss": 0.073,
+      "theoretical_loss": 3.4759661695444892,
+      "tokens_seen": 1884815360
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043307655272026963,
+      "loss": 0.0714,
+      "theoretical_loss": 3.47592572686505,
+      "tokens_seen": 1885077504
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043299630877868724,
+      "loss": 0.0712,
+      "theoretical_loss": 3.4758852913837783,
+      "tokens_seen": 1885339648
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004329160648371048,
+      "loss": 0.0673,
+      "theoretical_loss": 3.475844863098393,
+      "tokens_seen": 1885601792
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043283582089552236,
+      "loss": 0.0685,
+      "theoretical_loss": 3.475804442006612,
+      "tokens_seen": 1885863936
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043275557695393997,
+      "loss": 0.072,
+      "theoretical_loss": 3.4757640281061573,
+      "tokens_seen": 1886126080
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004326753330123576,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4757236213947484,
+      "tokens_seen": 1886388224
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043259508907077514,
+      "loss": 0.069,
+      "theoretical_loss": 3.4756832218701095,
+      "tokens_seen": 1886650368
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043251484512919276,
+      "loss": 0.0686,
+      "theoretical_loss": 3.475642829529963,
+      "tokens_seen": 1886912512
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043243460118761037,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4756024443720337,
+      "tokens_seen": 1887174656
+    },
+    {
+      "epoch": 0.57,
+      "objective/train/advantage_avg": -6.266022683121264e-05,
+      "objective/train/docs_used": 688313,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5270166397094727,
+      "objective/train/original_loss": 1.5270166397094727,
+      "objective/train/theoretical_loss": 3.4755620663940467,
+      "objective/train/tokens_used": 1907896800,
+      "objective/train/value_avg": -0.005428314208984375,
+      "objective/train/value_loss": 0.00046849262434989214,
+      "objective/train/value_max": -3.147125244140625e-05,
+      "objective/train/value_min": -0.85888671875,
+      "objective/train/value_reward_corr": 0.546287861758024,
+      "objective/train/value_std": 0.0126953125,
+      "objective/train/weight_avg": 1.000136375427246,
+      "objective/train/weighted_lm_loss": 1.5277141332626343,
+      "objective/train/weights_max": 1.4028693437576294,
+      "objective/train/weights_min": 0.39049577713012695,
+      "theoretical_loss": 3.4755620663940467,
+      "tokens_seen": 1887436800
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043235435724602793,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4755620663940467,
+      "tokens_seen": 1887436800
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043227411330444554,
+      "loss": 0.0713,
+      "theoretical_loss": 3.4755216955937294,
+      "tokens_seen": 1887698944
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004321938693628631,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4754813319688087,
+      "tokens_seen": 1887961088
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004321136254212807,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4754409755170133,
+      "tokens_seen": 1888223232
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043203338147969827,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4754006262360733,
+      "tokens_seen": 1888485376
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004319531375381159,
+      "loss": 0.0639,
+      "theoretical_loss": 3.4753602841237186,
+      "tokens_seen": 1888747520
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004318728935965335,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4753199491776816,
+      "tokens_seen": 1889009664
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043179264965495105,
+      "loss": 0.0675,
+      "theoretical_loss": 3.475279621395695,
+      "tokens_seen": 1889271808
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043171240571336867,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4752393007754923,
+      "tokens_seen": 1889533952
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004316321617717862,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4751989873148084,
+      "tokens_seen": 1889796096
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043155191783020384,
+      "loss": 0.0718,
+      "theoretical_loss": 3.4751586810113793,
+      "tokens_seen": 1890058240
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004314716738886214,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4751183818629414,
+      "tokens_seen": 1890320384
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.000431391429947039,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4750780898672335,
+      "tokens_seen": 1890582528
+    },
+    {
+      "epoch": 0.57,
+      "objective/train/advantage_avg": 0.0010180543176829815,
+      "objective/train/docs_used": 689504,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3283250331878662,
+      "objective/train/original_loss": 1.3283251523971558,
+      "objective/train/theoretical_loss": 3.4750579465509466,
+      "objective/train/tokens_used": 1911173600,
+      "objective/train/value_avg": -0.007610321044921875,
+      "objective/train/value_loss": 0.0002268047392135486,
+      "objective/train/value_max": -3.618001937866211e-05,
+      "objective/train/value_min": -0.36572265625,
+      "objective/train/value_reward_corr": 0.656805733107698,
+      "objective/train/value_std": 0.01404571533203125,
+      "objective/train/weight_avg": 1.0011183023452759,
+      "objective/train/weighted_lm_loss": 1.3299323320388794,
+      "objective/train/weights_max": 1.2525979280471802,
+      "objective/train/weights_min": 0.37432897090911865,
+      "theoretical_loss": 3.4750579465509466,
+      "tokens_seen": 1890713600
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004313111860054566,
+      "loss": 0.0654,
+      "theoretical_loss": 3.475037805021994,
+      "tokens_seen": 1890844672
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004312309420638742,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4749975273249625,
+      "tokens_seen": 1891106816
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004311506981222918,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4749572567738807,
+      "tokens_seen": 1891368960
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004310704541807094,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4749169933664903,
+      "tokens_seen": 1891631104
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004309902102391269,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4748767371005345,
+      "tokens_seen": 1891893248
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004309099662975445,
+      "loss": 0.0686,
+      "theoretical_loss": 3.474836487973757,
+      "tokens_seen": 1892155392
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043082972235596213,
+      "loss": 0.0699,
+      "theoretical_loss": 3.474796245983903,
+      "tokens_seen": 1892417536
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043074947841437975,
+      "loss": 0.0704,
+      "theoretical_loss": 3.4747560111287195,
+      "tokens_seen": 1892679680
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004306692344727973,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4747157834059523,
+      "tokens_seen": 1892941824
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004305889905312149,
+      "loss": 0.068,
+      "theoretical_loss": 3.47467556281335,
+      "tokens_seen": 1893203968
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043050874658963253,
+      "loss": 0.07,
+      "theoretical_loss": 3.474635349348662,
+      "tokens_seen": 1893466112
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043042850264805003,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4745951430096387,
+      "tokens_seen": 1893728256
+    },
+    {
+      "epoch": 0.57,
+      "objective/train/advantage_avg": 0.0006099278689362109,
+      "objective/train/docs_used": 690600,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4099087715148926,
+      "objective/train/original_loss": 1.4099085330963135,
+      "objective/train/theoretical_loss": 3.474554943794031,
+      "objective/train/tokens_used": 1914450400,
+      "objective/train/value_avg": -0.0048828125,
+      "objective/train/value_loss": 9.77657618932426e-05,
+      "objective/train/value_max": -5.346536636352539e-05,
+      "objective/train/value_min": -0.19580078125,
+      "objective/train/value_reward_corr": 0.629735219499377,
+      "objective/train/value_std": 0.00786590576171875,
+      "objective/train/weight_avg": 1.000654697418213,
+      "objective/train/weighted_lm_loss": 1.4121841192245483,
+      "objective/train/weights_max": 1.0596908330917358,
+      "objective/train/weights_min": 0.3795166611671448,
+      "theoretical_loss": 3.474554943794031,
+      "tokens_seen": 1893990400
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043034825870646765,
+      "loss": 0.0689,
+      "theoretical_loss": 3.474554943794031,
+      "tokens_seen": 1893990400
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043026801476488526,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4745147516995916,
+      "tokens_seen": 1894252544
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043018777082330287,
+      "loss": 0.0689,
+      "theoretical_loss": 3.474474566724073,
+      "tokens_seen": 1894514688
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043010752688172043,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4744343888652303,
+      "tokens_seen": 1894776832
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00043002728294013804,
+      "loss": 0.0665,
+      "theoretical_loss": 3.474394218120818,
+      "tokens_seen": 1895038976
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00042994703899855565,
+      "loss": 0.0691,
+      "theoretical_loss": 3.474354054488593,
+      "tokens_seen": 1895301120
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00042986679505697316,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4743138979663133,
+      "tokens_seen": 1895563264
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00042978655111539077,
+      "loss": 0.0721,
+      "theoretical_loss": 3.4742737485517354,
+      "tokens_seen": 1895825408
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004297063071738084,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4742336062426205,
+      "tokens_seen": 1896087552
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.000429626063232226,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4741934710367284,
+      "tokens_seen": 1896349696
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00042954581929064356,
+      "loss": 0.0695,
+      "theoretical_loss": 3.47415334293182,
+      "tokens_seen": 1896611840
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00042946557534906117,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4741132219256583,
+      "tokens_seen": 1896873984
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.0004293853314074788,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4740731080160066,
+      "tokens_seen": 1897136128
+    },
+    {
+      "epoch": 0.57,
+      "objective/train/advantage_avg": 0.0012116730213165283,
+      "objective/train/docs_used": 691800,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4639194011688232,
+      "objective/train/original_loss": 1.4639195203781128,
+      "objective/train/theoretical_loss": 3.474053053721673,
+      "objective/train/tokens_used": 1917727200,
+      "objective/train/value_avg": -0.00799560546875,
+      "objective/train/value_loss": 0.0002491466875653714,
+      "objective/train/value_max": -0.0001233816146850586,
+      "objective/train/value_min": -0.67626953125,
+      "objective/train/value_reward_corr": 0.6715546928250509,
+      "objective/train/value_std": 0.01512908935546875,
+      "objective/train/weight_avg": 1.001329779624939,
+      "objective/train/weighted_lm_loss": 1.4651811122894287,
+      "objective/train/weights_max": 1.9665279388427734,
+      "objective/train/weights_min": 0.37346750497817993,
+      "theoretical_loss": 3.474053053721673,
+      "tokens_seen": 1897267200
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 0.00042930508746589634,
+      "loss": 0.0719,
+      "theoretical_loss": 3.4740330012006293,
+      "tokens_seen": 1897398272
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004292248435243139,
+      "loss": 0.0682,
+      "theoretical_loss": 3.473992901477292,
+      "tokens_seen": 1897660416
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004291445995827315,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4739528088437606,
+      "tokens_seen": 1897922560
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004290643556411491,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4739127232978033,
+      "tokens_seen": 1898184704
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004289841116995667,
+      "loss": 0.0656,
+      "theoretical_loss": 3.473872644837189,
+      "tokens_seen": 1898446848
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004289038677579843,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4738325734596858,
+      "tokens_seen": 1898708992
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004288236238164019,
+      "loss": 0.0652,
+      "theoretical_loss": 3.473792509163066,
+      "tokens_seen": 1898971136
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042874337987481946,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4737524519450993,
+      "tokens_seen": 1899233280
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.000428663135933237,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4737124018035597,
+      "tokens_seen": 1899495424
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042858289199165464,
+      "loss": 0.0673,
+      "theoretical_loss": 3.47367235873622,
+      "tokens_seen": 1899757568
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004285026480500722,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4736323227408548,
+      "tokens_seen": 1900019712
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004284224041084898,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4735922938152397,
+      "tokens_seen": 1900281856
+    },
+    {
+      "epoch": 0.58,
+      "objective/train/advantage_avg": -0.00030587660148739815,
+      "objective/train/docs_used": 693007,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.30275297164917,
+      "objective/train/original_loss": 1.302753210067749,
+      "objective/train/theoretical_loss": 3.4735522719571517,
+      "objective/train/tokens_used": 1921004000,
+      "objective/train/value_avg": -0.006526947021484375,
+      "objective/train/value_loss": 0.00030584644991904497,
+      "objective/train/value_max": -2.8192996978759766e-05,
+      "objective/train/value_min": -0.94970703125,
+      "objective/train/value_reward_corr": 0.6970315136260118,
+      "objective/train/value_std": 0.01509857177734375,
+      "objective/train/weight_avg": 0.9998296499252319,
+      "objective/train/weighted_lm_loss": 1.3028963804244995,
+      "objective/train/weights_max": 1.5824660062789917,
+      "objective/train/weights_min": 0.3694700598716736,
+      "theoretical_loss": 3.4735522719571517,
+      "tokens_seen": 1900544000
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004283421601669074,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4735522719571517,
+      "tokens_seen": 1900544000
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042826191622532503,
+      "loss": 0.0669,
+      "theoretical_loss": 3.473512257164368,
+      "tokens_seen": 1900806144
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004281816722837426,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4734722494346673,
+      "tokens_seen": 1901068288
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042810142834216015,
+      "loss": 0.0683,
+      "theoretical_loss": 3.473432248765829,
+      "tokens_seen": 1901330432
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042802118440057776,
+      "loss": 0.0692,
+      "theoretical_loss": 3.473392255155634,
+      "tokens_seen": 1901592576
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004279409404589953,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4733522686018636,
+      "tokens_seen": 1901854720
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042786069651741293,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4733122891023007,
+      "tokens_seen": 1902116864
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042778045257583054,
+      "loss": 0.0679,
+      "theoretical_loss": 3.473272316654729,
+      "tokens_seen": 1902379008
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042770020863424816,
+      "loss": 0.0709,
+      "theoretical_loss": 3.4732323512569323,
+      "tokens_seen": 1902641152
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004276199646926657,
+      "loss": 0.0677,
+      "theoretical_loss": 3.473192392906697,
+      "tokens_seen": 1902903296
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042753972075108333,
+      "loss": 0.0688,
+      "theoretical_loss": 3.473152441601809,
+      "tokens_seen": 1903165440
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004274594768095009,
+      "loss": 0.0721,
+      "theoretical_loss": 3.473112497340057,
+      "tokens_seen": 1903427584
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042737923286791845,
+      "loss": 0.0684,
+      "theoretical_loss": 3.473072560119229,
+      "tokens_seen": 1903689728
+    },
+    {
+      "epoch": 0.58,
+      "objective/train/advantage_avg": 0.002401332836598158,
+      "objective/train/docs_used": 694210,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3564651012420654,
+      "objective/train/original_loss": 1.356465220451355,
+      "objective/train/theoretical_loss": 3.4730525941484705,
+      "objective/train/tokens_used": 1924280800,
+      "objective/train/value_avg": -0.00836181640625,
+      "objective/train/value_loss": 0.0003155863960273564,
+      "objective/train/value_max": -3.5643577575683594e-05,
+      "objective/train/value_min": -0.505859375,
+      "objective/train/value_reward_corr": 0.528134268748609,
+      "objective/train/value_std": 0.01629638671875,
+      "objective/train/weight_avg": 1.002550482749939,
+      "objective/train/weighted_lm_loss": 1.359156608581543,
+      "objective/train/weights_max": 1.6584101915359497,
+      "objective/train/weights_min": 0.3761554956436157,
+      "theoretical_loss": 3.4730525941484705,
+      "tokens_seen": 1903820800
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042729898892633606,
+      "loss": 0.0699,
+      "theoretical_loss": 3.473032629937114,
+      "tokens_seen": 1903951872
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042721874498475367,
+      "loss": 0.0679,
+      "theoretical_loss": 3.472992706791504,
+      "tokens_seen": 1904214016
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004271385010431713,
+      "loss": 0.0686,
+      "theoretical_loss": 3.472952790680189,
+      "tokens_seen": 1904476160
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042705825710158884,
+      "loss": 0.0688,
+      "theoretical_loss": 3.472912881600963,
+      "tokens_seen": 1904738304
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042697801316000645,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4728729795516182,
+      "tokens_seen": 1905000448
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.000426897769218424,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4728330845299507,
+      "tokens_seen": 1905262592
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042681752527684157,
+      "loss": 0.0692,
+      "theoretical_loss": 3.472793196533755,
+      "tokens_seen": 1905524736
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004267372813352592,
+      "loss": 0.0705,
+      "theoretical_loss": 3.472753315560828,
+      "tokens_seen": 1905786880
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004266570373936768,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4727134416089678,
+      "tokens_seen": 1906049024
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042657679345209435,
+      "loss": 0.0712,
+      "theoretical_loss": 3.472673574675972,
+      "tokens_seen": 1906311168
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042649654951051197,
+      "loss": 0.0667,
+      "theoretical_loss": 3.472633714759641,
+      "tokens_seen": 1906573312
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004264163055689296,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4725938618577743,
+      "tokens_seen": 1906835456
+    },
+    {
+      "epoch": 0.58,
+      "objective/train/advantage_avg": 0.0019277179380878806,
+      "objective/train/docs_used": 695522,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3308924436569214,
+      "objective/train/original_loss": 1.3308923244476318,
+      "objective/train/theoretical_loss": 3.4725540159681745,
+      "objective/train/tokens_used": 1927557600,
+      "objective/train/value_avg": -0.006687164306640625,
+      "objective/train/value_loss": 0.00017574307275936007,
+      "objective/train/value_max": -3.510713577270508e-05,
+      "objective/train/value_min": -0.62060546875,
+      "objective/train/value_reward_corr": 0.7532226507887959,
+      "objective/train/value_std": 0.016326904296875,
+      "objective/train/weight_avg": 1.0020089149475098,
+      "objective/train/weighted_lm_loss": 1.3338085412979126,
+      "objective/train/weights_max": 1.386031150817871,
+      "objective/train/weights_min": 0.36823785305023193,
+      "theoretical_loss": 3.4725540159681745,
+      "tokens_seen": 1907097600
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004263360616273472,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4725540159681745,
+      "tokens_seen": 1907097600
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004262558176857647,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4725141770886436,
+      "tokens_seen": 1907359744
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004261755737441823,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4724743452169857,
+      "tokens_seen": 1907621888
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004260953298025999,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4724345203510047,
+      "tokens_seen": 1907884032
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004260150858610175,
+      "loss": 0.0679,
+      "theoretical_loss": 3.472394702488506,
+      "tokens_seen": 1908146176
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004259348419194351,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4723548916272966,
+      "tokens_seen": 1908408320
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004258545979778527,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4723150877651836,
+      "tokens_seen": 1908670464
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004257743540362703,
+      "loss": 0.068,
+      "theoretical_loss": 3.472275290899976,
+      "tokens_seen": 1908932608
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004256941100946878,
+      "loss": 0.069,
+      "theoretical_loss": 3.472235501029483,
+      "tokens_seen": 1909194752
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042561386615310543,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4721957181515144,
+      "tokens_seen": 1909456896
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042553362221152305,
+      "loss": 0.0695,
+      "theoretical_loss": 3.472155942263883,
+      "tokens_seen": 1909719040
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004254533782699406,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4721161733643995,
+      "tokens_seen": 1909981184
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004253731343283582,
+      "loss": 0.0674,
+      "theoretical_loss": 3.472076411450878,
+      "tokens_seen": 1910243328
+    },
+    {
+      "epoch": 0.58,
+      "objective/train/advantage_avg": 0.0001537144707981497,
+      "objective/train/docs_used": 696705,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2774269580841064,
+      "objective/train/original_loss": 1.2774269580841064,
+      "objective/train/theoretical_loss": 3.4720565331131708,
+      "objective/train/tokens_used": 1930834400,
+      "objective/train/value_avg": -0.008636474609375,
+      "objective/train/value_loss": 0.00040704250568524003,
+      "objective/train/value_max": -3.534555435180664e-05,
+      "objective/train/value_min": -0.73388671875,
+      "objective/train/value_reward_corr": 0.7005505863181416,
+      "objective/train/value_std": 0.01953125,
+      "objective/train/weight_avg": 1.0003374814987183,
+      "objective/train/weighted_lm_loss": 1.2776480913162231,
+      "objective/train/weights_max": 1.9289265871047974,
+      "objective/train/weights_min": 0.3710477352142334,
+      "theoretical_loss": 3.4720565331131708,
+      "tokens_seen": 1910374400
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042529289038677583,
+      "loss": 0.0678,
+      "theoretical_loss": 3.472036656521134,
+      "tokens_seen": 1910505472
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042521264644519344,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4719969085729816,
+      "tokens_seen": 1910767616
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042513240250361095,
+      "loss": 0.0715,
+      "theoretical_loss": 3.4719571676042373,
+      "tokens_seen": 1911029760
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042505215856202856,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4719174336127185,
+      "tokens_seen": 1911291904
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004249719146204462,
+      "loss": 0.0691,
+      "theoretical_loss": 3.471877706596244,
+      "tokens_seen": 1911554048
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042489167067886373,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4718379865526323,
+      "tokens_seen": 1911816192
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042481142673728134,
+      "loss": 0.0704,
+      "theoretical_loss": 3.4717982734797044,
+      "tokens_seen": 1912078336
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042473118279569896,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4717585673752813,
+      "tokens_seen": 1912340480
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004246509388541165,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4717188682371853,
+      "tokens_seen": 1912602624
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042457069491253413,
+      "loss": 0.0678,
+      "theoretical_loss": 3.471679176063239,
+      "tokens_seen": 1912864768
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004244904509709517,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4716394908512678,
+      "tokens_seen": 1913126912
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004244102070293693,
+      "loss": 0.0659,
+      "theoretical_loss": 3.471599812599095,
+      "tokens_seen": 1913389056
+    },
+    {
+      "epoch": 0.58,
+      "objective/train/advantage_avg": -5.6982433306984603e-05,
+      "objective/train/docs_used": 697951,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4868850708007812,
+      "objective/train/original_loss": 1.4868848323822021,
+      "objective/train/theoretical_loss": 3.4715601413045487,
+      "objective/train/tokens_used": 1934111200,
+      "objective/train/value_avg": -0.00873565673828125,
+      "objective/train/value_loss": 0.00032236077822744846,
+      "objective/train/value_max": -2.586841583251953e-05,
+      "objective/train/value_min": -0.73828125,
+      "objective/train/value_reward_corr": 0.5984469149683921,
+      "objective/train/value_std": 0.0132904052734375,
+      "objective/train/weight_avg": 1.0000770092010498,
+      "objective/train/weighted_lm_loss": 1.4859241247177124,
+      "objective/train/weights_max": 1.268136978149414,
+      "objective/train/weights_min": 0.22390565276145935,
+      "theoretical_loss": 3.4715601413045487,
+      "tokens_seen": 1913651200
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042432996308778686,
+      "loss": 0.0713,
+      "theoretical_loss": 3.4715601413045487,
+      "tokens_seen": 1913651200
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042424971914620447,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4715204769654555,
+      "tokens_seen": 1913913344
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004241694752046221,
+      "loss": 0.0693,
+      "theoretical_loss": 3.471480819579643,
+      "tokens_seen": 1914175488
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042408923126303964,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4714411691449403,
+      "tokens_seen": 1914437632
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042400898732145725,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4714015256591777,
+      "tokens_seen": 1914699776
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004239287433798748,
+      "loss": 0.0715,
+      "theoretical_loss": 3.4713618891201863,
+      "tokens_seen": 1914961920
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004238484994382924,
+      "loss": 0.0693,
+      "theoretical_loss": 3.471322259525798,
+      "tokens_seen": 1915224064
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042376825549671,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4712826368738456,
+      "tokens_seen": 1915486208
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004236880115551276,
+      "loss": 0.0684,
+      "theoretical_loss": 3.471243021162163,
+      "tokens_seen": 1915748352
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004236077676135452,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4712034123885855,
+      "tokens_seen": 1916010496
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042352752367196277,
+      "loss": 0.066,
+      "theoretical_loss": 3.471163810550949,
+      "tokens_seen": 1916272640
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004234472797303804,
+      "loss": 0.0695,
+      "theoretical_loss": 3.47112421564709,
+      "tokens_seen": 1916534784
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042336703578879794,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4710846276748466,
+      "tokens_seen": 1916796928
+    },
+    {
+      "epoch": 0.58,
+      "objective/train/advantage_avg": 0.0002990372304338962,
+      "objective/train/docs_used": 699035,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3075224161148071,
+      "objective/train/original_loss": 1.3075222969055176,
+      "objective/train/theoretical_loss": 3.471064836287405,
+      "objective/train/tokens_used": 1937388000,
+      "objective/train/value_avg": -0.006961822509765625,
+      "objective/train/value_loss": 0.000213931969483383,
+      "objective/train/value_max": -4.100799560546875e-05,
+      "objective/train/value_min": -0.340087890625,
+      "objective/train/value_reward_corr": 0.5923001788406644,
+      "objective/train/value_std": 0.01080322265625,
+      "objective/train/weight_avg": 1.0003918409347534,
+      "objective/train/weighted_lm_loss": 1.3085472583770752,
+      "objective/train/weights_max": 1.317122220993042,
+      "objective/train/weights_min": 0.38067081570625305,
+      "theoretical_loss": 3.471064836287405,
+      "tokens_seen": 1916928000
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042328679184721555,
+      "loss": 0.068,
+      "theoretical_loss": 3.4710450466320575,
+      "tokens_seen": 1917059072
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004232065479056331,
+      "loss": 0.0678,
+      "theoretical_loss": 3.471005472516562,
+      "tokens_seen": 1917321216
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004231263039640507,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4709659053262016,
+      "tokens_seen": 1917583360
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042304606002246833,
+      "loss": 0.0713,
+      "theoretical_loss": 3.470926345058818,
+      "tokens_seen": 1917845504
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004229658160808859,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4708867917122532,
+      "tokens_seen": 1918107648
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004228855721393035,
+      "loss": 0.0695,
+      "theoretical_loss": 3.470847245284351,
+      "tokens_seen": 1918369792
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004228053281977211,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4708077057729567,
+      "tokens_seen": 1918631936
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004227250842561386,
+      "loss": 0.069,
+      "theoretical_loss": 3.4707681731759155,
+      "tokens_seen": 1918894080
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042264484031455623,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4707286474910735,
+      "tokens_seen": 1919156224
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042256459637297385,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4706891287162787,
+      "tokens_seen": 1919418368
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042248435243139146,
+      "loss": 0.069,
+      "theoretical_loss": 3.470649616849379,
+      "tokens_seen": 1919680512
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.000422404108489809,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4706101118882247,
+      "tokens_seen": 1919942656
+    },
+    {
+      "epoch": 0.58,
+      "objective/train/advantage_avg": 0.0010100979125127196,
+      "objective/train/docs_used": 700307,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4696931838989258,
+      "objective/train/original_loss": 1.4696931838989258,
+      "objective/train/theoretical_loss": 3.4705706138306653,
+      "objective/train/tokens_used": 1940664800,
+      "objective/train/value_avg": -0.004924774169921875,
+      "objective/train/value_loss": 0.0002015416685026139,
+      "objective/train/value_max": -4.57763671875e-05,
+      "objective/train/value_min": -0.230712890625,
+      "objective/train/value_reward_corr": 0.49469162323894733,
+      "objective/train/value_std": 0.00820159912109375,
+      "objective/train/weight_avg": 1.0010887384414673,
+      "objective/train/weighted_lm_loss": 1.4720734357833862,
+      "objective/train/weights_max": 1.1817450523376465,
+      "objective/train/weights_min": 0.22504496574401855,
+      "theoretical_loss": 3.4705706138306653,
+      "tokens_seen": 1920204800
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042232386454822663,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4705706138306653,
+      "tokens_seen": 1920204800
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042224362060664424,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4705311226745525,
+      "tokens_seen": 1920466944
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042216337666506175,
+      "loss": 0.0678,
+      "theoretical_loss": 3.470491638417739,
+      "tokens_seen": 1920729088
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042208313272347936,
+      "loss": 0.0665,
+      "theoretical_loss": 3.470452161058078,
+      "tokens_seen": 1920991232
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042200288878189697,
+      "loss": 0.0695,
+      "theoretical_loss": 3.470412690593423,
+      "tokens_seen": 1921253376
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004219226448403146,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4703732270216303,
+      "tokens_seen": 1921515520
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042184240089873214,
+      "loss": 0.067,
+      "theoretical_loss": 3.470333770340555,
+      "tokens_seen": 1921777664
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042176215695714976,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4702943205480548,
+      "tokens_seen": 1922039808
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042168191301556737,
+      "loss": 0.0678,
+      "theoretical_loss": 3.470254877641988,
+      "tokens_seen": 1922301952
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042160166907398493,
+      "loss": 0.0687,
+      "theoretical_loss": 3.470215441620213,
+      "tokens_seen": 1922564096
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004215214251324025,
+      "loss": 0.0714,
+      "theoretical_loss": 3.47017601248059,
+      "tokens_seen": 1922826240
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004214411811908201,
+      "loss": 0.068,
+      "theoretical_loss": 3.4701365902209798,
+      "tokens_seen": 1923088384
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004213609372492377,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4700971748392453,
+      "tokens_seen": 1923350528
+    },
+    {
+      "epoch": 0.58,
+      "objective/train/advantage_avg": 4.2215957364533097e-05,
+      "objective/train/docs_used": 701572,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3477340936660767,
+      "objective/train/original_loss": 1.3477338552474976,
+      "objective/train/theoretical_loss": 3.4700774697269128,
+      "objective/train/tokens_used": 1943941600,
+      "objective/train/value_avg": -0.0045928955078125,
+      "objective/train/value_loss": 0.0001257354742847383,
+      "objective/train/value_max": -3.3736228942871094e-05,
+      "objective/train/value_min": -0.418701171875,
+      "objective/train/value_reward_corr": 0.56594521479194,
+      "objective/train/value_std": 0.007476806640625,
+      "objective/train/weight_avg": 1.000099539756775,
+      "objective/train/weighted_lm_loss": 1.3483197689056396,
+      "objective/train/weights_max": 1.2155424356460571,
+      "objective/train/weights_min": 0.3688117563724518,
+      "theoretical_loss": 3.4700774697269128,
+      "tokens_seen": 1923481600
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042128069330765527,
+      "loss": 0.0677,
+      "theoretical_loss": 3.470057766333248,
+      "tokens_seen": 1923612672
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004212004493660729,
+      "loss": 0.0706,
+      "theoretical_loss": 3.4700183647008522,
+      "tokens_seen": 1923874816
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004211202054244905,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4699789699399233,
+      "tokens_seen": 1924136960
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042103996148290805,
+      "loss": 0.0659,
+      "theoretical_loss": 3.469939582048326,
+      "tokens_seen": 1924399104
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004209597175413256,
+      "loss": 0.0676,
+      "theoretical_loss": 3.469900201023928,
+      "tokens_seen": 1924661248
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004208794735997432,
+      "loss": 0.065,
+      "theoretical_loss": 3.469860826864596,
+      "tokens_seen": 1924923392
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004207992296581608,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4698214595681995,
+      "tokens_seen": 1925185536
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004207189857165784,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4697820991326074,
+      "tokens_seen": 1925447680
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.000420638741774996,
+      "loss": 0.0658,
+      "theoretical_loss": 3.46974274555569,
+      "tokens_seen": 1925709824
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004205584978334136,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4697033988353194,
+      "tokens_seen": 1925971968
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004204782538918312,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4696640589693675,
+      "tokens_seen": 1926234112
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042039800995024874,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4696247259557076,
+      "tokens_seen": 1926496256
+    },
+    {
+      "epoch": 0.58,
+      "objective/train/advantage_avg": 0.0005393868195824325,
+      "objective/train/docs_used": 702709,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5641299486160278,
+      "objective/train/original_loss": 1.5641303062438965,
+      "objective/train/theoretical_loss": 3.469585399792215,
+      "objective/train/tokens_used": 1947218400,
+      "objective/train/value_avg": -0.01006317138671875,
+      "objective/train/value_loss": 0.0004965075058862567,
+      "objective/train/value_max": -3.24249267578125e-05,
+      "objective/train/value_min": -0.732421875,
+      "objective/train/value_reward_corr": 0.7204884420073756,
+      "objective/train/value_std": 0.0225067138671875,
+      "objective/train/weight_avg": 1.0007641315460205,
+      "objective/train/weighted_lm_loss": 1.5642176866531372,
+      "objective/train/weights_max": 1.5964356660842896,
+      "objective/train/weights_min": 0.3893653154373169,
+      "theoretical_loss": 3.469585399792215,
+      "tokens_seen": 1926758400
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042031776600866635,
+      "loss": 0.0706,
+      "theoretical_loss": 3.469585399792215,
+      "tokens_seen": 1926758400
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004202375220670839,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4695460804767633,
+      "tokens_seen": 1927020544
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004201572781255015,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4695067680072293,
+      "tokens_seen": 1927282688
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00042007703418391913,
+      "loss": 0.0708,
+      "theoretical_loss": 3.469467462381491,
+      "tokens_seen": 1927544832
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00041999679024233675,
+      "loss": 0.0682,
+      "theoretical_loss": 3.469428163597426,
+      "tokens_seen": 1927806976
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004199165463007543,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4693888716529124,
+      "tokens_seen": 1928069120
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004198363023591719,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4693495865458313,
+      "tokens_seen": 1928331264
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.0004197560584175895,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4693103082740633,
+      "tokens_seen": 1928593408
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00041967581447600703,
+      "loss": 0.0718,
+      "theoretical_loss": 3.4692710368354898,
+      "tokens_seen": 1928855552
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00041959557053442465,
+      "loss": 0.0683,
+      "theoretical_loss": 3.469231772227994,
+      "tokens_seen": 1929117696
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00041951532659284226,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4691925144494604,
+      "tokens_seen": 1929379840
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00041943508265125987,
+      "loss": 0.0716,
+      "theoretical_loss": 3.4691532634977724,
+      "tokens_seen": 1929641984
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00041935483870967743,
+      "loss": 0.069,
+      "theoretical_loss": 3.469114019370816,
+      "tokens_seen": 1929904128
+    },
+    {
+      "epoch": 0.58,
+      "objective/train/advantage_avg": 0.0007605802384205163,
+      "objective/train/docs_used": 703903,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4023168087005615,
+      "objective/train/original_loss": 1.4023168087005615,
+      "objective/train/theoretical_loss": 3.4690943998659516,
+      "objective/train/tokens_used": 1950495200,
+      "objective/train/value_avg": -0.00800323486328125,
+      "objective/train/value_loss": 0.00025014992570504546,
+      "objective/train/value_max": -4.684925079345703e-05,
+      "objective/train/value_min": -0.84228515625,
+      "objective/train/value_reward_corr": 0.686540956128751,
+      "objective/train/value_std": 0.0177459716796875,
+      "objective/train/weight_avg": 1.0008822679519653,
+      "objective/train/weighted_lm_loss": 1.4031236171722412,
+      "objective/train/weights_max": 2.208437204360962,
+      "objective/train/weights_min": 0.37882235646247864,
+      "theoretical_loss": 3.4690943998659516,
+      "tokens_seen": 1930035200
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 0.00041927459476809504,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4690747820664782,
+      "tokens_seen": 1930166272
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004191943508265126,
+      "loss": 0.0673,
+      "theoretical_loss": 3.469035551582646,
+      "tokens_seen": 1930428416
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041911410688493016,
+      "loss": 0.068,
+      "theoretical_loss": 3.468996327917208,
+      "tokens_seen": 1930690560
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041903386294334777,
+      "loss": 0.0695,
+      "theoretical_loss": 3.468957111068054,
+      "tokens_seen": 1930952704
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004189536190017654,
+      "loss": 0.067,
+      "theoretical_loss": 3.468917901033074,
+      "tokens_seen": 1931214848
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.000418873375060183,
+      "loss": 0.0695,
+      "theoretical_loss": 3.468878697810159,
+      "tokens_seen": 1931476992
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041879313111860056,
+      "loss": 0.0699,
+      "theoretical_loss": 3.468839501397202,
+      "tokens_seen": 1931739136
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041871288717701817,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4688003117920956,
+      "tokens_seen": 1932001280
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004186326432354357,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4687611289927336,
+      "tokens_seen": 1932263424
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004185523992938533,
+      "loss": 0.0683,
+      "theoretical_loss": 3.468721952997012,
+      "tokens_seen": 1932525568
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004184721553522709,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4686827838028256,
+      "tokens_seen": 1932787712
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004183919114106885,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4686436214080727,
+      "tokens_seen": 1933049856
+    },
+    {
+      "epoch": 0.59,
+      "objective/train/advantage_avg": 0.00045794775360263884,
+      "objective/train/docs_used": 705162,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3477123975753784,
+      "objective/train/original_loss": 1.3477123975753784,
+      "objective/train/theoretical_loss": 3.4686044658106496,
+      "objective/train/tokens_used": 1953772000,
+      "objective/train/value_avg": -0.0085296630859375,
+      "objective/train/value_loss": 0.00025771657237783074,
+      "objective/train/value_max": -6.401538848876953e-05,
+      "objective/train/value_min": -0.287841796875,
+      "objective/train/value_reward_corr": 0.7388518675620712,
+      "objective/train/value_std": 0.016693115234375,
+      "objective/train/weight_avg": 1.000576376914978,
+      "objective/train/weighted_lm_loss": 1.3479864597320557,
+      "objective/train/weights_max": 1.1910806894302368,
+      "objective/train/weights_min": 0.36963358521461487,
+      "theoretical_loss": 3.4686044658106496,
+      "tokens_seen": 1933312000
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041831166746910607,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4686044658106496,
+      "tokens_seen": 1933312000
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004182314235275237,
+      "loss": 0.0669,
+      "theoretical_loss": 3.468565317008456,
+      "tokens_seen": 1933574144
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004181511795859413,
+      "loss": 0.0713,
+      "theoretical_loss": 3.4685261749993908,
+      "tokens_seen": 1933836288
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004180709356443589,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4684870397813556,
+      "tokens_seen": 1934098432
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004179906917027764,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4684479113522517,
+      "tokens_seen": 1934360576
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.000417910447761194,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4684087897099816,
+      "tokens_seen": 1934622720
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041783020381961164,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4683696748524486,
+      "tokens_seen": 1934884864
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004177499598780292,
+      "loss": 0.0709,
+      "theoretical_loss": 3.468330566777557,
+      "tokens_seen": 1935147008
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004176697159364468,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4682914654832118,
+      "tokens_seen": 1935409152
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004175894719948644,
+      "loss": 0.0651,
+      "theoretical_loss": 3.46825237096732,
+      "tokens_seen": 1935671296
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041750922805328203,
+      "loss": 0.067,
+      "theoretical_loss": 3.4682132832277883,
+      "tokens_seen": 1935933440
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041742898411169954,
+      "loss": 0.069,
+      "theoretical_loss": 3.4681742022625253,
+      "tokens_seen": 1936195584
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041734874017011715,
+      "loss": 0.0737,
+      "theoretical_loss": 3.468135128069439,
+      "tokens_seen": 1936457728
+    },
+    {
+      "epoch": 0.59,
+      "objective/train/advantage_avg": 0.0009253130410797894,
+      "objective/train/docs_used": 706032,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.303908348083496,
+      "objective/train/original_loss": 1.303908109664917,
+      "objective/train/theoretical_loss": 3.4681155935118095,
+      "objective/train/tokens_used": 1957048800,
+      "objective/train/value_avg": -0.00952911376953125,
+      "objective/train/value_loss": 0.00012180649355286732,
+      "objective/train/value_max": -6.812810897827148e-05,
+      "objective/train/value_min": -0.23779296875,
+      "objective/train/value_reward_corr": 0.7694273438548433,
+      "objective/train/value_std": 0.01409149169921875,
+      "objective/train/weight_avg": 1.0009859800338745,
+      "objective/train/weighted_lm_loss": 1.3054463863372803,
+      "objective/train/weights_max": 1.2155041694641113,
+      "objective/train/weights_min": 0.8359750509262085,
+      "theoretical_loss": 3.4681155935118095,
+      "tokens_seen": 1936588800
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041726849622853476,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4680960606464404,
+      "tokens_seen": 1936719872
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004171882522869523,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4680569999914397,
+      "tokens_seen": 1936982016
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041710800834536993,
+      "loss": 0.0678,
+      "theoretical_loss": 3.468017946102349,
+      "tokens_seen": 1937244160
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041702776440378755,
+      "loss": 0.0692,
+      "theoretical_loss": 3.467978898977081,
+      "tokens_seen": 1937506304
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041694752046220516,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4679398586135486,
+      "tokens_seen": 1937768448
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004168672765206227,
+      "loss": 0.0678,
+      "theoretical_loss": 3.467900825009668,
+      "tokens_seen": 1938030592
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004167870325790403,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4678617981633533,
+      "tokens_seen": 1938292736
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004167067886374579,
+      "loss": 0.0682,
+      "theoretical_loss": 3.467822778072521,
+      "tokens_seen": 1938554880
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041662654469587545,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4677837647350898,
+      "tokens_seen": 1938817024
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041654630075429306,
+      "loss": 0.0681,
+      "theoretical_loss": 3.467744758148976,
+      "tokens_seen": 1939079168
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041646605681271067,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4677057583121007,
+      "tokens_seen": 1939341312
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041638581287112823,
+      "loss": 0.0688,
+      "theoretical_loss": 3.467666765222383,
+      "tokens_seen": 1939603456
+    },
+    {
+      "epoch": 0.59,
+      "objective/train/advantage_avg": 0.00017141261196229607,
+      "objective/train/docs_used": 707259,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3940285444259644,
+      "objective/train/original_loss": 1.3940284252166748,
+      "objective/train/theoretical_loss": 3.4676277788777443,
+      "objective/train/tokens_used": 1960325600,
+      "objective/train/value_avg": -0.009735107421875,
+      "objective/train/value_loss": 0.00045788646093569696,
+      "objective/train/value_max": -4.035234451293945e-05,
+      "objective/train/value_min": -0.48828125,
+      "objective/train/value_reward_corr": 0.5960321884519733,
+      "objective/train/value_std": 0.0169525146484375,
+      "objective/train/weight_avg": 1.0003608465194702,
+      "objective/train/weighted_lm_loss": 1.394312858581543,
+      "objective/train/weights_max": 1.5537357330322266,
+      "objective/train/weights_min": 0.23123404383659363,
+      "theoretical_loss": 3.4676277788777443,
+      "tokens_seen": 1939865600
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041630556892954584,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4676277788777443,
+      "tokens_seen": 1939865600
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004162253249879634,
+      "loss": 0.0683,
+      "theoretical_loss": 3.467588799276106,
+      "tokens_seen": 1940127744
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.000416145081046381,
+      "loss": 0.068,
+      "theoretical_loss": 3.4675498264153912,
+      "tokens_seen": 1940389888
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041606483710479857,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4675108602935243,
+      "tokens_seen": 1940652032
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004159845931632162,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4674719009084294,
+      "tokens_seen": 1940914176
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004159043492216338,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4674329482580326,
+      "tokens_seen": 1941176320
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041582410528005135,
+      "loss": 0.0666,
+      "theoretical_loss": 3.46739400234026,
+      "tokens_seen": 1941438464
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041574386133846897,
+      "loss": 0.0705,
+      "theoretical_loss": 3.467355063153039,
+      "tokens_seen": 1941700608
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004156636173968865,
+      "loss": 0.0647,
+      "theoretical_loss": 3.467316130694299,
+      "tokens_seen": 1941962752
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041558337345530414,
+      "loss": 0.0664,
+      "theoretical_loss": 3.467277204961968,
+      "tokens_seen": 1942224896
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004155031295137217,
+      "loss": 0.067,
+      "theoretical_loss": 3.467238285953977,
+      "tokens_seen": 1942487040
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004154228855721393,
+      "loss": 0.0696,
+      "theoretical_loss": 3.467199373668257,
+      "tokens_seen": 1942749184
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004153426416305569,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4671604681027404,
+      "tokens_seen": 1943011328
+    },
+    {
+      "epoch": 0.59,
+      "objective/train/advantage_avg": 0.0002338496851734817,
+      "objective/train/docs_used": 708379,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3313257694244385,
+      "objective/train/original_loss": 1.3313257694244385,
+      "objective/train/theoretical_loss": 3.4671410178394115,
+      "objective/train/tokens_used": 1963602400,
+      "objective/train/value_avg": -0.0081024169921875,
+      "objective/train/value_loss": 0.0004016077728010714,
+      "objective/train/value_max": -6.866455078125e-05,
+      "objective/train/value_min": -0.88134765625,
+      "objective/train/value_reward_corr": 0.6732135705871113,
+      "objective/train/value_std": 0.0167083740234375,
+      "objective/train/weight_avg": 1.000407338142395,
+      "objective/train/weighted_lm_loss": 1.3311095237731934,
+      "objective/train/weights_max": 1.3798331022262573,
+      "objective/train/weights_min": 0.3778006136417389,
+      "theoretical_loss": 3.4671410178394115,
+      "tokens_seen": 1943142400
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004152623976889745,
+      "loss": 0.0706,
+      "theoretical_loss": 3.467121569255359,
+      "tokens_seen": 1943273472
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004151821537473921,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4670826771240484,
+      "tokens_seen": 1943535616
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004151019098058097,
+      "loss": 0.071,
+      "theoretical_loss": 3.4670437917067423,
+      "tokens_seen": 1943797760
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041502166586422726,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4670049130013765,
+      "tokens_seen": 1944059904
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004149414219226448,
+      "loss": 0.0657,
+      "theoretical_loss": 3.466966041005888,
+      "tokens_seen": 1944322048
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041486117798106244,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4669271757182143,
+      "tokens_seen": 1944584192
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041478093403948005,
+      "loss": 0.0688,
+      "theoretical_loss": 3.466888317136293,
+      "tokens_seen": 1944846336
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004147006900978976,
+      "loss": 0.0678,
+      "theoretical_loss": 3.466849465258065,
+      "tokens_seen": 1945108480
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004146204461563152,
+      "loss": 0.068,
+      "theoretical_loss": 3.4668106200814695,
+      "tokens_seen": 1945370624
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041454020221473283,
+      "loss": 0.0677,
+      "theoretical_loss": 3.466771781604448,
+      "tokens_seen": 1945632768
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041445995827315034,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4667329498249426,
+      "tokens_seen": 1945894912
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041437971433156795,
+      "loss": 0.0673,
+      "theoretical_loss": 3.466694124740896,
+      "tokens_seen": 1946157056
+    },
+    {
+      "epoch": 0.59,
+      "objective/train/advantage_avg": 0.0006498626316897571,
+      "objective/train/docs_used": 709628,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3628689050674438,
+      "objective/train/original_loss": 1.3628687858581543,
+      "objective/train/theoretical_loss": 3.466655306350253,
+      "objective/train/tokens_used": 1966879200,
+      "objective/train/value_avg": -0.00696563720703125,
+      "objective/train/value_loss": 0.0002535603125579655,
+      "objective/train/value_max": -5.829334259033203e-05,
+      "objective/train/value_min": -0.6953125,
+      "objective/train/value_reward_corr": 0.7165104923699819,
+      "objective/train/value_std": 0.017547607421875,
+      "objective/train/weight_avg": 1.0007662773132324,
+      "objective/train/weighted_lm_loss": 1.3636950254440308,
+      "objective/train/weights_max": 1.7379988431930542,
+      "objective/train/weights_min": 0.38149651885032654,
+      "theoretical_loss": 3.466655306350253,
+      "tokens_seen": 1946419200
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041429947038998556,
+      "loss": 0.0681,
+      "theoretical_loss": 3.466655306350253,
+      "tokens_seen": 1946419200
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004142192264484032,
+      "loss": 0.068,
+      "theoretical_loss": 3.4666164946509572,
+      "tokens_seen": 1946681344
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041413898250682073,
+      "loss": 0.0695,
+      "theoretical_loss": 3.466577689640955,
+      "tokens_seen": 1946943488
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041405873856523834,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4665388913181934,
+      "tokens_seen": 1947205632
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041397849462365596,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4665000996806192,
+      "tokens_seen": 1947467776
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041389825068207346,
+      "loss": 0.0685,
+      "theoretical_loss": 3.466461314726182,
+      "tokens_seen": 1947729920
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004138180067404911,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4664225364528294,
+      "tokens_seen": 1947992064
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004137377627989087,
+      "loss": 0.0696,
+      "theoretical_loss": 3.466383764858513,
+      "tokens_seen": 1948254208
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004136575188573263,
+      "loss": 0.0689,
+      "theoretical_loss": 3.466344999941184,
+      "tokens_seen": 1948516352
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041357727491574386,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4663062416987938,
+      "tokens_seen": 1948778496
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041349703097416147,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4662674901292956,
+      "tokens_seen": 1949040640
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004134167870325791,
+      "loss": 0.0638,
+      "theoretical_loss": 3.4662287452306435,
+      "tokens_seen": 1949302784
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041333654309099664,
+      "loss": 0.0681,
+      "theoretical_loss": 3.466190007000792,
+      "tokens_seen": 1949564928
+    },
+    {
+      "epoch": 0.59,
+      "objective/train/advantage_avg": 0.0006561456248164177,
+      "objective/train/docs_used": 710798,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3777148723602295,
+      "objective/train/original_loss": 1.3777148723602295,
+      "objective/train/theoretical_loss": 3.4661706403860277,
+      "objective/train/tokens_used": 1970156000,
+      "objective/train/value_avg": -0.00855255126953125,
+      "objective/train/value_loss": 0.00019631710893008858,
+      "objective/train/value_max": -5.346536636352539e-05,
+      "objective/train/value_min": -0.379150390625,
+      "objective/train/value_reward_corr": 0.7235816315595381,
+      "objective/train/value_std": 0.0145263671875,
+      "objective/train/weight_avg": 1.0007461309432983,
+      "objective/train/weighted_lm_loss": 1.379103422164917,
+      "objective/train/weights_max": 1.2311376333236694,
+      "objective/train/weights_min": 0.3972630202770233,
+      "theoretical_loss": 3.4661706403860277,
+      "tokens_seen": 1949696000
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004132562991494142,
+      "loss": 0.0712,
+      "theoretical_loss": 3.466151275437697,
+      "tokens_seen": 1949827072
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004131760552078318,
+      "loss": 0.066,
+      "theoretical_loss": 3.4661125505393153,
+      "tokens_seen": 1950089216
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004130958112662494,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4660738323036036,
+      "tokens_seen": 1950351360
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.000413015567324667,
+      "loss": 0.0704,
+      "theoretical_loss": 3.466035120728521,
+      "tokens_seen": 1950613504
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004129353233830846,
+      "loss": 0.0669,
+      "theoretical_loss": 3.465996415812027,
+      "tokens_seen": 1950875648
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004128550794415022,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4659577175520813,
+      "tokens_seen": 1951137792
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041277483549991977,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4659190259466444,
+      "tokens_seen": 1951399936
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004126945915583373,
+      "loss": 0.0672,
+      "theoretical_loss": 3.46588034099368,
+      "tokens_seen": 1951662080
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041261434761675494,
+      "loss": 0.069,
+      "theoretical_loss": 3.465841662691149,
+      "tokens_seen": 1951924224
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004125341036751725,
+      "loss": 0.0673,
+      "theoretical_loss": 3.465802991037016,
+      "tokens_seen": 1952186368
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004124538597335901,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4657643260292463,
+      "tokens_seen": 1952448512
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004123736157920077,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4657256676658053,
+      "tokens_seen": 1952710656
+    },
+    {
+      "epoch": 0.59,
+      "objective/train/advantage_avg": 0.0003945839125663042,
+      "objective/train/docs_used": 712021,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.313441514968872,
+      "objective/train/original_loss": 1.3134416341781616,
+      "objective/train/theoretical_loss": 3.4656870159446584,
+      "objective/train/tokens_used": 1973432800,
+      "objective/train/value_avg": -0.00843048095703125,
+      "objective/train/value_loss": 0.0001438377657905221,
+      "objective/train/value_max": -4.988908767700195e-05,
+      "objective/train/value_min": -0.321044921875,
+      "objective/train/value_reward_corr": 0.715787808037206,
+      "objective/train/value_std": 0.0128021240234375,
+      "objective/train/weight_avg": 1.0004626512527466,
+      "objective/train/weighted_lm_loss": 1.313431739807129,
+      "objective/train/weights_max": 1.2099876403808594,
+      "objective/train/weights_min": 0.4170147180557251,
+      "theoretical_loss": 3.4656870159446584,
+      "tokens_seen": 1952972800
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041229337185042533,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4656870159446584,
+      "tokens_seen": 1952972800
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004122131279088429,
+      "loss": 0.0676,
+      "theoretical_loss": 3.465648370863774,
+      "tokens_seen": 1953234944
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004121328839672605,
+      "loss": 0.0663,
+      "theoretical_loss": 3.46560973242112,
+      "tokens_seen": 1953497088
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041205264002567806,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4655711006146657,
+      "tokens_seen": 1953759232
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004119723960840956,
+      "loss": 0.0703,
+      "theoretical_loss": 3.465532475442381,
+      "tokens_seen": 1954021376
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041189215214251323,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4654938569022375,
+      "tokens_seen": 1954283520
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041181190820093085,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4654552449922056,
+      "tokens_seen": 1954545664
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041173166425934846,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4654166397102593,
+      "tokens_seen": 1954807808
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.000411651420317766,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4653780410543717,
+      "tokens_seen": 1955069952
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041157117637618363,
+      "loss": 0.0688,
+      "theoretical_loss": 3.465339449022517,
+      "tokens_seen": 1955332096
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004114909324346012,
+      "loss": 0.0703,
+      "theoretical_loss": 3.4653008636126716,
+      "tokens_seen": 1955594240
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041141068849301875,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4652622848228107,
+      "tokens_seen": 1955856384
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041133044455143636,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4652237126509124,
+      "tokens_seen": 1956118528
+    },
+    {
+      "epoch": 0.59,
+      "objective/train/advantage_avg": -0.0006969102541916072,
+      "objective/train/docs_used": 713197,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3060723543167114,
+      "objective/train/original_loss": 1.3060722351074219,
+      "objective/train/theoretical_loss": 3.465204429046067,
+      "objective/train/tokens_used": 1976709600,
+      "objective/train/value_avg": -0.006160736083984375,
+      "objective/train/value_loss": 0.0003185392706654966,
+      "objective/train/value_max": -4.267692565917969e-05,
+      "objective/train/value_min": -0.393798828125,
+      "objective/train/value_reward_corr": 0.663370985817207,
+      "objective/train/value_std": 0.01360321044921875,
+      "objective/train/weight_avg": 0.9994413256645203,
+      "objective/train/weighted_lm_loss": 1.3050378561019897,
+      "objective/train/weights_max": 1.3730344772338867,
+      "objective/train/weights_min": 0.3702559173107147,
+      "theoretical_loss": 3.465204429046067,
+      "tokens_seen": 1956249600
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041125020060985397,
+      "loss": 0.0693,
+      "theoretical_loss": 3.465185147094954,
+      "tokens_seen": 1956380672
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004111699566682716,
+      "loss": 0.0694,
+      "theoretical_loss": 3.465146588152915,
+      "tokens_seen": 1956642816
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041108971272668914,
+      "loss": 0.068,
+      "theoretical_loss": 3.465108035822775,
+      "tokens_seen": 1956904960
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041100946878510676,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4650694901025147,
+      "tokens_seen": 1957167104
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004109292248435243,
+      "loss": 0.067,
+      "theoretical_loss": 3.4650309509901156,
+      "tokens_seen": 1957429248
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004108489809019419,
+      "loss": 0.0722,
+      "theoretical_loss": 3.4649924184835603,
+      "tokens_seen": 1957691392
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004107687369603595,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4649538925808328,
+      "tokens_seen": 1957953536
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004106884930187771,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4649153732799167,
+      "tokens_seen": 1958215680
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004106082490771947,
+      "loss": 0.0674,
+      "theoretical_loss": 3.464876860578797,
+      "tokens_seen": 1958477824
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00041052800513561227,
+      "loss": 0.0702,
+      "theoretical_loss": 3.46483835447546,
+      "tokens_seen": 1958739968
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004104477611940299,
+      "loss": 0.0676,
+      "theoretical_loss": 3.464799854967893,
+      "tokens_seen": 1959002112
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004103675172524475,
+      "loss": 0.0697,
+      "theoretical_loss": 3.464761362054084,
+      "tokens_seen": 1959264256
+    },
+    {
+      "epoch": 0.59,
+      "objective/train/advantage_avg": -0.0016641940455883741,
+      "objective/train/docs_used": 714425,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4409244060516357,
+      "objective/train/original_loss": 1.4409245252609253,
+      "objective/train/theoretical_loss": 3.4647228757320203,
+      "objective/train/tokens_used": 1979986400,
+      "objective/train/value_avg": -0.0134735107421875,
+      "objective/train/value_loss": 0.0007892985595390201,
+      "objective/train/value_max": -4.100799560546875e-05,
+      "objective/train/value_min": -0.98095703125,
+      "objective/train/value_reward_corr": 0.7946839372874429,
+      "objective/train/value_std": 0.033355712890625,
+      "objective/train/weight_avg": 0.9987018704414368,
+      "objective/train/weighted_lm_loss": 1.4399670362472534,
+      "objective/train/weights_max": 2.14656138420105,
+      "objective/train/weights_min": 0.3712770938873291,
+      "theoretical_loss": 3.4647228757320203,
+      "tokens_seen": 1959526400
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.000410287273310865,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4647228757320203,
+      "tokens_seen": 1959526400
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004102070293692826,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4646843959996927,
+      "tokens_seen": 1959788544
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004101267854277002,
+      "loss": 0.0676,
+      "theoretical_loss": 3.464645922855091,
+      "tokens_seen": 1960050688
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004100465414861178,
+      "loss": 0.0699,
+      "theoretical_loss": 3.464607456296207,
+      "tokens_seen": 1960312832
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004099662975445354,
+      "loss": 0.0708,
+      "theoretical_loss": 3.464568996321033,
+      "tokens_seen": 1960574976
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.000409886053602953,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4645305429275624,
+      "tokens_seen": 1960837120
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004098058096613706,
+      "loss": 0.0694,
+      "theoretical_loss": 3.464492096113788,
+      "tokens_seen": 1961099264
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004097255657197881,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4644536558777057,
+      "tokens_seen": 1961361408
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00040964532177820574,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4644152222173106,
+      "tokens_seen": 1961623552
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00040956507783662335,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4643767951305997,
+      "tokens_seen": 1961885696
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004094848338950409,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4643383746155703,
+      "tokens_seen": 1962147840
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.0004094045899534585,
+      "loss": 0.0677,
+      "theoretical_loss": 3.464299960670221,
+      "tokens_seen": 1962409984
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00040932434601187613,
+      "loss": 0.0692,
+      "theoretical_loss": 3.464261553292551,
+      "tokens_seen": 1962672128
+    },
+    {
+      "epoch": 0.59,
+      "objective/train/advantage_avg": 9.944574412656948e-05,
+      "objective/train/docs_used": 715564,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2323715686798096,
+      "objective/train/original_loss": 1.2323713302612305,
+      "objective/train/theoretical_loss": 3.464242352065971,
+      "objective/train/tokens_used": 1983263200,
+      "objective/train/value_avg": -0.0050811767578125,
+      "objective/train/value_loss": 0.00013055642193648964,
+      "objective/train/value_max": -4.13060188293457e-05,
+      "objective/train/value_min": -0.2900390625,
+      "objective/train/value_reward_corr": 0.657465772340293,
+      "objective/train/value_std": 0.00870513916015625,
+      "objective/train/weight_avg": 1.0001602172851562,
+      "objective/train/weighted_lm_loss": 1.2325642108917236,
+      "objective/train/weights_max": 1.0780946016311646,
+      "objective/train/weights_min": 0.3836330473423004,
+      "theoretical_loss": 3.464242352065971,
+      "tokens_seen": 1962803200
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00040924410207029375,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4642231524805607,
+      "tokens_seen": 1962934272
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 0.00040916385812871125,
+      "loss": 0.0704,
+      "theoretical_loss": 3.4641847582322507,
+      "tokens_seen": 1963196416
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040908361418712886,
+      "loss": 0.0711,
+      "theoretical_loss": 3.4641463705456226,
+      "tokens_seen": 1963458560
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004090033702455465,
+      "loss": 0.0704,
+      "theoretical_loss": 3.4641079894186797,
+      "tokens_seen": 1963720704
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040892312630396403,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4640696148494254,
+      "tokens_seen": 1963982848
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040884288236238165,
+      "loss": 0.068,
+      "theoretical_loss": 3.464031246835864,
+      "tokens_seen": 1964244992
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040876263842079926,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4639928853760016,
+      "tokens_seen": 1964507136
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040868239447921687,
+      "loss": 0.0689,
+      "theoretical_loss": 3.463954530467844,
+      "tokens_seen": 1964769280
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040860215053763443,
+      "loss": 0.0671,
+      "theoretical_loss": 3.463916182109398,
+      "tokens_seen": 1965031424
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.000408521906596052,
+      "loss": 0.0686,
+      "theoretical_loss": 3.463877840298672,
+      "tokens_seen": 1965293568
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004084416626544696,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4638395050336745,
+      "tokens_seen": 1965555712
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040836141871288716,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4638011763124164,
+      "tokens_seen": 1965817856
+    },
+    {
+      "debugging/Compilability": 1.0,
+      "debugging/distinct-1-grams": 0.7394139103740143,
+      "debugging/entropy-1-grams": 5.1329642330953185,
+      "debugging/length": 486.55555555555554,
+      "debugging/num_segments": 9,
+      "debugging/raw_token_scores_avg": 0.005027866922318935,
+      "debugging/raw_token_scores_std": 0.022307263687253,
+      "debugging/score": 0.00915515707107717,
+      "debugging/score_std": 0.01661134296838666,
+      "epoch": 0.6,
+      "objective/train/advantage_avg": 0.0009648797567933798,
+      "objective/train/docs_used": 716760,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4830039739608765,
+      "objective/train/original_loss": 1.483003854751587,
+      "objective/train/theoretical_loss": 3.4637628541329066,
+      "objective/train/tokens_used": 1986540000,
+      "objective/train/value_avg": -0.005992889404296875,
+      "objective/train/value_loss": 0.000370719499187544,
+      "objective/train/value_max": -4.947185516357422e-05,
+      "objective/train/value_min": -0.92724609375,
+      "objective/train/value_reward_corr": 0.5103701637010936,
+      "objective/train/value_std": 0.0126495361328125,
+      "objective/train/weight_avg": 1.0010908842086792,
+      "objective/train/weighted_lm_loss": 1.4844170808792114,
+      "objective/train/weights_max": 2.2417259216308594,
+      "objective/train/weights_min": 0.06188541650772095,
+      "theoretical_loss": 3.4637628541329066,
+      "tokens_seen": 1966080000
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040828117477130477,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4637628541329066,
+      "tokens_seen": 1966080000
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004082009308297224,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4637245384931576,
+      "tokens_seen": 1966342144
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040812068688813994,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4636862293911816,
+      "tokens_seen": 1966604288
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040804044294655756,
+      "loss": 0.0678,
+      "theoretical_loss": 3.463647926824992,
+      "tokens_seen": 1966866432
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004079601990049751,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4636096307926016,
+      "tokens_seen": 1967128576
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004078799550633927,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4635713412920275,
+      "tokens_seen": 1967390720
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004077997111218103,
+      "loss": 0.0703,
+      "theoretical_loss": 3.4635330583212838,
+      "tokens_seen": 1967652864
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004077194671802279,
+      "loss": 0.0693,
+      "theoretical_loss": 3.463494781878388,
+      "tokens_seen": 1967915008
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004076392232386455,
+      "loss": 0.0716,
+      "theoretical_loss": 3.4634565119613576,
+      "tokens_seen": 1968177152
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040755897929706307,
+      "loss": 0.0698,
+      "theoretical_loss": 3.463418248568211,
+      "tokens_seen": 1968439296
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004074787353554807,
+      "loss": 0.0672,
+      "theoretical_loss": 3.463379991696967,
+      "tokens_seen": 1968701440
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004073984914138983,
+      "loss": 0.0682,
+      "theoretical_loss": 3.463341741345646,
+      "tokens_seen": 1968963584
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040731824747231585,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4633034975122694,
+      "tokens_seen": 1969225728
+    },
+    {
+      "epoch": 0.6,
+      "objective/train/advantage_avg": 0.0015214718878269196,
+      "objective/train/docs_used": 718017,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4403096437454224,
+      "objective/train/original_loss": 1.4403096437454224,
+      "objective/train/theoretical_loss": 3.4632843780391918,
+      "objective/train/tokens_used": 1989816800,
+      "objective/train/value_avg": -0.005222320556640625,
+      "objective/train/value_loss": 0.0001249258202733472,
+      "objective/train/value_max": -1.7344951629638672e-05,
+      "objective/train/value_min": -0.24560546875,
+      "objective/train/value_reward_corr": 0.6257195538265397,
+      "objective/train/value_std": 0.00986480712890625,
+      "objective/train/weight_avg": 1.0015816688537598,
+      "objective/train/weighted_lm_loss": 1.442775845527649,
+      "objective/train/weights_max": 1.1413390636444092,
+      "objective/train/weights_min": 0.7176816463470459,
+      "theoretical_loss": 3.4632843780391918,
+      "tokens_seen": 1969356800
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004072380035307334,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4632652601948593,
+      "tokens_seen": 1969487872
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.000407157759589151,
+      "loss": 0.0684,
+      "theoretical_loss": 3.463227029391437,
+      "tokens_seen": 1969750016
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040707751564756864,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4631888051000272,
+      "tokens_seen": 1970012160
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004069972717059862,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4631505873186548,
+      "tokens_seen": 1970274304
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004069170277644038,
+      "loss": 0.0671,
+      "theoretical_loss": 3.463112376045344,
+      "tokens_seen": 1970536448
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004068367838228214,
+      "loss": 0.0678,
+      "theoretical_loss": 3.463074171278122,
+      "tokens_seen": 1970798592
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.000406756539881239,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4630359730150153,
+      "tokens_seen": 1971060736
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040667629593965654,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4629977812540518,
+      "tokens_seen": 1971322880
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040659605199807415,
+      "loss": 0.0711,
+      "theoretical_loss": 3.46295959599326,
+      "tokens_seen": 1971585024
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040651580805649176,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4629214172306706,
+      "tokens_seen": 1971847168
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004064355641149093,
+      "loss": 0.0704,
+      "theoretical_loss": 3.462883244964313,
+      "tokens_seen": 1972109312
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040635532017332693,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4628450791922187,
+      "tokens_seen": 1972371456
+    },
+    {
+      "epoch": 0.6,
+      "objective/train/advantage_avg": 0.0003923242911696434,
+      "objective/train/docs_used": 719209,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2555794715881348,
+      "objective/train/original_loss": 1.2555794715881348,
+      "objective/train/theoretical_loss": 3.4628069199124205,
+      "objective/train/tokens_used": 1993093600,
+      "objective/train/value_avg": -0.0065765380859375,
+      "objective/train/value_loss": 0.000190413833479397,
+      "objective/train/value_max": -4.13060188293457e-05,
+      "objective/train/value_min": -0.201171875,
+      "objective/train/value_reward_corr": 0.557396649936048,
+      "objective/train/value_std": 0.009490966796875,
+      "objective/train/weight_avg": 1.000475287437439,
+      "objective/train/weighted_lm_loss": 1.256247878074646,
+      "objective/train/weights_max": 1.151779294013977,
+      "objective/train/weights_min": 0.36899471282958984,
+      "theoretical_loss": 3.4628069199124205,
+      "tokens_seen": 1972633600
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040627507623174455,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4628069199124205,
+      "tokens_seen": 1972633600
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040619483229016205,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4627687671229515,
+      "tokens_seen": 1972895744
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040611458834857966,
+      "loss": 0.0686,
+      "theoretical_loss": 3.462730620821845,
+      "tokens_seen": 1973157888
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004060343444069973,
+      "loss": 0.0695,
+      "theoretical_loss": 3.462692481007136,
+      "tokens_seen": 1973420032
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004059541004654149,
+      "loss": 0.0678,
+      "theoretical_loss": 3.46265434767686,
+      "tokens_seen": 1973682176
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040587385652383245,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4626162208290534,
+      "tokens_seen": 1973944320
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040579361258225006,
+      "loss": 0.0691,
+      "theoretical_loss": 3.462578100461754,
+      "tokens_seen": 1974206464
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040571336864066767,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4625399865730007,
+      "tokens_seen": 1974468608
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040563312469908523,
+      "loss": 0.0671,
+      "theoretical_loss": 3.462501879160831,
+      "tokens_seen": 1974730752
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004055528807575028,
+      "loss": 0.0701,
+      "theoretical_loss": 3.462463778223285,
+      "tokens_seen": 1974992896
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004054726368159204,
+      "loss": 0.0682,
+      "theoretical_loss": 3.462425683758404,
+      "tokens_seen": 1975255040
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.000405392392874338,
+      "loss": 0.0723,
+      "theoretical_loss": 3.4623875957642305,
+      "tokens_seen": 1975517184
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040531214893275557,
+      "loss": 0.0699,
+      "theoretical_loss": 3.462349514238805,
+      "tokens_seen": 1975779328
+    },
+    {
+      "epoch": 0.6,
+      "objective/train/advantage_avg": 0.0036880422849208117,
+      "objective/train/docs_used": 720450,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5445085763931274,
+      "objective/train/original_loss": 1.544508457183838,
+      "objective/train/theoretical_loss": 3.462330475901262,
+      "objective/train/tokens_used": 1996370400,
+      "objective/train/value_avg": -0.01209259033203125,
+      "objective/train/value_loss": 0.0005820911610499024,
+      "objective/train/value_max": -5.519390106201172e-05,
+      "objective/train/value_min": -0.8037109375,
+      "objective/train/value_reward_corr": 0.5987648461866432,
+      "objective/train/value_std": 0.0245361328125,
+      "objective/train/weight_avg": 1.0039633512496948,
+      "objective/train/weighted_lm_loss": 1.5502508878707886,
+      "objective/train/weights_max": 2.015129804611206,
+      "objective/train/weights_min": 0.34483224153518677,
+      "theoretical_loss": 3.462330475901262,
+      "tokens_seen": 1975910400
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004052319049911732,
+      "loss": 0.0705,
+      "theoretical_loss": 3.462311439180173,
+      "tokens_seen": 1976041472
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004051516610495908,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4622733705863764,
+      "tokens_seen": 1976303616
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040507141710800835,
+      "loss": 0.0653,
+      "theoretical_loss": 3.462235308455462,
+      "tokens_seen": 1976565760
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004049911731664259,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4621972527854745,
+      "tokens_seen": 1976827904
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004049109292248435,
+      "loss": 0.0705,
+      "theoretical_loss": 3.462159203574461,
+      "tokens_seen": 1977090048
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040483068528326114,
+      "loss": 0.0686,
+      "theoretical_loss": 3.46212116082047,
+      "tokens_seen": 1977352192
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004047504413416787,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4620831245215484,
+      "tokens_seen": 1977614336
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004046701974000963,
+      "loss": 0.069,
+      "theoretical_loss": 3.462045094675747,
+      "tokens_seen": 1977876480
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004045899534585139,
+      "loss": 0.0683,
+      "theoretical_loss": 3.462007071281114,
+      "tokens_seen": 1978138624
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004045097095169315,
+      "loss": 0.0666,
+      "theoretical_loss": 3.461969054335703,
+      "tokens_seen": 1978400768
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040442946557534904,
+      "loss": 0.0687,
+      "theoretical_loss": 3.461931043837563,
+      "tokens_seen": 1978662912
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040434922163376665,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4618930397847487,
+      "tokens_seen": 1978925056
+    },
+    {
+      "epoch": 0.6,
+      "objective/train/advantage_avg": 0.0012460628058761358,
+      "objective/train/docs_used": 721537,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4055533409118652,
+      "objective/train/original_loss": 1.4055533409118652,
+      "objective/train/theoretical_loss": 3.461855042175312,
+      "objective/train/tokens_used": 1999647200,
+      "objective/train/value_avg": -0.005329132080078125,
+      "objective/train/value_loss": 0.0001690825738478452,
+      "objective/train/value_max": -6.014108657836914e-05,
+      "objective/train/value_min": -0.61181640625,
+      "objective/train/value_reward_corr": 0.6281343766334138,
+      "objective/train/value_std": 0.01203155517578125,
+      "objective/train/weight_avg": 1.0013235807418823,
+      "objective/train/weighted_lm_loss": 1.4071850776672363,
+      "objective/train/weights_max": 1.8437772989273071,
+      "objective/train/weights_min": 0.37389513850212097,
+      "theoretical_loss": 3.461855042175312,
+      "tokens_seen": 1979187200
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004042689776921842,
+      "loss": 0.0678,
+      "theoretical_loss": 3.461855042175312,
+      "tokens_seen": 1979187200
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004041887337506018,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4618170510073085,
+      "tokens_seen": 1979449344
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040410848980901944,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4617790662787935,
+      "tokens_seen": 1979711488
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040402824586743705,
+      "loss": 0.0725,
+      "theoretical_loss": 3.4617410879878223,
+      "tokens_seen": 1979973632
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004039480019258546,
+      "loss": 0.0694,
+      "theoretical_loss": 3.461703116132452,
+      "tokens_seen": 1980235776
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004038677579842722,
+      "loss": 0.0696,
+      "theoretical_loss": 3.46166515071074,
+      "tokens_seen": 1980497920
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004037875140426898,
+      "loss": 0.0684,
+      "theoretical_loss": 3.461627191720745,
+      "tokens_seen": 1980760064
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040370727010110734,
+      "loss": 0.0694,
+      "theoretical_loss": 3.461589239160528,
+      "tokens_seen": 1981022208
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040362702615952495,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4615512930281467,
+      "tokens_seen": 1981284352
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040354678221794256,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4615133533216635,
+      "tokens_seen": 1981546496
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004034665382763602,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4614754200391404,
+      "tokens_seen": 1981808640
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040338629433477773,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4614374931786402,
+      "tokens_seen": 1982070784
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040330605039319534,
+      "loss": 0.0668,
+      "theoretical_loss": 3.461399572738226,
+      "tokens_seen": 1982332928
+    },
+    {
+      "epoch": 0.6,
+      "objective/train/advantage_avg": 0.0009382269927300513,
+      "objective/train/docs_used": 722739,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.329759955406189,
+      "objective/train/original_loss": 1.3297600746154785,
+      "objective/train/theoretical_loss": 3.4613806149249466,
+      "objective/train/tokens_used": 2002924000,
+      "objective/train/value_avg": -0.0085906982421875,
+      "objective/train/value_loss": 0.00015004277520347387,
+      "objective/train/value_max": -3.218650817871094e-05,
+      "objective/train/value_min": -0.362548828125,
+      "objective/train/value_reward_corr": 0.81791631267802,
+      "objective/train/value_std": 0.0171661376953125,
+      "objective/train/weight_avg": 1.0010125637054443,
+      "objective/train/weighted_lm_loss": 1.3312724828720093,
+      "objective/train/weights_max": 1.30023992061615,
+      "objective/train/weights_min": 0.8231871724128723,
+      "theoretical_loss": 3.4613806149249466,
+      "tokens_seen": 1982464000
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004032258064516129,
+      "loss": 0.0687,
+      "theoretical_loss": 3.461361658715963,
+      "tokens_seen": 1982595072
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040314556251003046,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4613237511099157,
+      "tokens_seen": 1982857216
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004030653185684481,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4612858499181502,
+      "tokens_seen": 1983119360
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004029850746268657,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4612479551387345,
+      "tokens_seen": 1983381504
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004029048306852833,
+      "loss": 0.0691,
+      "theoretical_loss": 3.461210066769736,
+      "tokens_seen": 1983643648
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040282458674370086,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4611721848092225,
+      "tokens_seen": 1983905792
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040274434280211847,
+      "loss": 0.0691,
+      "theoretical_loss": 3.461134309255265,
+      "tokens_seen": 1984167936
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004026640988605361,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4610964401059325,
+      "tokens_seen": 1984430080
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004025838549189536,
+      "loss": 0.0692,
+      "theoretical_loss": 3.461058577359297,
+      "tokens_seen": 1984692224
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004025036109773712,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4610207210134294,
+      "tokens_seen": 1984954368
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004024233670357888,
+      "loss": 0.07,
+      "theoretical_loss": 3.4609828710664035,
+      "tokens_seen": 1985216512
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004023431230942064,
+      "loss": 0.0674,
+      "theoretical_loss": 3.460945027516293,
+      "tokens_seen": 1985478656
+    },
+    {
+      "epoch": 0.6,
+      "objective/train/advantage_avg": 0.0006598348263651133,
+      "objective/train/docs_used": 723995,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3495676517486572,
+      "objective/train/original_loss": 1.3495676517486572,
+      "objective/train/theoretical_loss": 3.460907190361172,
+      "objective/train/tokens_used": 2006200800,
+      "objective/train/value_avg": -0.01053619384765625,
+      "objective/train/value_loss": 0.0005898875533603132,
+      "objective/train/value_max": -5.4776668548583984e-05,
+      "objective/train/value_min": -0.9677734375,
+      "objective/train/value_reward_corr": 0.6037339334248357,
+      "objective/train/value_std": 0.02178955078125,
+      "objective/train/weight_avg": 1.0009301900863647,
+      "objective/train/weighted_lm_loss": 1.3499261140823364,
+      "objective/train/weights_max": 2.356096029281616,
+      "objective/train/weights_min": 0.3713054358959198,
+      "theoretical_loss": 3.460907190361172,
+      "tokens_seen": 1985740800
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.000402262879152624,
+      "loss": 0.0675,
+      "theoretical_loss": 3.460907190361172,
+      "tokens_seen": 1985740800
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004021826352110416,
+      "loss": 0.068,
+      "theoretical_loss": 3.460869359599116,
+      "tokens_seen": 1986002944
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004021023912694592,
+      "loss": 0.0682,
+      "theoretical_loss": 3.460831535228201,
+      "tokens_seen": 1986265088
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004020221473278767,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4607937172465046,
+      "tokens_seen": 1986527232
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004019419033862943,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4607559056521033,
+      "tokens_seen": 1986789376
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040186165944471194,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4607181004430774,
+      "tokens_seen": 1987051520
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004017814155031295,
+      "loss": 0.0683,
+      "theoretical_loss": 3.460680301617505,
+      "tokens_seen": 1987313664
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004017011715615471,
+      "loss": 0.069,
+      "theoretical_loss": 3.460642509173468,
+      "tokens_seen": 1987575808
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004016209276199647,
+      "loss": 0.0719,
+      "theoretical_loss": 3.4606047231090455,
+      "tokens_seen": 1987837952
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040154068367838233,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4605669434223216,
+      "tokens_seen": 1988100096
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040146043973679984,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4605291701113776,
+      "tokens_seen": 1988362240
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040138019579521745,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4604914031742977,
+      "tokens_seen": 1988624384
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040129995185363506,
+      "loss": 0.0672,
+      "theoretical_loss": 3.460453642609166,
+      "tokens_seen": 1988886528
+    },
+    {
+      "epoch": 0.6,
+      "objective/train/advantage_avg": 0.0007240816485136747,
+      "objective/train/docs_used": 725238,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5036556720733643,
+      "objective/train/original_loss": 1.5036556720733643,
+      "objective/train/theoretical_loss": 3.4604347647154827,
+      "objective/train/tokens_used": 2009477600,
+      "objective/train/value_avg": -0.005214691162109375,
+      "objective/train/value_loss": 8.880392124410719e-05,
+      "objective/train/value_max": -2.6285648345947266e-05,
+      "objective/train/value_min": -0.1871337890625,
+      "objective/train/value_reward_corr": 0.6710600827464257,
+      "objective/train/value_std": 0.00882720947265625,
+      "objective/train/weight_avg": 1.000767707824707,
+      "objective/train/weighted_lm_loss": 1.5045437812805176,
+      "objective/train/weights_max": 1.109333872795105,
+      "objective/train/weights_min": 0.7807214856147766,
+      "theoretical_loss": 3.4604347647154827,
+      "tokens_seen": 1989017600
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004012197079120526,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4604158884140683,
+      "tokens_seen": 1989148672
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040113946397047023,
+      "loss": 0.0685,
+      "theoretical_loss": 3.460378140587091,
+      "tokens_seen": 1989410816
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040105922002888785,
+      "loss": 0.0694,
+      "theoretical_loss": 3.46034039912632,
+      "tokens_seen": 1989672960
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040097897608730546,
+      "loss": 0.0662,
+      "theoretical_loss": 3.460302664029844,
+      "tokens_seen": 1989935104
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.000400898732145723,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4602649352957515,
+      "tokens_seen": 1990197248
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004008184882041406,
+      "loss": 0.0676,
+      "theoretical_loss": 3.460227212922131,
+      "tokens_seen": 1990459392
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004007382442625582,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4601894969070743,
+      "tokens_seen": 1990721536
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040065800032097575,
+      "loss": 0.067,
+      "theoretical_loss": 3.460151787248672,
+      "tokens_seen": 1990983680
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040057775637939336,
+      "loss": 0.0682,
+      "theoretical_loss": 3.460114083945015,
+      "tokens_seen": 1991245824
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040049751243781097,
+      "loss": 0.069,
+      "theoretical_loss": 3.4600763869941966,
+      "tokens_seen": 1991507968
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004004172684962286,
+      "loss": 0.0687,
+      "theoretical_loss": 3.460038696394311,
+      "tokens_seen": 1991770112
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00040033702455464614,
+      "loss": 0.0706,
+      "theoretical_loss": 3.4600010121434517,
+      "tokens_seen": 1992032256
+    },
+    {
+      "epoch": 0.6,
+      "objective/train/advantage_avg": 0.00020876179041806608,
+      "objective/train/docs_used": 726441,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.400139570236206,
+      "objective/train/original_loss": 1.400139570236206,
+      "objective/train/theoretical_loss": 3.4599633342397142,
+      "objective/train/tokens_used": 2012754400,
+      "objective/train/value_avg": -0.008514404296875,
+      "objective/train/value_loss": 0.00025450316024944186,
+      "objective/train/value_max": -5.9664249420166016e-05,
+      "objective/train/value_min": -0.28369140625,
+      "objective/train/value_reward_corr": 0.7200322547456051,
+      "objective/train/value_std": 0.0171051025390625,
+      "objective/train/weight_avg": 1.0003241300582886,
+      "objective/train/weighted_lm_loss": 1.4006654024124146,
+      "objective/train/weights_max": 1.277927041053772,
+      "objective/train/weights_min": 0.36856982111930847,
+      "theoretical_loss": 3.4599633342397142,
+      "tokens_seen": 1992294400
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004002567806130637,
+      "loss": 0.068,
+      "theoretical_loss": 3.4599633342397142,
+      "tokens_seen": 1992294400
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004001765366714813,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4599256626811945,
+      "tokens_seen": 1992556544
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004000962927298989,
+      "loss": 0.0682,
+      "theoretical_loss": 3.45988799746599,
+      "tokens_seen": 1992818688
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0004000160487883165,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4598503385921977,
+      "tokens_seen": 1993080832
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0003999358048467341,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4598126860579166,
+      "tokens_seen": 1993342976
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00039985556090515166,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4597750398612455,
+      "tokens_seen": 1993605120
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00039977531696356927,
+      "loss": 0.0702,
+      "theoretical_loss": 3.459737400000284,
+      "tokens_seen": 1993867264
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00039969507302198683,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4596997664731344,
+      "tokens_seen": 1994129408
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00039961482908040444,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4596621392778983,
+      "tokens_seen": 1994391552
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.000399534585138822,
+      "loss": 0.0686,
+      "theoretical_loss": 3.459624518412677,
+      "tokens_seen": 1994653696
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0003994543411972396,
+      "loss": 0.0667,
+      "theoretical_loss": 3.459586903875575,
+      "tokens_seen": 1994915840
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0003993740972556572,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4595492956646963,
+      "tokens_seen": 1995177984
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0003992938533140748,
+      "loss": 0.0675,
+      "theoretical_loss": 3.459511693778146,
+      "tokens_seen": 1995440128
+    },
+    {
+      "epoch": 0.6,
+      "objective/train/advantage_avg": 0.0013966768747195601,
+      "objective/train/docs_used": 727673,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2055573463439941,
+      "objective/train/original_loss": 1.2055573463439941,
+      "objective/train/theoretical_loss": 3.4594928952059014,
+      "objective/train/tokens_used": 2016031200,
+      "objective/train/value_avg": -0.00856781005859375,
+      "objective/train/value_loss": 0.00022097454348113388,
+      "objective/train/value_max": -4.57763671875e-05,
+      "objective/train/value_min": -0.63623046875,
+      "objective/train/value_reward_corr": 0.7161873647776913,
+      "objective/train/value_std": 0.01617431640625,
+      "objective/train/weight_avg": 1.0014972686767578,
+      "objective/train/weighted_lm_loss": 1.2068300247192383,
+      "objective/train/weights_max": 1.8893455266952515,
+      "objective/train/weights_min": 0.3683340847492218,
+      "theoretical_loss": 3.4594928952059014,
+      "tokens_seen": 1995571200
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.0003992136093724924,
+      "loss": 0.068,
+      "theoretical_loss": 3.4594740982140295,
+      "tokens_seen": 1995702272
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00039913336543091,
+      "loss": 0.0699,
+      "theoretical_loss": 3.459436508970454,
+      "tokens_seen": 1995964416
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 0.00039905312148932757,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4593989260455267,
+      "tokens_seen": 1996226560
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003989728775477451,
+      "loss": 0.0678,
+      "theoretical_loss": 3.459361349437356,
+      "tokens_seen": 1996488704
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039889263360616274,
+      "loss": 0.0696,
+      "theoretical_loss": 3.459323779144051,
+      "tokens_seen": 1996750848
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039881238966458035,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4592862151637216,
+      "tokens_seen": 1997012992
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003987321457229979,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4592486574944785,
+      "tokens_seen": 1997275136
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003986519017814155,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4592111061344335,
+      "tokens_seen": 1997537280
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039857165783983313,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4591735610816983,
+      "tokens_seen": 1997799424
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003984914138982507,
+      "loss": 0.069,
+      "theoretical_loss": 3.459136022334387,
+      "tokens_seen": 1998061568
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039841116995666825,
+      "loss": 0.0715,
+      "theoretical_loss": 3.4590984898906134,
+      "tokens_seen": 1998323712
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039833092601508586,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4590609637484913,
+      "tokens_seen": 1998585856
+    },
+    {
+      "epoch": 0.61,
+      "objective/train/advantage_avg": -0.0010950957657769322,
+      "objective/train/docs_used": 728818,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.404526948928833,
+      "objective/train/original_loss": 1.404526948928833,
+      "objective/train/theoretical_loss": 3.459023443906138,
+      "objective/train/tokens_used": 2019308000,
+      "objective/train/value_avg": -0.0084686279296875,
+      "objective/train/value_loss": 0.00020435434998944402,
+      "objective/train/value_max": -2.014636993408203e-05,
+      "objective/train/value_min": -0.281494140625,
+      "objective/train/value_reward_corr": 0.7389789184439145,
+      "objective/train/value_std": 0.014190673828125,
+      "objective/train/weight_avg": 0.9990012645721436,
+      "objective/train/weighted_lm_loss": 1.4025185108184814,
+      "objective/train/weights_max": 1.1698321104049683,
+      "objective/train/weights_min": 0.3728525638580322,
+      "theoretical_loss": 3.459023443906138,
+      "tokens_seen": 1998848000
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003982506820735035,
+      "loss": 0.0686,
+      "theoretical_loss": 3.459023443906138,
+      "tokens_seen": 1998848000
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039817043813192103,
+      "loss": 0.0676,
+      "theoretical_loss": 3.458985930361668,
+      "tokens_seen": 1999110144
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039809019419033865,
+      "loss": 0.068,
+      "theoretical_loss": 3.4589484231132,
+      "tokens_seen": 1999372288
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039800995024875626,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4589109221588514,
+      "tokens_seen": 1999634432
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003979297063071738,
+      "loss": 0.0721,
+      "theoretical_loss": 3.4588734274967416,
+      "tokens_seen": 1999896576
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003978494623655914,
+      "loss": 0.07,
+      "theoretical_loss": 3.4588359391249894,
+      "tokens_seen": 2000158720
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.000397769218424009,
+      "loss": 0.0681,
+      "theoretical_loss": 3.458798457041716,
+      "tokens_seen": 2000420864
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003976889744824266,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4587609812450424,
+      "tokens_seen": 2000683008
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039760873054084416,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4587235117330906,
+      "tokens_seen": 2000945152
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039752848659926177,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4586860485039836,
+      "tokens_seen": 2001207296
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003974482426576794,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4586485915558454,
+      "tokens_seen": 2001469440
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039736799871609694,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4586111408868,
+      "tokens_seen": 2001731584
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003972877547745145,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4585736964949727,
+      "tokens_seen": 2001993728
+    },
+    {
+      "epoch": 0.61,
+      "objective/train/advantage_avg": 0.0014532208442687988,
+      "objective/train/docs_used": 729965,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.308320164680481,
+      "objective/train/original_loss": 1.308320164680481,
+      "objective/train/theoretical_loss": 3.4585549766524304,
+      "objective/train/tokens_used": 2022584800,
+      "objective/train/value_avg": -0.00588226318359375,
+      "objective/train/value_loss": 0.00016835334827192128,
+      "objective/train/value_max": -4.7206878662109375e-05,
+      "objective/train/value_min": -0.2203369140625,
+      "objective/train/value_reward_corr": 0.6780754803158118,
+      "objective/train/value_std": 0.01114654541015625,
+      "objective/train/weight_avg": 1.0015290975570679,
+      "objective/train/weighted_lm_loss": 1.310949683189392,
+      "objective/train/weights_max": 1.2337805032730103,
+      "objective/train/weights_min": 0.3807870149612427,
+      "theoretical_loss": 3.4585549766524304,
+      "tokens_seen": 2002124800
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003972075108329321,
+      "loss": 0.0666,
+      "theoretical_loss": 3.45853625837849,
+      "tokens_seen": 2002255872
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003971272668913497,
+      "loss": 0.0663,
+      "theoretical_loss": 3.458498826535479,
+      "tokens_seen": 2002518016
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003970470229497673,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4584614009640666,
+      "tokens_seen": 2002780160
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003969667790081849,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4584239816623823,
+      "tokens_seen": 2003042304
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003968865350666025,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4583865686285544,
+      "tokens_seen": 2003304448
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039680629112502007,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4583491618607134,
+      "tokens_seen": 2003566592
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039672604718343763,
+      "loss": 0.0687,
+      "theoretical_loss": 3.458311761356991,
+      "tokens_seen": 2003828736
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039664580324185524,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4582743671155183,
+      "tokens_seen": 2004090880
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039656555930027285,
+      "loss": 0.0694,
+      "theoretical_loss": 3.458236979134428,
+      "tokens_seen": 2004353024
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003964853153586904,
+      "loss": 0.0671,
+      "theoretical_loss": 3.458199597411853,
+      "tokens_seen": 2004615168
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.000396405071417108,
+      "loss": 0.069,
+      "theoretical_loss": 3.4581622219459276,
+      "tokens_seen": 2004877312
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039632482747552564,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4581248527347874,
+      "tokens_seen": 2005139456
+    },
+    {
+      "epoch": 0.61,
+      "objective/train/advantage_avg": 0.00025961673236452043,
+      "objective/train/docs_used": 731237,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3400050401687622,
+      "objective/train/original_loss": 1.3400051593780518,
+      "objective/train/theoretical_loss": 3.458087489776567,
+      "objective/train/tokens_used": 2025861600,
+      "objective/train/value_avg": -0.0065155029296875,
+      "objective/train/value_loss": 0.00015271175652742386,
+      "objective/train/value_max": -1.7642974853515625e-05,
+      "objective/train/value_min": -0.496826171875,
+      "objective/train/value_reward_corr": 0.73502062968239,
+      "objective/train/value_std": 0.0140533447265625,
+      "objective/train/weight_avg": 1.0003293752670288,
+      "objective/train/weighted_lm_loss": 1.3405778408050537,
+      "objective/train/weights_max": 1.4610846042633057,
+      "objective/train/weights_min": 0.40551888942718506,
+      "theoretical_loss": 3.458087489776567,
+      "tokens_seen": 2005401600
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003962445835339432,
+      "loss": 0.0674,
+      "theoretical_loss": 3.458087489776567,
+      "tokens_seen": 2005401600
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003961643395923608,
+      "loss": 0.0677,
+      "theoretical_loss": 3.458050133069404,
+      "tokens_seen": 2005663744
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039608409565077837,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4580127826114353,
+      "tokens_seen": 2005925888
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003960038517091959,
+      "loss": 0.0672,
+      "theoretical_loss": 3.457975438400799,
+      "tokens_seen": 2006188032
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039592360776761354,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4579381004356344,
+      "tokens_seen": 2006450176
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039584336382603115,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4579007687140804,
+      "tokens_seen": 2006712320
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039576311988444876,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4578634432342783,
+      "tokens_seen": 2006974464
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003956828759428663,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4578261239943693,
+      "tokens_seen": 2007236608
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039560263200128393,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4577888109924952,
+      "tokens_seen": 2007498752
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003955223880597015,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4577515042267994,
+      "tokens_seen": 2007760896
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039544214411811905,
+      "loss": 0.0655,
+      "theoretical_loss": 3.457714203695425,
+      "tokens_seen": 2008023040
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039536190017653666,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4576769093965174,
+      "tokens_seen": 2008285184
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003952816562349543,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4576396213282212,
+      "tokens_seen": 2008547328
+    },
+    {
+      "epoch": 0.61,
+      "objective/train/advantage_avg": 0.000601315638050437,
+      "objective/train/docs_used": 732466,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3620326519012451,
+      "objective/train/original_loss": 1.3620326519012451,
+      "objective/train/theoretical_loss": 3.457620979629973,
+      "objective/train/tokens_used": 2029138400,
+      "objective/train/value_avg": -0.00832366943359375,
+      "objective/train/value_loss": 0.00017946993466466665,
+      "objective/train/value_max": -8.821487426757812e-05,
+      "objective/train/value_min": -0.301513671875,
+      "objective/train/value_reward_corr": 0.65227393135886,
+      "objective/train/value_std": 0.012939453125,
+      "objective/train/weight_avg": 1.0006853342056274,
+      "objective/train/weighted_lm_loss": 1.3613041639328003,
+      "objective/train/weights_max": 1.1302473545074463,
+      "objective/train/weights_min": 0.3697379231452942,
+      "theoretical_loss": 3.457620979629973,
+      "tokens_seen": 2008678400
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003952014122933719,
+      "loss": 0.0702,
+      "theoretical_loss": 3.457602339488682,
+      "tokens_seen": 2008809472
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039512116835178945,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4575650638760482,
+      "tokens_seen": 2009071616
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039504092441020706,
+      "loss": 0.0687,
+      "theoretical_loss": 3.457527794488466,
+      "tokens_seen": 2009333760
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003949606804686246,
+      "loss": 0.0685,
+      "theoretical_loss": 3.457490531324085,
+      "tokens_seen": 2009595904
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003948804365270422,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4574532743810535,
+      "tokens_seen": 2009858048
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003948001925854598,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4574160236575224,
+      "tokens_seen": 2010120192
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003947199486438774,
+      "loss": 0.0685,
+      "theoretical_loss": 3.457378779151642,
+      "tokens_seen": 2010382336
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.000394639704702295,
+      "loss": 0.0675,
+      "theoretical_loss": 3.457341540861564,
+      "tokens_seen": 2010644480
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039455946076071257,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4573043087854414,
+      "tokens_seen": 2010906624
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003944792168191302,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4572670829214265,
+      "tokens_seen": 2011168768
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003943989728775478,
+      "loss": 0.0682,
+      "theoretical_loss": 3.457229863267674,
+      "tokens_seen": 2011430912
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003943187289359653,
+      "loss": 0.0678,
+      "theoretical_loss": 3.457192649822338,
+      "tokens_seen": 2011693056
+    },
+    {
+      "epoch": 0.61,
+      "objective/train/advantage_avg": 0.0002623085747472942,
+      "objective/train/docs_used": 733698,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4967702627182007,
+      "objective/train/original_loss": 1.4967701435089111,
+      "objective/train/theoretical_loss": 3.4571554425835753,
+      "objective/train/tokens_used": 2032415200,
+      "objective/train/value_avg": -0.007495880126953125,
+      "objective/train/value_loss": 0.00019993532623630017,
+      "objective/train/value_max": -4.363059997558594e-05,
+      "objective/train/value_min": -0.908203125,
+      "objective/train/value_reward_corr": 0.6953280027505602,
+      "objective/train/value_std": 0.013702392578125,
+      "objective/train/weight_avg": 1.0003552436828613,
+      "objective/train/weighted_lm_loss": 1.497015357017517,
+      "objective/train/weights_max": 1.4672387838363647,
+      "objective/train/weights_min": 0.3828728199005127,
+      "theoretical_loss": 3.4571554425835753,
+      "tokens_seen": 2011955200
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003942384849943829,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4571554425835753,
+      "tokens_seen": 2011955200
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003941582410528005,
+      "loss": 0.0688,
+      "theoretical_loss": 3.457118241549541,
+      "tokens_seen": 2012217344
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039407799711121814,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4570810467183932,
+      "tokens_seen": 2012479488
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003939977531696357,
+      "loss": 0.0696,
+      "theoretical_loss": 3.45704385808829,
+      "tokens_seen": 2012741632
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003939175092280533,
+      "loss": 0.0704,
+      "theoretical_loss": 3.4570066756573885,
+      "tokens_seen": 2013003776
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003938372652864709,
+      "loss": 0.0675,
+      "theoretical_loss": 3.45696949942385,
+      "tokens_seen": 2013265920
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003937570213448884,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4569323293858334,
+      "tokens_seen": 2013528064
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039367677740330604,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4568951655415017,
+      "tokens_seen": 2013790208
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039359653346172365,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4568580078890143,
+      "tokens_seen": 2014052352
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003935162895201412,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4568208564265364,
+      "tokens_seen": 2014314496
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003934360455785588,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4567837111522293,
+      "tokens_seen": 2014576640
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039335580163697644,
+      "loss": 0.0676,
+      "theoretical_loss": 3.456746572064259,
+      "tokens_seen": 2014838784
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039327555769539405,
+      "loss": 0.0676,
+      "theoretical_loss": 3.456709439160789,
+      "tokens_seen": 2015100928
+    },
+    {
+      "epoch": 0.61,
+      "objective/train/advantage_avg": 0.00043751299381256104,
+      "objective/train/docs_used": 734924,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4263153076171875,
+      "objective/train/original_loss": 1.4263153076171875,
+      "objective/train/theoretical_loss": 3.456690875027669,
+      "objective/train/tokens_used": 2035692000,
+      "objective/train/value_avg": -0.00778961181640625,
+      "objective/train/value_loss": 0.00017772949649952352,
+      "objective/train/value_max": -4.13060188293457e-05,
+      "objective/train/value_min": -0.476318359375,
+      "objective/train/value_reward_corr": 0.6913995082057987,
+      "objective/train/value_std": 0.01406097412109375,
+      "objective/train/weight_avg": 1.0005217790603638,
+      "objective/train/weighted_lm_loss": 1.4272732734680176,
+      "objective/train/weights_max": 1.438776969909668,
+      "objective/train/weights_min": 0.3907938003540039,
+      "theoretical_loss": 3.456690875027669,
+      "tokens_seen": 2015232000
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003931953137538116,
+      "loss": 0.0678,
+      "theoretical_loss": 3.456672312439986,
+      "tokens_seen": 2015363072
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039311506981222916,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4566351919000167,
+      "tokens_seen": 2015625216
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003930348258706468,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4565980775390477,
+      "tokens_seen": 2015887360
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039295458192906434,
+      "loss": 0.0688,
+      "theoretical_loss": 3.456560969355248,
+      "tokens_seen": 2016149504
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039287433798748195,
+      "loss": 0.0673,
+      "theoretical_loss": 3.456523867346786,
+      "tokens_seen": 2016411648
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039279409404589956,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4564867715118313,
+      "tokens_seen": 2016673792
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003927138501043172,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4564496818485546,
+      "tokens_seen": 2016935936
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039263360616273473,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4564125983551275,
+      "tokens_seen": 2017198080
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003925533622211523,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4563755210297216,
+      "tokens_seen": 2017460224
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003924731182795699,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4563384498705094,
+      "tokens_seen": 2017722368
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039239287433798746,
+      "loss": 0.0679,
+      "theoretical_loss": 3.456301384875666,
+      "tokens_seen": 2017984512
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003923126303964051,
+      "loss": 0.0666,
+      "theoretical_loss": 3.456264326043364,
+      "tokens_seen": 2018246656
+    },
+    {
+      "epoch": 0.61,
+      "objective/train/advantage_avg": 0.0014328511897474527,
+      "objective/train/docs_used": 736110,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3064955472946167,
+      "objective/train/original_loss": 1.3064955472946167,
+      "objective/train/theoretical_loss": 3.4562272733717796,
+      "objective/train/tokens_used": 2038968800,
+      "objective/train/value_avg": -0.0081787109375,
+      "objective/train/value_loss": 0.00021994484995957464,
+      "objective/train/value_max": -4.506111145019531e-05,
+      "objective/train/value_min": -0.57373046875,
+      "objective/train/value_reward_corr": 0.7439695048999607,
+      "objective/train/value_std": 0.017333984375,
+      "objective/train/weight_avg": 1.001534104347229,
+      "objective/train/weighted_lm_loss": 1.3077045679092407,
+      "objective/train/weights_max": 1.2325594425201416,
+      "objective/train/weights_min": 0.3684840798377991,
+      "theoretical_loss": 3.4562272733717796,
+      "tokens_seen": 2018508800
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003922323864548227,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4562272733717796,
+      "tokens_seen": 2018508800
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003921521425132403,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4561902268590883,
+      "tokens_seen": 2018770944
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039207189857165786,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4561531865034665,
+      "tokens_seen": 2019033088
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003919916546300754,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4561161523030925,
+      "tokens_seen": 2019295232
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039191141068849303,
+      "loss": 0.0672,
+      "theoretical_loss": 3.456079124256145,
+      "tokens_seen": 2019557376
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003918311667469106,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4560421023608012,
+      "tokens_seen": 2019819520
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003917509228053282,
+      "loss": 0.0706,
+      "theoretical_loss": 3.4560050866152423,
+      "tokens_seen": 2020081664
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003916706788637458,
+      "loss": 0.0665,
+      "theoretical_loss": 3.455968077017649,
+      "tokens_seen": 2020343808
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039159043492216337,
+      "loss": 0.0675,
+      "theoretical_loss": 3.455931073566202,
+      "tokens_seen": 2020605952
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.000391510190980581,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4558940762590837,
+      "tokens_seen": 2020868096
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003914299470389986,
+      "loss": 0.0698,
+      "theoretical_loss": 3.455857085094477,
+      "tokens_seen": 2021130240
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039134970309741615,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4558201000705653,
+      "tokens_seen": 2021392384
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003912694591558337,
+      "loss": 0.0701,
+      "theoretical_loss": 3.455783121185534,
+      "tokens_seen": 2021654528
+    },
+    {
+      "epoch": 0.61,
+      "objective/train/advantage_avg": 0.0006677538040094078,
+      "objective/train/docs_used": 737310,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3411515951156616,
+      "objective/train/original_loss": 1.3411513566970825,
+      "objective/train/theoretical_loss": 3.455764634044531,
+      "objective/train/tokens_used": 2042245600,
+      "objective/train/value_avg": -0.006855010986328125,
+      "objective/train/value_loss": 0.00017928997112903744,
+      "objective/train/value_max": -5.561113357543945e-05,
+      "objective/train/value_min": -0.269287109375,
+      "objective/train/value_reward_corr": 0.717935934624405,
+      "objective/train/value_std": 0.01328277587890625,
+      "objective/train/weight_avg": 1.0007518529891968,
+      "objective/train/weighted_lm_loss": 1.3418787717819214,
+      "objective/train/weights_max": 1.1867854595184326,
+      "objective/train/weights_min": 0.3734874427318573,
+      "theoretical_loss": 3.455764634044531,
+      "tokens_seen": 2021785600
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003911892152142513,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4557461484375676,
+      "tokens_seen": 2021916672
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039110897127266894,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4557091818248518,
+      "tokens_seen": 2022178816
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003910287273310865,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4556722213455737,
+      "tokens_seen": 2022440960
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003909484833895041,
+      "loss": 0.0672,
+      "theoretical_loss": 3.455635266997921,
+      "tokens_seen": 2022703104
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003908682394479217,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4555983187800825,
+      "tokens_seen": 2022965248
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003907879955063393,
+      "loss": 0.0682,
+      "theoretical_loss": 3.455561376690246,
+      "tokens_seen": 2023227392
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039070775156475684,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4555244407266024,
+      "tokens_seen": 2023489536
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039062750762317445,
+      "loss": 0.0688,
+      "theoretical_loss": 3.455487510887342,
+      "tokens_seen": 2023751680
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039054726368159206,
+      "loss": 0.0689,
+      "theoretical_loss": 3.455450587170656,
+      "tokens_seen": 2024013824
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003904670197400096,
+      "loss": 0.0705,
+      "theoretical_loss": 3.455413669574737,
+      "tokens_seen": 2024275968
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039038677579842723,
+      "loss": 0.068,
+      "theoretical_loss": 3.4553767580977777,
+      "tokens_seen": 2024538112
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039030653185684485,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4553398527379717,
+      "tokens_seen": 2024800256
+    },
+    {
+      "epoch": 0.61,
+      "objective/train/advantage_avg": 0.00035192776704207063,
+      "objective/train/docs_used": 738608,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2800827026367188,
+      "objective/train/original_loss": 1.2800827026367188,
+      "objective/train/theoretical_loss": 3.4553029534935136,
+      "objective/train/tokens_used": 2045522400,
+      "objective/train/value_avg": -0.010284423828125,
+      "objective/train/value_loss": 0.00023136693926062435,
+      "objective/train/value_max": -9.995698928833008e-05,
+      "objective/train/value_min": -0.76904296875,
+      "objective/train/value_reward_corr": 0.884760049339832,
+      "objective/train/value_std": 0.0257110595703125,
+      "objective/train/weight_avg": 1.0004581212997437,
+      "objective/train/weighted_lm_loss": 1.2805683612823486,
+      "objective/train/weights_max": 1.1941192150115967,
+      "objective/train/weights_min": 0.37031102180480957,
+      "theoretical_loss": 3.4553029534935136,
+      "tokens_seen": 2025062400
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003902262879152624,
+      "loss": 0.067,
+      "theoretical_loss": 3.4553029534935136,
+      "tokens_seen": 2025062400
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00039014604397367996,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4552660603625984,
+      "tokens_seen": 2025324544
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003900658000320976,
+      "loss": 0.0669,
+      "theoretical_loss": 3.455229173343423,
+      "tokens_seen": 2025586688
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003899855560905152,
+      "loss": 0.068,
+      "theoretical_loss": 3.455192292434183,
+      "tokens_seen": 2025848832
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00038990531214893275,
+      "loss": 0.0658,
+      "theoretical_loss": 3.455155417633076,
+      "tokens_seen": 2026110976
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00038982506820735036,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4551185489383007,
+      "tokens_seen": 2026373120
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00038974482426576797,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4550816863480565,
+      "tokens_seen": 2026635264
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00038966458032418553,
+      "loss": 0.0679,
+      "theoretical_loss": 3.455044829860543,
+      "tokens_seen": 2026897408
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003895843363826031,
+      "loss": 0.0691,
+      "theoretical_loss": 3.45500797947396,
+      "tokens_seen": 2027159552
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003895040924410207,
+      "loss": 0.0679,
+      "theoretical_loss": 3.45497113518651,
+      "tokens_seen": 2027421696
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003894238484994383,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4549342969963943,
+      "tokens_seen": 2027683840
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003893436045578559,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4548974649018165,
+      "tokens_seen": 2027945984
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003892633606162735,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4548606389009793,
+      "tokens_seen": 2028208128
+    },
+    {
+      "epoch": 0.61,
+      "objective/train/advantage_avg": 0.0008364830864593387,
+      "objective/train/docs_used": 739750,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.382369041442871,
+      "objective/train/original_loss": 1.382369041442871,
+      "objective/train/theoretical_loss": 3.4548422281851527,
+      "objective/train/tokens_used": 2048799200,
+      "objective/train/value_avg": -0.00458526611328125,
+      "objective/train/value_loss": 0.000101262099633459,
+      "objective/train/value_max": -3.510713577270508e-05,
+      "objective/train/value_min": -0.1873779296875,
+      "objective/train/value_reward_corr": 0.6766283418160592,
+      "objective/train/value_std": 0.00801849365234375,
+      "objective/train/weight_avg": 1.000882863998413,
+      "objective/train/weighted_lm_loss": 1.384031891822815,
+      "objective/train/weights_max": 1.1359182596206665,
+      "objective/train/weights_min": 0.3874214291572571,
+      "theoretical_loss": 3.4548422281851527,
+      "tokens_seen": 2028339200
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003891831166746911,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4548238189920877,
+      "tokens_seen": 2028470272
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00038910287273310866,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4547870051733467,
+      "tokens_seen": 2028732416
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.0003890226287915262,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4547501974429626,
+      "tokens_seen": 2028994560
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 0.00038894238484994383,
+      "loss": 0.0661,
+      "theoretical_loss": 3.454713395799142,
+      "tokens_seen": 2029256704
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038886214090836144,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4546766002400915,
+      "tokens_seen": 2029518848
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.000388781896966779,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4546398107640197,
+      "tokens_seen": 2029780992
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003887016530251966,
+      "loss": 0.07,
+      "theoretical_loss": 3.4546030273691364,
+      "tokens_seen": 2030043136
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003886214090836142,
+      "loss": 0.066,
+      "theoretical_loss": 3.4545662500536505,
+      "tokens_seen": 2030305280
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003885411651420318,
+      "loss": 0.071,
+      "theoretical_loss": 3.4545294788157728,
+      "tokens_seen": 2030567424
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038846092120044934,
+      "loss": 0.0686,
+      "theoretical_loss": 3.454492713653714,
+      "tokens_seen": 2030829568
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038838067725886695,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4544559545656863,
+      "tokens_seen": 2031091712
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038830043331728457,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4544192015499027,
+      "tokens_seen": 2031353856
+    },
+    {
+      "epoch": 0.62,
+      "objective/train/advantage_avg": 0.00042441111872904,
+      "objective/train/docs_used": 741020,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3329555988311768,
+      "objective/train/original_loss": 1.3329553604125977,
+      "objective/train/theoretical_loss": 3.454382454604577,
+      "objective/train/tokens_used": 2052076000,
+      "objective/train/value_avg": -0.0072784423828125,
+      "objective/train/value_loss": 0.00047027276013977826,
+      "objective/train/value_max": -1.9848346710205078e-05,
+      "objective/train/value_min": -0.70166015625,
+      "objective/train/value_reward_corr": 0.7057014467850871,
+      "objective/train/value_std": 0.018829345703125,
+      "objective/train/weight_avg": 1.0006253719329834,
+      "objective/train/weighted_lm_loss": 1.3332040309906006,
+      "objective/train/weights_max": 1.614859700202942,
+      "objective/train/weights_min": 0.3689524829387665,
+      "theoretical_loss": 3.454382454604577,
+      "tokens_seen": 2031616000
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003882201893757021,
+      "loss": 0.0668,
+      "theoretical_loss": 3.454382454604577,
+      "tokens_seen": 2031616000
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038813994543411974,
+      "loss": 0.0666,
+      "theoretical_loss": 3.454345713727923,
+      "tokens_seen": 2031878144
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038805970149253735,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4543089789181556,
+      "tokens_seen": 2032140288
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003879794575509549,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4542722501734904,
+      "tokens_seen": 2032402432
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003878992136093725,
+      "loss": 0.0682,
+      "theoretical_loss": 3.454235527492145,
+      "tokens_seen": 2032664576
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003878189696677901,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4541988108723354,
+      "tokens_seen": 2032926720
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038773872572620764,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4541621003122804,
+      "tokens_seen": 2033188864
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038765848178462525,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4541253958101983,
+      "tokens_seen": 2033451008
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038757823784304286,
+      "loss": 0.0677,
+      "theoretical_loss": 3.454088697364309,
+      "tokens_seen": 2033713152
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003874979939014605,
+      "loss": 0.0651,
+      "theoretical_loss": 3.454052004972833,
+      "tokens_seen": 2033975296
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038741774995987803,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4540153186339912,
+      "tokens_seen": 2034237440
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038733750601829565,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4539786383460047,
+      "tokens_seen": 2034499584
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003872572620767132,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4539419641070968,
+      "tokens_seen": 2034761728
+    },
+    {
+      "epoch": 0.62,
+      "objective/train/advantage_avg": 0.0009591991547495127,
+      "objective/train/docs_used": 742098,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4628486633300781,
+      "objective/train/original_loss": 1.4628486633300781,
+      "objective/train/theoretical_loss": 3.453923629255492,
+      "objective/train/tokens_used": 2055352800,
+      "objective/train/value_avg": -0.00998687744140625,
+      "objective/train/value_loss": 0.00025923250359483063,
+      "objective/train/value_max": -5.561113357543945e-05,
+      "objective/train/value_min": -0.90869140625,
+      "objective/train/value_reward_corr": 0.6947016315842284,
+      "objective/train/value_std": 0.0157470703125,
+      "objective/train/weight_avg": 1.0010782480239868,
+      "objective/train/weighted_lm_loss": 1.463948369026184,
+      "objective/train/weights_max": 1.1520448923110962,
+      "objective/train/weights_min": 0.36821871995925903,
+      "theoretical_loss": 3.453923629255492,
+      "tokens_seen": 2034892800
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038717701813513076,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4539052959154906,
+      "tokens_seen": 2035023872
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003870967741935484,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4538686337694102,
+      "tokens_seen": 2035286016
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.000387016530251966,
+      "loss": 0.0687,
+      "theoretical_loss": 3.45383197766708,
+      "tokens_seen": 2035548160
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003869362863103836,
+      "loss": 0.0645,
+      "theoretical_loss": 3.453795327606726,
+      "tokens_seen": 2035810304
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038685604236880116,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4537586835865746,
+      "tokens_seen": 2036072448
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038677579842721877,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4537220456048523,
+      "tokens_seen": 2036334592
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003866955544856364,
+      "loss": 0.0685,
+      "theoretical_loss": 3.453685413659788,
+      "tokens_seen": 2036596736
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003866153105440539,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4536487877496085,
+      "tokens_seen": 2036858880
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003865350666024715,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4536121678725444,
+      "tokens_seen": 2037121024
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003864548226608891,
+      "loss": 0.0678,
+      "theoretical_loss": 3.453575554026825,
+      "tokens_seen": 2037383168
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003863745787193067,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4535389462106822,
+      "tokens_seen": 2037645312
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003862943347777243,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4535023444223465,
+      "tokens_seen": 2037907456
+    },
+    {
+      "epoch": 0.62,
+      "objective/train/advantage_avg": -0.0006078130681999028,
+      "objective/train/docs_used": 743253,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2707901000976562,
+      "objective/train/original_loss": 1.2707901000976562,
+      "objective/train/theoretical_loss": 3.453465748660051,
+      "objective/train/tokens_used": 2058629600,
+      "objective/train/value_avg": -0.007007598876953125,
+      "objective/train/value_loss": 0.00013605756976176053,
+      "objective/train/value_max": -5.269050598144531e-05,
+      "objective/train/value_min": -0.456787109375,
+      "objective/train/value_reward_corr": 0.7220390549582588,
+      "objective/train/value_std": 0.01114654541015625,
+      "objective/train/weight_avg": 0.9994558691978455,
+      "objective/train/weighted_lm_loss": 1.2703384160995483,
+      "objective/train/weights_max": 1.1364604234695435,
+      "objective/train/weights_min": 0.3713054358959198,
+      "theoretical_loss": 3.453465748660051,
+      "tokens_seen": 2038169600
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003862140908361419,
+      "loss": 0.0676,
+      "theoretical_loss": 3.453465748660051,
+      "tokens_seen": 2038169600
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003861338468945595,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4534291589220274,
+      "tokens_seen": 2038431744
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.000386053602952977,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4533925752065104,
+      "tokens_seen": 2038693888
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038597335901139463,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4533559975117347,
+      "tokens_seen": 2038956032
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038589311506981224,
+      "loss": 0.0696,
+      "theoretical_loss": 3.453319425835935,
+      "tokens_seen": 2039218176
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038581287112822985,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4532828601773478,
+      "tokens_seen": 2039480320
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003857326271866474,
+      "loss": 0.0664,
+      "theoretical_loss": 3.453246300534209,
+      "tokens_seen": 2039742464
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.000385652383245065,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4532097469047573,
+      "tokens_seen": 2040004608
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038557213930348264,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4531731992872303,
+      "tokens_seen": 2040266752
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038549189536190014,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4531366576798668,
+      "tokens_seen": 2040528896
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038541165142031775,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4531001220809068,
+      "tokens_seen": 2040791040
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038533140747873537,
+      "loss": 0.0685,
+      "theoretical_loss": 3.453063592488591,
+      "tokens_seen": 2041053184
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003852511635371529,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4530270689011595,
+      "tokens_seen": 2041315328
+    },
+    {
+      "epoch": 0.62,
+      "objective/train/advantage_avg": 0.0007066431571729481,
+      "objective/train/docs_used": 744490,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.501568078994751,
+      "objective/train/original_loss": 1.501568078994751,
+      "objective/train/theoretical_loss": 3.4530088093587263,
+      "objective/train/tokens_used": 2061906400,
+      "objective/train/value_avg": -0.007480621337890625,
+      "objective/train/value_loss": 0.0003073073457926512,
+      "objective/train/value_max": -4.267692565917969e-05,
+      "objective/train/value_min": -0.266845703125,
+      "objective/train/value_reward_corr": 0.5380204613048158,
+      "objective/train/value_std": 0.01178741455078125,
+      "objective/train/weight_avg": 1.0008387565612793,
+      "objective/train/weighted_lm_loss": 1.501943826675415,
+      "objective/train/weights_max": 1.2743744850158691,
+      "objective/train/weights_min": 0.37047773599624634,
+      "theoretical_loss": 3.4530088093587263,
+      "tokens_seen": 2041446400
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038517091959557054,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4529905513168555,
+      "tokens_seen": 2041577472
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038509067565398815,
+      "loss": 0.0675,
+      "theoretical_loss": 3.452954039733921,
+      "tokens_seen": 2041839616
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038501043171240576,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4529175341505995,
+      "tokens_seen": 2042101760
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003849301877708233,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4528810345651357,
+      "tokens_seen": 2042363904
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003848499438292409,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4528445409757738,
+      "tokens_seen": 2042626048
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003847696998876585,
+      "loss": 0.0691,
+      "theoretical_loss": 3.45280805338076,
+      "tokens_seen": 2042888192
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038468945594607605,
+      "loss": 0.0659,
+      "theoretical_loss": 3.45277157177834,
+      "tokens_seen": 2043150336
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038460921200449366,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4527350961667613,
+      "tokens_seen": 2043412480
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003845289680629113,
+      "loss": 0.0669,
+      "theoretical_loss": 3.452698626544272,
+      "tokens_seen": 2043674624
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003844487241213289,
+      "loss": 0.0674,
+      "theoretical_loss": 3.45266216290912,
+      "tokens_seen": 2043936768
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038436848017974645,
+      "loss": 0.0683,
+      "theoretical_loss": 3.452625705259556,
+      "tokens_seen": 2044198912
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.000384288236238164,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4525892535938283,
+      "tokens_seen": 2044461056
+    },
+    {
+      "epoch": 0.62,
+      "objective/train/advantage_avg": 0.0002817703934852034,
+      "objective/train/docs_used": 745676,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1205445528030396,
+      "objective/train/original_loss": 1.1205445528030396,
+      "objective/train/theoretical_loss": 3.4525528079101884,
+      "objective/train/tokens_used": 2065183200,
+      "objective/train/value_avg": -0.006732940673828125,
+      "objective/train/value_loss": 0.00019041493942495435,
+      "objective/train/value_max": -6.300210952758789e-05,
+      "objective/train/value_min": -0.61474609375,
+      "objective/train/value_reward_corr": 0.7468571501310386,
+      "objective/train/value_std": 0.01427459716796875,
+      "objective/train/weight_avg": 1.0003718137741089,
+      "objective/train/weighted_lm_loss": 1.1208994388580322,
+      "objective/train/weights_max": 1.6867209672927856,
+      "objective/train/weights_min": 0.49503588676452637,
+      "theoretical_loss": 3.4525528079101884,
+      "tokens_seen": 2044723200
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003842079922965816,
+      "loss": 0.068,
+      "theoretical_loss": 3.4525528079101884,
+      "tokens_seen": 2044723200
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003841277483549992,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4525163682068882,
+      "tokens_seen": 2044985344
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003840475044134168,
+      "loss": 0.0706,
+      "theoretical_loss": 3.45247993448218,
+      "tokens_seen": 2045247488
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003839672604718344,
+      "loss": 0.067,
+      "theoretical_loss": 3.4524435067343164,
+      "tokens_seen": 2045509632
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.000383887016530252,
+      "loss": 0.0644,
+      "theoretical_loss": 3.452407084961551,
+      "tokens_seen": 2045771776
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038380677258866957,
+      "loss": 0.0643,
+      "theoretical_loss": 3.452370669162139,
+      "tokens_seen": 2046033920
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038372652864708713,
+      "loss": 0.0677,
+      "theoretical_loss": 3.452334259334335,
+      "tokens_seen": 2046296064
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038364628470550474,
+      "loss": 0.0677,
+      "theoretical_loss": 3.452297855476395,
+      "tokens_seen": 2046558208
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003835660407639223,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4522614575865753,
+      "tokens_seen": 2046820352
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003834857968223399,
+      "loss": 0.0647,
+      "theoretical_loss": 3.452225065663134,
+      "tokens_seen": 2047082496
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003834055528807575,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4521886797043293,
+      "tokens_seen": 2047344640
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003833253089391751,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4521522997084197,
+      "tokens_seen": 2047606784
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003832450649975927,
+      "loss": 0.0682,
+      "theoretical_loss": 3.452115925673665,
+      "tokens_seen": 2047868928
+    },
+    {
+      "epoch": 0.62,
+      "objective/train/advantage_avg": 0.0006643827073276043,
+      "objective/train/docs_used": 746888,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2512800693511963,
+      "objective/train/original_loss": 1.2512800693511963,
+      "objective/train/theoretical_loss": 3.4520977408911766,
+      "objective/train/tokens_used": 2068460000,
+      "objective/train/value_avg": -0.004673004150390625,
+      "objective/train/value_loss": 9.969693201128393e-05,
+      "objective/train/value_max": -3.451108932495117e-05,
+      "objective/train/value_min": -0.271240234375,
+      "objective/train/value_reward_corr": 0.6657257095051546,
+      "objective/train/value_std": 0.0092926025390625,
+      "objective/train/weight_avg": 1.000710368156433,
+      "objective/train/weighted_lm_loss": 1.2529733180999756,
+      "objective/train/weights_max": 1.1873843669891357,
+      "objective/train/weights_min": 0.393546462059021,
+      "theoretical_loss": 3.4520977408911766,
+      "tokens_seen": 2048000000
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003831648210560103,
+      "loss": 0.063,
+      "theoretical_loss": 3.4520795575983247,
+      "tokens_seen": 2048131072
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038308457711442787,
+      "loss": 0.0709,
+      "theoretical_loss": 3.4520431954806607,
+      "tokens_seen": 2048393216
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003830043331728454,
+      "loss": 0.0681,
+      "theoretical_loss": 3.452006839318935,
+      "tokens_seen": 2048655360
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038292408923126304,
+      "loss": 0.0686,
+      "theoretical_loss": 3.45197048911141,
+      "tokens_seen": 2048917504
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038284384528968065,
+      "loss": 0.0693,
+      "theoretical_loss": 3.451934144856348,
+      "tokens_seen": 2049179648
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003827636013480982,
+      "loss": 0.0679,
+      "theoretical_loss": 3.451897806552014,
+      "tokens_seen": 2049441792
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003826833574065158,
+      "loss": 0.0692,
+      "theoretical_loss": 3.451861474196672,
+      "tokens_seen": 2049703936
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038260311346493344,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4518251477885884,
+      "tokens_seen": 2049966080
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.000382522869523351,
+      "loss": 0.0713,
+      "theoretical_loss": 3.4517888273260287,
+      "tokens_seen": 2050228224
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038244262558176855,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4517525128072593,
+      "tokens_seen": 2050490368
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038236238164018616,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4517162042305483,
+      "tokens_seen": 2050752512
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003822821376986038,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4516799015941646,
+      "tokens_seen": 2051014656
+    },
+    {
+      "epoch": 0.62,
+      "objective/train/advantage_avg": -0.0004441846103873104,
+      "objective/train/docs_used": 748156,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.271807312965393,
+      "objective/train/original_loss": 1.271807312965393,
+      "objective/train/theoretical_loss": 3.451643604896377,
+      "objective/train/tokens_used": 2071736800,
+      "objective/train/value_avg": -0.008514404296875,
+      "objective/train/value_loss": 0.00032554796780459583,
+      "objective/train/value_max": -2.86102294921875e-05,
+      "objective/train/value_min": -0.8369140625,
+      "objective/train/value_reward_corr": 0.689176385382612,
+      "objective/train/value_std": 0.0157318115234375,
+      "objective/train/weight_avg": 0.9996981024742126,
+      "objective/train/weighted_lm_loss": 1.270479679107666,
+      "objective/train/weights_max": 1.6649008989334106,
+      "objective/train/weights_min": 0.36991289258003235,
+      "theoretical_loss": 3.451643604896377,
+      "tokens_seen": 2051276800
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038220189375702134,
+      "loss": 0.0683,
+      "theoretical_loss": 3.451643604896377,
+      "tokens_seen": 2051276800
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038212164981543895,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4516073141354546,
+      "tokens_seen": 2051538944
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038204140587385656,
+      "loss": 0.069,
+      "theoretical_loss": 3.451571029309668,
+      "tokens_seen": 2051801088
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003819611619322742,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4515347504172893,
+      "tokens_seen": 2052063232
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003818809179906917,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4514984774565898,
+      "tokens_seen": 2052325376
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003818006740491093,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4514622104258423,
+      "tokens_seen": 2052587520
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003817204301075269,
+      "loss": 0.0653,
+      "theoretical_loss": 3.451425949323321,
+      "tokens_seen": 2052849664
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038164018616594446,
+      "loss": 0.0666,
+      "theoretical_loss": 3.451389694147298,
+      "tokens_seen": 2053111808
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003815599422243621,
+      "loss": 0.0679,
+      "theoretical_loss": 3.45135344489605,
+      "tokens_seen": 2053373952
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003814796982827797,
+      "loss": 0.069,
+      "theoretical_loss": 3.4513172015678526,
+      "tokens_seen": 2053636096
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038139945434119725,
+      "loss": 0.0651,
+      "theoretical_loss": 3.451280964160981,
+      "tokens_seen": 2053898240
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003813192103996148,
+      "loss": 0.0677,
+      "theoretical_loss": 3.451244732673713,
+      "tokens_seen": 2054160384
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003812389664580324,
+      "loss": 0.0655,
+      "theoretical_loss": 3.451208507104326,
+      "tokens_seen": 2054422528
+    },
+    {
+      "epoch": 0.62,
+      "objective/train/advantage_avg": 0.0003811193164438009,
+      "objective/train/docs_used": 749325,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3119323253631592,
+      "objective/train/original_loss": 1.3119322061538696,
+      "objective/train/theoretical_loss": 3.4511903965382995,
+      "objective/train/tokens_used": 2075013600,
+      "objective/train/value_avg": -0.006206512451171875,
+      "objective/train/value_loss": 0.00013376564311329275,
+      "objective/train/value_max": -3.88026237487793e-05,
+      "objective/train/value_min": -0.200927734375,
+      "objective/train/value_reward_corr": 0.6514871179032515,
+      "objective/train/value_std": 0.00937652587890625,
+      "objective/train/weight_avg": 1.000443458557129,
+      "objective/train/weighted_lm_loss": 1.3129256963729858,
+      "objective/train/weights_max": 1.1168068647384644,
+      "objective/train/weights_min": 0.39825838804244995,
+      "theoretical_loss": 3.4511903965382995,
+      "tokens_seen": 2054553600
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038115872251645003,
+      "loss": 0.0674,
+      "theoretical_loss": 3.451172287451098,
+      "tokens_seen": 2054684672
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003810784785748676,
+      "loss": 0.0697,
+      "theoretical_loss": 3.45113607371231,
+      "tokens_seen": 2054946816
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003809982346332852,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4510998658862397,
+      "tokens_seen": 2055208960
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003809179906917028,
+      "loss": 0.0672,
+      "theoretical_loss": 3.451063663971169,
+      "tokens_seen": 2055471104
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038083774675012037,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4510274679653787,
+      "tokens_seen": 2055733248
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038075750280853793,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4509912778671517,
+      "tokens_seen": 2055995392
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038067725886695554,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4509550936747697,
+      "tokens_seen": 2056257536
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038059701492537315,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4509189153865165,
+      "tokens_seen": 2056519680
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003805167709837907,
+      "loss": 0.0684,
+      "theoretical_loss": 3.450882743000677,
+      "tokens_seen": 2056781824
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003804365270422083,
+      "loss": 0.0706,
+      "theoretical_loss": 3.450846576515535,
+      "tokens_seen": 2057043968
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038035628310062594,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4508104159293773,
+      "tokens_seen": 2057306112
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003802760391590435,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4507742612404897,
+      "tokens_seen": 2057568256
+    },
+    {
+      "epoch": 0.62,
+      "objective/train/advantage_avg": 0.0004177381924819201,
+      "objective/train/docs_used": 750571,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4458907842636108,
+      "objective/train/original_loss": 1.4458907842636108,
+      "objective/train/theoretical_loss": 3.4507381124471594,
+      "objective/train/tokens_used": 2078290400,
+      "objective/train/value_avg": -0.006328582763671875,
+      "objective/train/value_loss": 0.00010109315917361528,
+      "objective/train/value_max": -4.100799560546875e-05,
+      "objective/train/value_min": -0.20556640625,
+      "objective/train/value_reward_corr": 0.7573714555918173,
+      "objective/train/value_std": 0.01088714599609375,
+      "objective/train/weight_avg": 1.00046706199646,
+      "objective/train/weighted_lm_loss": 1.4466280937194824,
+      "objective/train/weights_max": 1.1097402572631836,
+      "objective/train/weights_min": 0.7170053720474243,
+      "theoretical_loss": 3.4507381124471594,
+      "tokens_seen": 2057830400
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003801957952174611,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4507381124471594,
+      "tokens_seen": 2057830400
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00038011555127587867,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4507019695476737,
+      "tokens_seen": 2058092544
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003800353073342963,
+      "loss": 0.0685,
+      "theoretical_loss": 3.450665832540322,
+      "tokens_seen": 2058354688
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00037995506339271384,
+      "loss": 0.0671,
+      "theoretical_loss": 3.450629701423393,
+      "tokens_seen": 2058616832
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00037987481945113145,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4505935761951765,
+      "tokens_seen": 2058878976
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00037979457550954906,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4505574568539634,
+      "tokens_seen": 2059141120
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003797143315679666,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4505213433980453,
+      "tokens_seen": 2059403264
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00037963408762638423,
+      "loss": 0.07,
+      "theoretical_loss": 3.450485235825714,
+      "tokens_seen": 2059665408
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003795538436848018,
+      "loss": 0.0669,
+      "theoretical_loss": 3.450449134135262,
+      "tokens_seen": 2059927552
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00037947359974321935,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4504130383249834,
+      "tokens_seen": 2060189696
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00037939335580163696,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4503769483931723,
+      "tokens_seen": 2060451840
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003793131118600546,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4503408643381235,
+      "tokens_seen": 2060713984
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003792328679184722,
+      "loss": 0.066,
+      "theoretical_loss": 3.4503047861581324,
+      "tokens_seen": 2060976128
+    },
+    {
+      "epoch": 0.62,
+      "objective/train/advantage_avg": 8.919855463318527e-05,
+      "objective/train/docs_used": 751685,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3986313343048096,
+      "objective/train/original_loss": 1.39863121509552,
+      "objective/train/theoretical_loss": 3.450286749270751,
+      "objective/train/tokens_used": 2081567200,
+      "objective/train/value_avg": -0.0098876953125,
+      "objective/train/value_loss": 0.0007862445199862123,
+      "objective/train/value_max": -4.4345855712890625e-05,
+      "objective/train/value_min": -0.83544921875,
+      "objective/train/value_reward_corr": 0.5908050268196616,
+      "objective/train/value_std": 0.0193023681640625,
+      "objective/train/weight_avg": 1.0004451274871826,
+      "objective/train/weighted_lm_loss": 1.4001002311706543,
+      "objective/train/weights_max": 1.6845263242721558,
+      "objective/train/weights_min": 0.37146979570388794,
+      "theoretical_loss": 3.450286749270751,
+      "tokens_seen": 2061107200
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00037915262397688975,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4502687138514956,
+      "tokens_seen": 2061238272
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.00037907238003530736,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4502326474165104,
+      "tokens_seen": 2061500416
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003789921360937249,
+      "loss": 0.0694,
+      "theoretical_loss": 3.450196586851474,
+      "tokens_seen": 2061762560
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003789118921521425,
+      "loss": 0.069,
+      "theoretical_loss": 3.450160532154685,
+      "tokens_seen": 2062024704
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 0.0003788316482105601,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4501244833244438,
+      "tokens_seen": 2062286848
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003787514042689777,
+      "loss": 0.0688,
+      "theoretical_loss": 3.450088440359049,
+      "tokens_seen": 2062548992
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003786711603273953,
+      "loss": 0.0665,
+      "theoretical_loss": 3.450052403256801,
+      "tokens_seen": 2062811136
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003785909163858129,
+      "loss": 0.0673,
+      "theoretical_loss": 3.450016372016002,
+      "tokens_seen": 2063073280
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003785106724442305,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4499803466349537,
+      "tokens_seen": 2063335424
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003784304285026481,
+      "loss": 0.0696,
+      "theoretical_loss": 3.449944327111959,
+      "tokens_seen": 2063597568
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003783501845610656,
+      "loss": 0.0677,
+      "theoretical_loss": 3.449908313445321,
+      "tokens_seen": 2063859712
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003782699406194832,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4498723056333445,
+      "tokens_seen": 2064121856
+    },
+    {
+      "epoch": 0.63,
+      "objective/train/advantage_avg": 0.001222593942657113,
+      "objective/train/docs_used": 752866,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3336776494979858,
+      "objective/train/original_loss": 1.3336775302886963,
+      "objective/train/theoretical_loss": 3.4498363036743336,
+      "objective/train/tokens_used": 2084844000,
+      "objective/train/value_avg": -0.0068817138671875,
+      "objective/train/value_loss": 0.0002020899555645883,
+      "objective/train/value_max": -1.996755599975586e-05,
+      "objective/train/value_min": -0.223388671875,
+      "objective/train/value_reward_corr": 0.5299720695477663,
+      "objective/train/value_std": 0.0103607177734375,
+      "objective/train/weight_avg": 1.0013130903244019,
+      "objective/train/weighted_lm_loss": 1.3344902992248535,
+      "objective/train/weights_max": 1.1334251165390015,
+      "objective/train/weights_min": 0.3700779974460602,
+      "theoretical_loss": 3.4498363036743336,
+      "tokens_seen": 2064384000
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037818969667790083,
+      "loss": 0.0716,
+      "theoretical_loss": 3.4498363036743336,
+      "tokens_seen": 2064384000
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037810945273631844,
+      "loss": 0.0667,
+      "theoretical_loss": 3.449800307566594,
+      "tokens_seen": 2064646144
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.000378029208794736,
+      "loss": 0.0679,
+      "theoretical_loss": 3.449764317308432,
+      "tokens_seen": 2064908288
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003779489648531536,
+      "loss": 0.0701,
+      "theoretical_loss": 3.449728332898155,
+      "tokens_seen": 2065170432
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003778687209115712,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4496923543340703,
+      "tokens_seen": 2065432576
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037778847696998873,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4496563816144867,
+      "tokens_seen": 2065694720
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037770823302840634,
+      "loss": 0.0676,
+      "theoretical_loss": 3.449620414737713,
+      "tokens_seen": 2065956864
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037762798908682395,
+      "loss": 0.0661,
+      "theoretical_loss": 3.449584453702059,
+      "tokens_seen": 2066219008
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003775477451452415,
+      "loss": 0.0674,
+      "theoretical_loss": 3.449548498505834,
+      "tokens_seen": 2066481152
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003774675012036591,
+      "loss": 0.0714,
+      "theoretical_loss": 3.4495125491473515,
+      "tokens_seen": 2066743296
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037738725726207674,
+      "loss": 0.0672,
+      "theoretical_loss": 3.449476605624922,
+      "tokens_seen": 2067005440
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037730701332049435,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4494406679368583,
+      "tokens_seen": 2067267584
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003772267693789119,
+      "loss": 0.0696,
+      "theoretical_loss": 3.449404736081474,
+      "tokens_seen": 2067529728
+    },
+    {
+      "epoch": 0.63,
+      "objective/train/advantage_avg": 0.001186740817502141,
+      "objective/train/docs_used": 753921,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5087406635284424,
+      "objective/train/original_loss": 1.5087406635284424,
+      "objective/train/theoretical_loss": 3.4493867723405094,
+      "objective/train/tokens_used": 2088120800,
+      "objective/train/value_avg": -0.01195526123046875,
+      "objective/train/value_loss": 0.000287308037513867,
+      "objective/train/value_max": -2.3186206817626953e-05,
+      "objective/train/value_min": -0.3701171875,
+      "objective/train/value_reward_corr": 0.7702874414541618,
+      "objective/train/value_std": 0.023345947265625,
+      "objective/train/weight_avg": 1.0013281106948853,
+      "objective/train/weighted_lm_loss": 1.509456753730774,
+      "objective/train/weights_max": 1.3135128021240234,
+      "objective/train/weights_min": 0.6228498816490173,
+      "theoretical_loss": 3.4493867723405094,
+      "tokens_seen": 2067660800
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037714652543732947,
+      "loss": 0.0713,
+      "theoretical_loss": 3.4493688100570825,
+      "tokens_seen": 2067791872
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003770662814957471,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4493328898619993,
+      "tokens_seen": 2068054016
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037698603755416464,
+      "loss": 0.0677,
+      "theoretical_loss": 3.449296975494539,
+      "tokens_seen": 2068316160
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037690579361258225,
+      "loss": 0.0666,
+      "theoretical_loss": 3.449261066953018,
+      "tokens_seen": 2068578304
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037682554967099986,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4492251642357536,
+      "tokens_seen": 2068840448
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003767453057294175,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4491892673410627,
+      "tokens_seen": 2069102592
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037666506178783503,
+      "loss": 0.069,
+      "theoretical_loss": 3.449153376267264,
+      "tokens_seen": 2069364736
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003765848178462526,
+      "loss": 0.068,
+      "theoretical_loss": 3.449117491012676,
+      "tokens_seen": 2069626880
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003765045739046702,
+      "loss": 0.0673,
+      "theoretical_loss": 3.449081611575618,
+      "tokens_seen": 2069889024
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037642432996308776,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4490457379544113,
+      "tokens_seen": 2070151168
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003763440860215054,
+      "loss": 0.0706,
+      "theoretical_loss": 3.4490098701473757,
+      "tokens_seen": 2070413312
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.000376263842079923,
+      "loss": 0.0672,
+      "theoretical_loss": 3.448974008152834,
+      "tokens_seen": 2070675456
+    },
+    {
+      "epoch": 0.63,
+      "objective/train/advantage_avg": 0.0002597297716420144,
+      "objective/train/docs_used": 755047,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2841310501098633,
+      "objective/train/original_loss": 1.2841309309005737,
+      "objective/train/theoretical_loss": 3.4489381519691085,
+      "objective/train/tokens_used": 2091397600,
+      "objective/train/value_avg": -0.0073089599609375,
+      "objective/train/value_loss": 0.00023801308998372406,
+      "objective/train/value_max": -6.401538848876953e-05,
+      "objective/train/value_min": -0.2763671875,
+      "objective/train/value_reward_corr": 0.7057852962488432,
+      "objective/train/value_std": 0.013275146484375,
+      "objective/train/weight_avg": 1.0003691911697388,
+      "objective/train/weighted_lm_loss": 1.2843565940856934,
+      "objective/train/weights_max": 1.14552640914917,
+      "objective/train/weights_min": 0.3723408579826355,
+      "theoretical_loss": 3.4489381519691085,
+      "tokens_seen": 2070937600
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003761835981383406,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4489381519691085,
+      "tokens_seen": 2070937600
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037610335419675816,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4489023015945213,
+      "tokens_seen": 2071199744
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003760231102551757,
+      "loss": 0.0682,
+      "theoretical_loss": 3.448866457027397,
+      "tokens_seen": 2071461888
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037594286631359333,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4488306182660597,
+      "tokens_seen": 2071724032
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003758626223720109,
+      "loss": 0.0713,
+      "theoretical_loss": 3.448794785308835,
+      "tokens_seen": 2071986176
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003757823784304285,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4487589581540483,
+      "tokens_seen": 2072248320
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003757021344888461,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4487231368000266,
+      "tokens_seen": 2072510464
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003756218905472637,
+      "loss": 0.0674,
+      "theoretical_loss": 3.448687321245097,
+      "tokens_seen": 2072772608
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003755416466056813,
+      "loss": 0.066,
+      "theoretical_loss": 3.4486515114875873,
+      "tokens_seen": 2073034752
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003754614026640989,
+      "loss": 0.0704,
+      "theoretical_loss": 3.448615707525826,
+      "tokens_seen": 2073296896
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037538115872251646,
+      "loss": 0.0686,
+      "theoretical_loss": 3.448579909358143,
+      "tokens_seen": 2073559040
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.000375300914780934,
+      "loss": 0.0685,
+      "theoretical_loss": 3.448544116982868,
+      "tokens_seen": 2073821184
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037522067083935163,
+      "loss": 0.0696,
+      "theoretical_loss": 3.448508330398332,
+      "tokens_seen": 2074083328
+    },
+    {
+      "epoch": 0.63,
+      "objective/train/advantage_avg": 6.10949209658429e-05,
+      "objective/train/docs_used": 756110,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3281543254852295,
+      "objective/train/original_loss": 1.3281540870666504,
+      "objective/train/theoretical_loss": 3.4484904392770694,
+      "objective/train/tokens_used": 2094674400,
+      "objective/train/value_avg": -0.00695037841796875,
+      "objective/train/value_loss": 0.00031872920226305723,
+      "objective/train/value_max": -4.100799560546875e-05,
+      "objective/train/value_min": -0.329345703125,
+      "objective/train/value_reward_corr": 0.6203797536959921,
+      "objective/train/value_std": 0.0136566162109375,
+      "objective/train/weight_avg": 1.0001968145370483,
+      "objective/train/weighted_lm_loss": 1.328813076019287,
+      "objective/train/weights_max": 1.3900582790374756,
+      "objective/train/weights_min": 0.372374951839447,
+      "theoretical_loss": 3.4484904392770694,
+      "tokens_seen": 2074214400
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037514042689776924,
+      "loss": 0.0672,
+      "theoretical_loss": 3.448472549602866,
+      "tokens_seen": 2074345472
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003750601829561868,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4484367745948026,
+      "tokens_seen": 2074607616
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003749799390146044,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4484010053724736,
+      "tokens_seen": 2074869760
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.000374899695073022,
+      "loss": 0.0687,
+      "theoretical_loss": 3.448365241934214,
+      "tokens_seen": 2075131904
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003748194511314396,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4483294842783563,
+      "tokens_seen": 2075394048
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037473920718985714,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4482937324032368,
+      "tokens_seen": 2075656192
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037465896324827475,
+      "loss": 0.0709,
+      "theoretical_loss": 3.4482579863071905,
+      "tokens_seen": 2075918336
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037457871930669237,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4482222459885534,
+      "tokens_seen": 2076180480
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003744984753651099,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4481865114456625,
+      "tokens_seen": 2076442624
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037441823142352754,
+      "loss": 0.0671,
+      "theoretical_loss": 3.448150782676856,
+      "tokens_seen": 2076704768
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037433798748194515,
+      "loss": 0.0641,
+      "theoretical_loss": 3.4481150596804717,
+      "tokens_seen": 2076966912
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003742577435403627,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4480793424548493,
+      "tokens_seen": 2077229056
+    },
+    {
+      "epoch": 0.63,
+      "objective/train/advantage_avg": 0.0016278502298519015,
+      "objective/train/docs_used": 757304,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4312312602996826,
+      "objective/train/original_loss": 1.4312312602996826,
+      "objective/train/theoretical_loss": 3.4480436309983276,
+      "objective/train/tokens_used": 2097951200,
+      "objective/train/value_avg": -0.007061004638671875,
+      "objective/train/value_loss": 0.00019786780467256904,
+      "objective/train/value_max": -4.6133995056152344e-05,
+      "objective/train/value_min": -0.6015625,
+      "objective/train/value_reward_corr": 0.6384379863088263,
+      "objective/train/value_std": 0.01360321044921875,
+      "objective/train/weight_avg": 1.0017173290252686,
+      "objective/train/weighted_lm_loss": 1.4329873323440552,
+      "objective/train/weights_max": 1.4410042762756348,
+      "objective/train/weights_min": 0.3698126971721649,
+      "theoretical_loss": 3.4480436309983276,
+      "tokens_seen": 2077491200
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037417749959878027,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4480436309983276,
+      "tokens_seen": 2077491200
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003740972556571979,
+      "loss": 0.0688,
+      "theoretical_loss": 3.448007925309247,
+      "tokens_seen": 2077753344
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003740170117156155,
+      "loss": 0.0694,
+      "theoretical_loss": 3.447972225385949,
+      "tokens_seen": 2078015488
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037393676777403305,
+      "loss": 0.065,
+      "theoretical_loss": 3.447936531226776,
+      "tokens_seen": 2078277632
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037385652383245066,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4479008428300686,
+      "tokens_seen": 2078539776
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003737762798908683,
+      "loss": 0.0677,
+      "theoretical_loss": 3.447865160194171,
+      "tokens_seen": 2078801920
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003736960359492859,
+      "loss": 0.0687,
+      "theoretical_loss": 3.447829483317428,
+      "tokens_seen": 2079064064
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003736157920077034,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4477938121981824,
+      "tokens_seen": 2079326208
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.000373535548066121,
+      "loss": 0.0644,
+      "theoretical_loss": 3.44775814683478,
+      "tokens_seen": 2079588352
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003734553041245386,
+      "loss": 0.0665,
+      "theoretical_loss": 3.447722487225567,
+      "tokens_seen": 2079850496
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003733750601829562,
+      "loss": 0.064,
+      "theoretical_loss": 3.4476868333688904,
+      "tokens_seen": 2080112640
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003732948162413738,
+      "loss": 0.0685,
+      "theoretical_loss": 3.447651185263096,
+      "tokens_seen": 2080374784
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003732145722997914,
+      "loss": 0.0677,
+      "theoretical_loss": 3.447615542906532,
+      "tokens_seen": 2080636928
+    },
+    {
+      "epoch": 0.63,
+      "objective/train/advantage_avg": 0.0009759876993484795,
+      "objective/train/docs_used": 758504,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2792775630950928,
+      "objective/train/original_loss": 1.2792774438858032,
+      "objective/train/theoretical_loss": 3.447597723883696,
+      "objective/train/tokens_used": 2101228000,
+      "objective/train/value_avg": -0.0057220458984375,
+      "objective/train/value_loss": 0.00013313665112946182,
+      "objective/train/value_max": -7.367134094238281e-05,
+      "objective/train/value_min": -0.2230224609375,
+      "objective/train/value_reward_corr": 0.5767379735236927,
+      "objective/train/value_std": 0.00873565673828125,
+      "objective/train/weight_avg": 1.0010370016098022,
+      "objective/train/weighted_lm_loss": 1.2802315950393677,
+      "objective/train/weights_max": 1.1462258100509644,
+      "objective/train/weights_min": 0.3722386062145233,
+      "theoretical_loss": 3.447597723883696,
+      "tokens_seen": 2080768000
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037313432835820896,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4475799062975483,
+      "tokens_seen": 2080899072
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003730540844166265,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4475442754344927,
+      "tokens_seen": 2081161216
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037297384047504413,
+      "loss": 0.0641,
+      "theoretical_loss": 3.447508650315716,
+      "tokens_seen": 2081423360
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037289359653346174,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4474730309395687,
+      "tokens_seen": 2081685504
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003728133525918793,
+      "loss": 0.0708,
+      "theoretical_loss": 3.4474374173044025,
+      "tokens_seen": 2081947648
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003727331086502969,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4474018094085683,
+      "tokens_seen": 2082209792
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003726528647087145,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4473662072504196,
+      "tokens_seen": 2082471936
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003725726207671321,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4473306108283097,
+      "tokens_seen": 2082734080
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003724923768255497,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4472950201405923,
+      "tokens_seen": 2082996224
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037241213288396726,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4472594351856225,
+      "tokens_seen": 2083258368
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037233188894238487,
+      "loss": 0.0686,
+      "theoretical_loss": 3.447223855961756,
+      "tokens_seen": 2083520512
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003722516450008024,
+      "loss": 0.0675,
+      "theoretical_loss": 3.447188282467348,
+      "tokens_seen": 2083782656
+    },
+    {
+      "epoch": 0.63,
+      "objective/train/advantage_avg": 0.000588573922868818,
+      "objective/train/docs_used": 759560,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3722100257873535,
+      "objective/train/original_loss": 1.3722100257873535,
+      "objective/train/theoretical_loss": 3.4471527147007555,
+      "objective/train/tokens_used": 2104504800,
+      "objective/train/value_avg": -0.00565338134765625,
+      "objective/train/value_loss": 0.00010087020200444385,
+      "objective/train/value_max": -6.014108657836914e-05,
+      "objective/train/value_min": -0.7431640625,
+      "objective/train/value_reward_corr": 0.7463114972528345,
+      "objective/train/value_std": 0.01171875,
+      "objective/train/weight_avg": 1.0006376504898071,
+      "objective/train/weighted_lm_loss": 1.3735826015472412,
+      "objective/train/weights_max": 1.217918872833252,
+      "objective/train/weights_min": 0.534896731376648,
+      "theoretical_loss": 3.4471527147007555,
+      "tokens_seen": 2084044800
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037217140105922004,
+      "loss": 0.0711,
+      "theoretical_loss": 3.4471527147007555,
+      "tokens_seen": 2084044800
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037209115711763765,
+      "loss": 0.0671,
+      "theoretical_loss": 3.447117152660337,
+      "tokens_seen": 2084306944
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003720109131760552,
+      "loss": 0.0685,
+      "theoretical_loss": 3.447081596344449,
+      "tokens_seen": 2084569088
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003719306692344728,
+      "loss": 0.0679,
+      "theoretical_loss": 3.447046045751451,
+      "tokens_seen": 2084831232
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003718504252928904,
+      "loss": 0.0659,
+      "theoretical_loss": 3.447010500879703,
+      "tokens_seen": 2085093376
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.000371770181351308,
+      "loss": 0.065,
+      "theoretical_loss": 3.4469749617275642,
+      "tokens_seen": 2085355520
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037168993740972555,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4469394282933967,
+      "tokens_seen": 2085617664
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037160969346814316,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4469039005755606,
+      "tokens_seen": 2085879808
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003715294495265608,
+      "loss": 0.069,
+      "theoretical_loss": 3.446868378572419,
+      "tokens_seen": 2086141952
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037144920558497834,
+      "loss": 0.0692,
+      "theoretical_loss": 3.446832862282334,
+      "tokens_seen": 2086404096
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037136896164339595,
+      "loss": 0.0657,
+      "theoretical_loss": 3.44679735170367,
+      "tokens_seen": 2086666240
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003712887177018135,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4467618468347903,
+      "tokens_seen": 2086928384
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037120847376023107,
+      "loss": 0.0668,
+      "theoretical_loss": 3.44672634767406,
+      "tokens_seen": 2087190528
+    },
+    {
+      "epoch": 0.63,
+      "objective/train/advantage_avg": 0.0005934222135692835,
+      "objective/train/docs_used": 760861,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2901841402053833,
+      "objective/train/original_loss": 1.2901840209960938,
+      "objective/train/theoretical_loss": 3.446708600233741,
+      "objective/train/tokens_used": 2107781600,
+      "objective/train/value_avg": -0.00887298583984375,
+      "objective/train/value_loss": 0.0004026810929644853,
+      "objective/train/value_max": -5.918741226196289e-05,
+      "objective/train/value_min": -0.87939453125,
+      "objective/train/value_reward_corr": 0.670880919662618,
+      "objective/train/value_std": 0.016845703125,
+      "objective/train/weight_avg": 1.0007684230804443,
+      "objective/train/weighted_lm_loss": 1.2911221981048584,
+      "objective/train/weights_max": 1.1675496101379395,
+      "objective/train/weights_min": 0.3683558702468872,
+      "theoretical_loss": 3.446708600233741,
+      "tokens_seen": 2087321600
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003711282298186487,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4466908542198453,
+      "tokens_seen": 2087452672
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003710479858770663,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4466553664705124,
+      "tokens_seen": 2087714816
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003709677419354839,
+      "loss": 0.067,
+      "theoretical_loss": 3.446619884424427,
+      "tokens_seen": 2087976960
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037088749799390146,
+      "loss": 0.0678,
+      "theoretical_loss": 3.446584408079958,
+      "tokens_seen": 2088239104
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003708072540523191,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4465489374354727,
+      "tokens_seen": 2088501248
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003707270101107367,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4465134724893405,
+      "tokens_seen": 2088763392
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003706467661691542,
+      "loss": 0.0681,
+      "theoretical_loss": 3.446478013239931,
+      "tokens_seen": 2089025536
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003705665222275718,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4464425596856136,
+      "tokens_seen": 2089287680
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003704862782859894,
+      "loss": 0.0691,
+      "theoretical_loss": 3.446407111824761,
+      "tokens_seen": 2089549824
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037040603434440703,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4463716696557425,
+      "tokens_seen": 2089811968
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003703257904028246,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4463362331769325,
+      "tokens_seen": 2090074112
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003702455464612422,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4463008023867028,
+      "tokens_seen": 2090336256
+    },
+    {
+      "epoch": 0.63,
+      "objective/train/advantage_avg": 0.0002297655591974035,
+      "objective/train/docs_used": 762048,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4020442962646484,
+      "objective/train/original_loss": 1.4020442962646484,
+      "objective/train/theoretical_loss": 3.4462653772834266,
+      "objective/train/tokens_used": 2111058400,
+      "objective/train/value_avg": -0.00919342041015625,
+      "objective/train/value_loss": 0.00034913301351480186,
+      "objective/train/value_max": -3.0219554901123047e-05,
+      "objective/train/value_min": -0.4814453125,
+      "objective/train/value_reward_corr": 0.747774063492229,
+      "objective/train/value_std": 0.01910400390625,
+      "objective/train/weight_avg": 1.0003926753997803,
+      "objective/train/weighted_lm_loss": 1.4031243324279785,
+      "objective/train/weights_max": 1.618411898612976,
+      "objective/train/weights_min": 0.3690573275089264,
+      "theoretical_loss": 3.4462653772834266,
+      "tokens_seen": 2090598400
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003701653025196598,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4462653772834266,
+      "tokens_seen": 2090598400
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003700850585780773,
+      "loss": 0.0679,
+      "theoretical_loss": 3.446229957865479,
+      "tokens_seen": 2090860544
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00037000481463649493,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4461945441312354,
+      "tokens_seen": 2091122688
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00036992457069491254,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4461591360790704,
+      "tokens_seen": 2091384832
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00036984432675333015,
+      "loss": 0.0657,
+      "theoretical_loss": 3.44612373370736,
+      "tokens_seen": 2091646976
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003697640828117477,
+      "loss": 0.0687,
+      "theoretical_loss": 3.446088337014482,
+      "tokens_seen": 2091909120
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003696838388701653,
+      "loss": 0.0685,
+      "theoretical_loss": 3.446052945998814,
+      "tokens_seen": 2092171264
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00036960359492858294,
+      "loss": 0.0681,
+      "theoretical_loss": 3.446017560658734,
+      "tokens_seen": 2092433408
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00036952335098700044,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4459821809926208,
+      "tokens_seen": 2092695552
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00036944310704541805,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4459468069988537,
+      "tokens_seen": 2092957696
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00036936286310383567,
+      "loss": 0.0667,
+      "theoretical_loss": 3.445911438675814,
+      "tokens_seen": 2093219840
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003692826191622532,
+      "loss": 0.0639,
+      "theoretical_loss": 3.445876076021882,
+      "tokens_seen": 2093481984
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00036920237522067084,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4458407190354388,
+      "tokens_seen": 2093744128
+    },
+    {
+      "epoch": 0.63,
+      "objective/train/advantage_avg": 0.00022437794541474432,
+      "objective/train/docs_used": 763180,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3575186729431152,
+      "objective/train/original_loss": 1.3575185537338257,
+      "objective/train/theoretical_loss": 3.4458230426670204,
+      "objective/train/tokens_used": 2114335200,
+      "objective/train/value_avg": -0.007144927978515625,
+      "objective/train/value_loss": 0.00013656125520356,
+      "objective/train/value_max": -6.204843521118164e-05,
+      "objective/train/value_min": -0.31640625,
+      "objective/train/value_reward_corr": 0.7161289931619385,
+      "objective/train/value_std": 0.01316070556640625,
+      "objective/train/weight_avg": 1.0002896785736084,
+      "objective/train/weighted_lm_loss": 1.3582242727279663,
+      "objective/train/weights_max": 1.2010502815246582,
+      "objective/train/weights_min": 0.4233749806880951,
+      "theoretical_loss": 3.4458230426670204,
+      "tokens_seen": 2093875200
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00036912213127908845,
+      "loss": 0.0666,
+      "theoretical_loss": 3.445805367714868,
+      "tokens_seen": 2094006272
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00036904188733750606,
+      "loss": 0.0662,
+      "theoretical_loss": 3.445770022058551,
+      "tokens_seen": 2094268416
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003689616433959236,
+      "loss": 0.0677,
+      "theoretical_loss": 3.445734682064873,
+      "tokens_seen": 2094530560
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003688813994543412,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4456993477322166,
+      "tokens_seen": 2094792704
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.0003688011555127588,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4456640190589676,
+      "tokens_seen": 2095054848
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 0.00036872091157117635,
+      "loss": 0.0686,
+      "theoretical_loss": 3.445628696043512,
+      "tokens_seen": 2095316992
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036864066762959396,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4455933786842348,
+      "tokens_seen": 2095579136
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003685604236880116,
+      "loss": 0.07,
+      "theoretical_loss": 3.4455580669795243,
+      "tokens_seen": 2095841280
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003684801797464292,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4455227609277674,
+      "tokens_seen": 2096103424
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036839993580484675,
+      "loss": 0.071,
+      "theoretical_loss": 3.445487460527352,
+      "tokens_seen": 2096365568
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003683196918632643,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4454521657766675,
+      "tokens_seen": 2096627712
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003682394479216819,
+      "loss": 0.0699,
+      "theoretical_loss": 3.445416876674104,
+      "tokens_seen": 2096889856
+    },
+    {
+      "epoch": 0.64,
+      "objective/train/advantage_avg": 0.0014173869276419282,
+      "objective/train/docs_used": 764413,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4790658950805664,
+      "objective/train/original_loss": 1.4790658950805664,
+      "objective/train/theoretical_loss": 3.44538159321805,
+      "objective/train/tokens_used": 2117612000,
+      "objective/train/value_avg": -0.007617950439453125,
+      "objective/train/value_loss": 0.00010832334373844787,
+      "objective/train/value_max": -7.033348083496094e-05,
+      "objective/train/value_min": -0.2322998046875,
+      "objective/train/value_reward_corr": 0.7426590004130484,
+      "objective/train/value_std": 0.013214111328125,
+      "objective/train/weight_avg": 1.0014710426330566,
+      "objective/train/weighted_lm_loss": 1.4811238050460815,
+      "objective/train/weights_max": 1.1480461359024048,
+      "objective/train/weights_min": 0.8195471167564392,
+      "theoretical_loss": 3.44538159321805,
+      "tokens_seen": 2097152000
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003681592039800995,
+      "loss": 0.0677,
+      "theoretical_loss": 3.44538159321805,
+      "tokens_seen": 2097152000
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003680789600385171,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4453463154068977,
+      "tokens_seen": 2097414144
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003679987160969347,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4453110432390384,
+      "tokens_seen": 2097676288
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003679184721553523,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4452757767128643,
+      "tokens_seen": 2097938432
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003678382282137699,
+      "loss": 0.0643,
+      "theoretical_loss": 3.445240515826768,
+      "tokens_seen": 2098200576
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003677579842721875,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4452052605791432,
+      "tokens_seen": 2098462720
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036767774033060504,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4451700109683836,
+      "tokens_seen": 2098724864
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003675974963890226,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4451347669928856,
+      "tokens_seen": 2098987008
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003675172524474402,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4450995286510424,
+      "tokens_seen": 2099249152
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036743700850585783,
+      "loss": 0.0688,
+      "theoretical_loss": 3.445064295941252,
+      "tokens_seen": 2099511296
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036735676456427544,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4450290688619103,
+      "tokens_seen": 2099773440
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.000367276520622693,
+      "loss": 0.0637,
+      "theoretical_loss": 3.444993847411415,
+      "tokens_seen": 2100035584
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003671962766811106,
+      "loss": 0.069,
+      "theoretical_loss": 3.4449586315881637,
+      "tokens_seen": 2100297728
+    },
+    {
+      "epoch": 0.64,
+      "objective/train/advantage_avg": 0.000818012107629329,
+      "objective/train/docs_used": 765555,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2233686447143555,
+      "objective/train/original_loss": 1.2233686447143555,
+      "objective/train/theoretical_loss": 3.4449410257862545,
+      "objective/train/tokens_used": 2120888800,
+      "objective/train/value_avg": -0.006542205810546875,
+      "objective/train/value_loss": 0.00015375840303022414,
+      "objective/train/value_max": -5.3048133850097656e-05,
+      "objective/train/value_min": -0.29150390625,
+      "objective/train/value_reward_corr": 0.6431444439728494,
+      "objective/train/value_std": 0.01126861572265625,
+      "objective/train/weight_avg": 1.000891923904419,
+      "objective/train/weighted_lm_loss": 1.2243032455444336,
+      "objective/train/weights_max": 1.3039275407791138,
+      "objective/train/weights_min": 0.6103137731552124,
+      "theoretical_loss": 3.4449410257862545,
+      "tokens_seen": 2100428800
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036711603273952817,
+      "loss": 0.065,
+      "theoretical_loss": 3.4449234213905564,
+      "tokens_seen": 2100559872
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036703578879794573,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4448882168169908,
+      "tokens_seen": 2100822016
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036695554485636334,
+      "loss": 0.0654,
+      "theoretical_loss": 3.444853017865869,
+      "tokens_seen": 2101084160
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036687530091478095,
+      "loss": 0.0646,
+      "theoretical_loss": 3.4448178245355896,
+      "tokens_seen": 2101346304
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003667950569731985,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4447826368245558,
+      "tokens_seen": 2101608448
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003667148130316161,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4447474547311683,
+      "tokens_seen": 2101870592
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036663456909003374,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4447122782538306,
+      "tokens_seen": 2102132736
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003665543251484513,
+      "loss": 0.0656,
+      "theoretical_loss": 3.444677107390946,
+      "tokens_seen": 2102394880
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036647408120686885,
+      "loss": 0.0644,
+      "theoretical_loss": 3.444641942140918,
+      "tokens_seen": 2102657024
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036639383726528647,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4446067825021514,
+      "tokens_seen": 2102919168
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003663135933237041,
+      "loss": 0.0657,
+      "theoretical_loss": 3.444571628473052,
+      "tokens_seen": 2103181312
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036623334938212164,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4445364800520255,
+      "tokens_seen": 2103443456
+    },
+    {
+      "epoch": 0.64,
+      "objective/train/advantage_avg": -0.0001629337202757597,
+      "objective/train/docs_used": 766743,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4270471334457397,
+      "objective/train/original_loss": 1.4270472526550293,
+      "objective/train/theoretical_loss": 3.444501337237478,
+      "objective/train/tokens_used": 2124165600,
+      "objective/train/value_avg": -0.01125335693359375,
+      "objective/train/value_loss": 0.0007498456398025155,
+      "objective/train/value_max": -4.988908767700195e-05,
+      "objective/train/value_min": -0.91162109375,
+      "objective/train/value_reward_corr": 0.6557396160471719,
+      "objective/train/value_std": 0.0236358642578125,
+      "objective/train/weight_avg": 1.000152587890625,
+      "objective/train/weighted_lm_loss": 1.4258575439453125,
+      "objective/train/weights_max": 1.6239049434661865,
+      "objective/train/weights_min": 0.23518386483192444,
+      "theoretical_loss": 3.444501337237478,
+      "tokens_seen": 2103705600
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036615310544053925,
+      "loss": 0.0694,
+      "theoretical_loss": 3.444501337237478,
+      "tokens_seen": 2103705600
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036607286149895686,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4444662000278177,
+      "tokens_seen": 2103967744
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003659926175573745,
+      "loss": 0.0626,
+      "theoretical_loss": 3.4444310684214514,
+      "tokens_seen": 2104229888
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.000365912373615792,
+      "loss": 0.0638,
+      "theoretical_loss": 3.444395942416789,
+      "tokens_seen": 2104492032
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003658321296742096,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4443608220122384,
+      "tokens_seen": 2104754176
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003657518857326272,
+      "loss": 0.0668,
+      "theoretical_loss": 3.44432570720621,
+      "tokens_seen": 2105016320
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036567164179104476,
+      "loss": 0.0622,
+      "theoretical_loss": 3.4442905979971146,
+      "tokens_seen": 2105278464
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003655913978494624,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4442554943833628,
+      "tokens_seen": 2105540608
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036551115390788,
+      "loss": 0.0673,
+      "theoretical_loss": 3.444220396363367,
+      "tokens_seen": 2105802752
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003654309099662976,
+      "loss": 0.0642,
+      "theoretical_loss": 3.444185303935539,
+      "tokens_seen": 2106064896
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003653506660247151,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4441502170982927,
+      "tokens_seen": 2106327040
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003652704220831327,
+      "loss": 0.0689,
+      "theoretical_loss": 3.444115135850041,
+      "tokens_seen": 2106589184
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036519017814155033,
+      "loss": 0.0685,
+      "theoretical_loss": 3.444080060189199,
+      "tokens_seen": 2106851328
+    },
+    {
+      "epoch": 0.64,
+      "objective/train/advantage_avg": 0.0003502106701489538,
+      "objective/train/docs_used": 767939,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3586885929107666,
+      "objective/train/original_loss": 1.3586885929107666,
+      "objective/train/theoretical_loss": 3.444062524453561,
+      "objective/train/tokens_used": 2127442400,
+      "objective/train/value_avg": -0.00652313232421875,
+      "objective/train/value_loss": 0.0001283576712012291,
+      "objective/train/value_max": -3.916025161743164e-05,
+      "objective/train/value_min": -0.291259765625,
+      "objective/train/value_reward_corr": 0.7288310765868844,
+      "objective/train/value_std": 0.01226806640625,
+      "objective/train/weight_avg": 1.000410556793213,
+      "objective/train/weighted_lm_loss": 1.3594164848327637,
+      "objective/train/weights_max": 1.1665211915969849,
+      "objective/train/weights_min": 0.3985501825809479,
+      "theoretical_loss": 3.444062524453561,
+      "tokens_seen": 2106982400
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003651099341999679,
+      "loss": 0.0692,
+      "theoretical_loss": 3.444044990114181,
+      "tokens_seen": 2107113472
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003650296902583855,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4440099256234036,
+      "tokens_seen": 2107375616
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003649494463168031,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4439748667152825,
+      "tokens_seen": 2107637760
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036486920237522067,
+      "loss": 0.0667,
+      "theoretical_loss": 3.443939813388235,
+      "tokens_seen": 2107899904
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036478895843363823,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4439047656406783,
+      "tokens_seen": 2108162048
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036470871449205584,
+      "loss": 0.0691,
+      "theoretical_loss": 3.443869723471031,
+      "tokens_seen": 2108424192
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036462847055047346,
+      "loss": 0.067,
+      "theoretical_loss": 3.443834686877712,
+      "tokens_seen": 2108686336
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.000364548226608891,
+      "loss": 0.0674,
+      "theoretical_loss": 3.443799655859141,
+      "tokens_seen": 2108948480
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036446798266730863,
+      "loss": 0.0673,
+      "theoretical_loss": 3.443764630413738,
+      "tokens_seen": 2109210624
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036438773872572624,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4437296105399238,
+      "tokens_seen": 2109472768
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003643074947841438,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4436945962361203,
+      "tokens_seen": 2109734912
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003642272508425614,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4436595875007487,
+      "tokens_seen": 2109997056
+    },
+    {
+      "epoch": 0.64,
+      "objective/train/advantage_avg": 6.14080490777269e-05,
+      "objective/train/docs_used": 769058,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2044825553894043,
+      "objective/train/original_loss": 1.2044825553894043,
+      "objective/train/theoretical_loss": 3.443624584332233,
+      "objective/train/tokens_used": 2130719200,
+      "objective/train/value_avg": -0.006946563720703125,
+      "objective/train/value_loss": 0.00027350004529580474,
+      "objective/train/value_max": -3.11732292175293e-05,
+      "objective/train/value_min": -0.69775390625,
+      "objective/train/value_reward_corr": 0.5902221741881584,
+      "objective/train/value_std": 0.011749267578125,
+      "objective/train/weight_avg": 1.0001745223999023,
+      "objective/train/weighted_lm_loss": 1.2041305303573608,
+      "objective/train/weights_max": 1.1966127157211304,
+      "objective/train/weights_min": 0.3685459494590759,
+      "theoretical_loss": 3.443624584332233,
+      "tokens_seen": 2110259200
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036414700690097897,
+      "loss": 0.0672,
+      "theoretical_loss": 3.443624584332233,
+      "tokens_seen": 2110259200
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003640667629593966,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4435895867289963,
+      "tokens_seen": 2110521344
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036398651901781414,
+      "loss": 0.0666,
+      "theoretical_loss": 3.443554594689462,
+      "tokens_seen": 2110783488
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036390627507623175,
+      "loss": 0.0638,
+      "theoretical_loss": 3.443519608212055,
+      "tokens_seen": 2111045632
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036382603113464937,
+      "loss": 0.064,
+      "theoretical_loss": 3.4434846272952013,
+      "tokens_seen": 2111307776
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003637457871930669,
+      "loss": 0.067,
+      "theoretical_loss": 3.4434496519373265,
+      "tokens_seen": 2111569920
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036366554325148454,
+      "loss": 0.0684,
+      "theoretical_loss": 3.443414682136857,
+      "tokens_seen": 2111832064
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003635852993099021,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4433797178922205,
+      "tokens_seen": 2112094208
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003635050553683197,
+      "loss": 0.0647,
+      "theoretical_loss": 3.4433447592018447,
+      "tokens_seen": 2112356352
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036342481142673727,
+      "loss": 0.0672,
+      "theoretical_loss": 3.443309806064158,
+      "tokens_seen": 2112618496
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003633445674851549,
+      "loss": 0.0664,
+      "theoretical_loss": 3.44327485847759,
+      "tokens_seen": 2112880640
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003632643235435725,
+      "loss": 0.0669,
+      "theoretical_loss": 3.44323991644057,
+      "tokens_seen": 2113142784
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036318407960199005,
+      "loss": 0.0649,
+      "theoretical_loss": 3.443204979951529,
+      "tokens_seen": 2113404928
+    },
+    {
+      "epoch": 0.64,
+      "objective/train/advantage_avg": 0.0007937829359434545,
+      "objective/train/docs_used": 770311,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2841191291809082,
+      "objective/train/original_loss": 1.2841191291809082,
+      "objective/train/theoretical_loss": 3.4431875137870103,
+      "objective/train/tokens_used": 2133996000,
+      "objective/train/value_avg": -0.0069122314453125,
+      "objective/train/value_loss": 0.00022702872229274362,
+      "objective/train/value_max": -7.659196853637695e-05,
+      "objective/train/value_min": -0.361328125,
+      "objective/train/value_reward_corr": 0.6486080520614381,
+      "objective/train/value_std": 0.0117950439453125,
+      "objective/train/weight_avg": 1.0008962154388428,
+      "objective/train/weighted_lm_loss": 1.2856719493865967,
+      "objective/train/weights_max": 1.2625762224197388,
+      "objective/train/weights_min": 0.3683098554611206,
+      "theoretical_loss": 3.4431875137870103,
+      "tokens_seen": 2113536000
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036310383566040766,
+      "loss": 0.0642,
+      "theoretical_loss": 3.4431700490088977,
+      "tokens_seen": 2113667072
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003630235917188253,
+      "loss": 0.0675,
+      "theoretical_loss": 3.443135123611108,
+      "tokens_seen": 2113929216
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003629433477772428,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4431002037565923,
+      "tokens_seen": 2114191360
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003628631038356604,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4430652894437834,
+      "tokens_seen": 2114453504
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.000362782859894078,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4430303806711158,
+      "tokens_seen": 2114715648
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003627026159524956,
+      "loss": 0.0675,
+      "theoretical_loss": 3.442995477437023,
+      "tokens_seen": 2114977792
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003626223720109132,
+      "loss": 0.066,
+      "theoretical_loss": 3.44296057973994,
+      "tokens_seen": 2115239936
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003625421280693308,
+      "loss": 0.0671,
+      "theoretical_loss": 3.442925687578302,
+      "tokens_seen": 2115502080
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003624618841277484,
+      "loss": 0.0672,
+      "theoretical_loss": 3.442890800950546,
+      "tokens_seen": 2115764224
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003623816401861659,
+      "loss": 0.0637,
+      "theoretical_loss": 3.442855919855109,
+      "tokens_seen": 2116026368
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003623013962445835,
+      "loss": 0.0634,
+      "theoretical_loss": 3.4428210442904277,
+      "tokens_seen": 2116288512
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036222115230300113,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4427861742549406,
+      "tokens_seen": 2116550656
+    },
+    {
+      "epoch": 0.64,
+      "objective/train/advantage_avg": 0.00024320276861544698,
+      "objective/train/docs_used": 771474,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2714860439300537,
+      "objective/train/original_loss": 1.2714862823486328,
+      "objective/train/theoretical_loss": 3.442751309747086,
+      "objective/train/tokens_used": 2137272800,
+      "objective/train/value_avg": -0.009307861328125,
+      "objective/train/value_loss": 0.00032071780879050493,
+      "objective/train/value_max": -4.988908767700195e-05,
+      "objective/train/value_min": -0.431396484375,
+      "objective/train/value_reward_corr": 0.735485788504759,
+      "objective/train/value_std": 0.0192718505859375,
+      "objective/train/weight_avg": 1.00038480758667,
+      "objective/train/weighted_lm_loss": 1.271644115447998,
+      "objective/train/weights_max": 1.3421837091445923,
+      "objective/train/weights_min": 0.3815314471721649,
+      "theoretical_loss": 3.442751309747086,
+      "tokens_seen": 2116812800
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036214090836141874,
+      "loss": 0.0647,
+      "theoretical_loss": 3.442751309747086,
+      "tokens_seen": 2116812800
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003620606644198363,
+      "loss": 0.0663,
+      "theoretical_loss": 3.442716450765304,
+      "tokens_seen": 2117074944
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003619804204782539,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4426815973080345,
+      "tokens_seen": 2117337088
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003619001765366715,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4426467493737176,
+      "tokens_seen": 2117599232
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036181993259508903,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4426119069607948,
+      "tokens_seen": 2117861376
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036173968865350664,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4425770700677085,
+      "tokens_seen": 2118123520
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036165944471192426,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4425422386929005,
+      "tokens_seen": 2118385664
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036157920077034187,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4425074128348148,
+      "tokens_seen": 2118647808
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003614989568287594,
+      "loss": 0.0664,
+      "theoretical_loss": 3.442472592491894,
+      "tokens_seen": 2118909952
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036141871288717704,
+      "loss": 0.0669,
+      "theoretical_loss": 3.442437777662584,
+      "tokens_seen": 2119172096
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036133846894559465,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4424029683453288,
+      "tokens_seen": 2119434240
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003612582250040122,
+      "loss": 0.0683,
+      "theoretical_loss": 3.442368164538575,
+      "tokens_seen": 2119696384
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036117798106242977,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4423333662407676,
+      "tokens_seen": 2119958528
+    },
+    {
+      "epoch": 0.64,
+      "objective/train/advantage_avg": -0.0005769357667304575,
+      "objective/train/docs_used": 772695,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3378933668136597,
+      "objective/train/original_loss": 1.3378934860229492,
+      "objective/train/theoretical_loss": 3.442315969157234,
+      "objective/train/tokens_used": 2140549600,
+      "objective/train/value_avg": -0.00926971435546875,
+      "objective/train/value_loss": 0.0003164195513818413,
+      "objective/train/value_max": -2.110004425048828e-05,
+      "objective/train/value_min": -0.26513671875,
+      "objective/train/value_reward_corr": 0.7086236395211777,
+      "objective/train/value_std": 0.0174713134765625,
+      "objective/train/weight_avg": 0.999565601348877,
+      "objective/train/weighted_lm_loss": 1.3366376161575317,
+      "objective/train/weights_max": 1.3036091327667236,
+      "objective/train/weights_min": 0.368646502494812,
+      "theoretical_loss": 3.442315969157234,
+      "tokens_seen": 2120089600
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003610977371208474,
+      "loss": 0.0673,
+      "theoretical_loss": 3.442298573450355,
+      "tokens_seen": 2120220672
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036101749317926494,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4422637861657837,
+      "tokens_seen": 2120482816
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036093724923768255,
+      "loss": 0.0657,
+      "theoretical_loss": 3.442229004385502,
+      "tokens_seen": 2120744960
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036085700529610017,
+      "loss": 0.0661,
+      "theoretical_loss": 3.44219422810796,
+      "tokens_seen": 2121007104
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003607767613545178,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4421594573316057,
+      "tokens_seen": 2121269248
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036069651741293534,
+      "loss": 0.0669,
+      "theoretical_loss": 3.44212469205489,
+      "tokens_seen": 2121531392
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003606162734713529,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4420899322762635,
+      "tokens_seen": 2121793536
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003605360295297705,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4420551779941775,
+      "tokens_seen": 2122055680
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036045578558818807,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4420204292070844,
+      "tokens_seen": 2122317824
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003603755416466057,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4419856859134357,
+      "tokens_seen": 2122579968
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003602952977050233,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4419509481116854,
+      "tokens_seen": 2122842112
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003602150537634409,
+      "loss": 0.0688,
+      "theoretical_loss": 3.441916215800288,
+      "tokens_seen": 2123104256
+    },
+    {
+      "epoch": 0.64,
+      "objective/train/advantage_avg": -0.00034259771928191185,
+      "objective/train/docs_used": 773937,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2587370872497559,
+      "objective/train/original_loss": 1.2587368488311768,
+      "objective/train/theoretical_loss": 3.441881488977697,
+      "objective/train/tokens_used": 2143826400,
+      "objective/train/value_avg": -0.01271820068359375,
+      "objective/train/value_loss": 0.00030873133800923824,
+      "objective/train/value_max": -4.363059997558594e-05,
+      "objective/train/value_min": -0.39013671875,
+      "objective/train/value_reward_corr": 0.75689359226375,
+      "objective/train/value_std": 0.0196685791015625,
+      "objective/train/weight_avg": 0.9998003840446472,
+      "objective/train/weighted_lm_loss": 1.2579768896102905,
+      "objective/train/weights_max": 1.375349760055542,
+      "objective/train/weights_min": 0.3696674406528473,
+      "theoretical_loss": 3.441881488977697,
+      "tokens_seen": 2123366400
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00036013480982185846,
+      "loss": 0.0635,
+      "theoretical_loss": 3.441881488977697,
+      "tokens_seen": 2123366400
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.000360054565880276,
+      "loss": 0.0633,
+      "theoretical_loss": 3.441846767642368,
+      "tokens_seen": 2123628544
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00035997432193869363,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4418120517927564,
+      "tokens_seen": 2123890688
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003598940779971112,
+      "loss": 0.0686,
+      "theoretical_loss": 3.441777341427319,
+      "tokens_seen": 2124152832
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003598138340555288,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4417426365445127,
+      "tokens_seen": 2124414976
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003597335901139464,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4417079371427945,
+      "tokens_seen": 2124677120
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00035965334617236403,
+      "loss": 0.0665,
+      "theoretical_loss": 3.441673243220624,
+      "tokens_seen": 2124939264
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003595731022307816,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4416385547764583,
+      "tokens_seen": 2125201408
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003594928582891992,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4416038718087583,
+      "tokens_seen": 2125463552
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00035941261434761676,
+      "loss": 0.067,
+      "theoretical_loss": 3.4415691943159836,
+      "tokens_seen": 2125725696
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003593323704060343,
+      "loss": 0.0672,
+      "theoretical_loss": 3.441534522296595,
+      "tokens_seen": 2125987840
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00035925212646445193,
+      "loss": 0.0643,
+      "theoretical_loss": 3.441499855749054,
+      "tokens_seen": 2126249984
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00035917188252286954,
+      "loss": 0.0648,
+      "theoretical_loss": 3.441465194671822,
+      "tokens_seen": 2126512128
+    },
+    {
+      "epoch": 0.64,
+      "objective/train/advantage_avg": 0.0008241595351137221,
+      "objective/train/docs_used": 774962,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2446796894073486,
+      "objective/train/original_loss": 1.244679570198059,
+      "objective/train/theoretical_loss": 3.441447866184092,
+      "objective/train/tokens_used": 2147103200,
+      "objective/train/value_avg": -0.0080718994140625,
+      "objective/train/value_loss": 0.0002672381524462253,
+      "objective/train/value_max": -6.556510925292969e-05,
+      "objective/train/value_min": -0.96728515625,
+      "objective/train/value_reward_corr": 0.7998461563914081,
+      "objective/train/value_std": 0.0204925537109375,
+      "objective/train/weight_avg": 1.0009392499923706,
+      "objective/train/weighted_lm_loss": 1.2461644411087036,
+      "objective/train/weights_max": 1.2684307098388672,
+      "objective/train/weights_min": 0.3747805058956146,
+      "theoretical_loss": 3.441447866184092,
+      "tokens_seen": 2126643200
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00035909163858128715,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4414305390633624,
+      "tokens_seen": 2126774272
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003590113946397047,
+      "loss": 0.0648,
+      "theoretical_loss": 3.4413958889221385,
+      "tokens_seen": 2127036416
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003589311506981223,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4413612442466133,
+      "tokens_seen": 2127298560
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.0003588509067565399,
+      "loss": 0.0668,
+      "theoretical_loss": 3.441326605035252,
+      "tokens_seen": 2127560704
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00035877066281495744,
+      "loss": 0.067,
+      "theoretical_loss": 3.441291971286519,
+      "tokens_seen": 2127822848
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00035869041887337506,
+      "loss": 0.0644,
+      "theoretical_loss": 3.44125734299888,
+      "tokens_seen": 2128084992
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 0.00035861017493179267,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4412227201708028,
+      "tokens_seen": 2128347136
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003585299309902102,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4411881028007527,
+      "tokens_seen": 2128609280
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035844968704862784,
+      "loss": 0.0646,
+      "theoretical_loss": 3.4411534908871984,
+      "tokens_seen": 2128871424
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035836944310704545,
+      "loss": 0.068,
+      "theoretical_loss": 3.4411188844286067,
+      "tokens_seen": 2129133568
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035828919916546306,
+      "loss": 0.0676,
+      "theoretical_loss": 3.441084283423448,
+      "tokens_seen": 2129395712
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035820895522388057,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4410496878701906,
+      "tokens_seen": 2129657856
+    },
+    {
+      "epoch": 0.65,
+      "objective/train/advantage_avg": 0.0012429844355210662,
+      "objective/train/docs_used": 775953,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.301253318786621,
+      "objective/train/original_loss": 1.301253080368042,
+      "objective/train/theoretical_loss": 3.441015097767305,
+      "objective/train/tokens_used": 2150380000,
+      "objective/train/value_avg": -0.0076141357421875,
+      "objective/train/value_loss": 0.00014564950834028423,
+      "objective/train/value_max": -8.285045623779297e-05,
+      "objective/train/value_min": -0.197021484375,
+      "objective/train/value_reward_corr": 0.7439541772547014,
+      "objective/train/value_std": 0.01273345947265625,
+      "objective/train/weight_avg": 1.001312017440796,
+      "objective/train/weighted_lm_loss": 1.3027527332305908,
+      "objective/train/weights_max": 1.1942163705825806,
+      "objective/train/weights_min": 0.3970448672771454,
+      "theoretical_loss": 3.441015097767305,
+      "tokens_seen": 2129920000
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003581287112822982,
+      "loss": 0.065,
+      "theoretical_loss": 3.441015097767305,
+      "tokens_seen": 2129920000
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003580484673407158,
+      "loss": 0.0689,
+      "theoretical_loss": 3.440980513113262,
+      "tokens_seen": 2130182144
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035796822339913335,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4409459339065327,
+      "tokens_seen": 2130444288
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035788797945755096,
+      "loss": 0.0704,
+      "theoretical_loss": 3.4409113601455887,
+      "tokens_seen": 2130706432
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003578077355159686,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4408767918289027,
+      "tokens_seen": 2130968576
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003577274915743862,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4408422289549483,
+      "tokens_seen": 2131230720
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003576472476328037,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4408076715221982,
+      "tokens_seen": 2131492864
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003575670036912213,
+      "loss": 0.0658,
+      "theoretical_loss": 3.440773119529128,
+      "tokens_seen": 2131755008
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003574867597496389,
+      "loss": 0.0676,
+      "theoretical_loss": 3.440738572974212,
+      "tokens_seen": 2132017152
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003574065158080565,
+      "loss": 0.0688,
+      "theoretical_loss": 3.440704031855926,
+      "tokens_seen": 2132279296
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003573262718664741,
+      "loss": 0.0694,
+      "theoretical_loss": 3.440669496172746,
+      "tokens_seen": 2132541440
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003572460279248917,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4406349659231488,
+      "tokens_seen": 2132803584
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003571657839833093,
+      "loss": 0.0638,
+      "theoretical_loss": 3.4406004411056124,
+      "tokens_seen": 2133065728
+    },
+    {
+      "epoch": 0.65,
+      "objective/train/advantage_avg": 0.0007250725175254047,
+      "objective/train/docs_used": 777226,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2577089071273804,
+      "objective/train/original_loss": 1.2577087879180908,
+      "objective/train/theoretical_loss": 3.440583180733391,
+      "objective/train/tokens_used": 2153656800,
+      "objective/train/value_avg": -0.0079498291015625,
+      "objective/train/value_loss": 0.00016377944848500192,
+      "objective/train/value_max": -7.033348083496094e-05,
+      "objective/train/value_min": -0.2166748046875,
+      "objective/train/value_reward_corr": 0.6768520456626519,
+      "objective/train/value_std": 0.01261138916015625,
+      "objective/train/weight_avg": 1.0007988214492798,
+      "objective/train/weighted_lm_loss": 1.2580339908599854,
+      "objective/train/weights_max": 1.2066720724105835,
+      "objective/train/weights_min": 0.3683168888092041,
+      "theoretical_loss": 3.440583180733391,
+      "tokens_seen": 2133196800
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003570855400417268,
+      "loss": 0.0642,
+      "theoretical_loss": 3.4405659217186138,
+      "tokens_seen": 2133327872
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035700529610014443,
+      "loss": 0.0641,
+      "theoretical_loss": 3.440531407760633,
+      "tokens_seen": 2133590016
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035692505215856204,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4404968992301477,
+      "tokens_seen": 2133852160
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003568448082169796,
+      "loss": 0.0639,
+      "theoretical_loss": 3.440462396125639,
+      "tokens_seen": 2134114304
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003567645642753972,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4404278984455874,
+      "tokens_seen": 2134376448
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035668432033381483,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4403934061884733,
+      "tokens_seen": 2134638592
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003566040763922324,
+      "loss": 0.068,
+      "theoretical_loss": 3.4403589193527786,
+      "tokens_seen": 2134900736
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035652383245065,
+      "loss": 0.0676,
+      "theoretical_loss": 3.440324437936986,
+      "tokens_seen": 2135162880
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035644358850906756,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4402899619395786,
+      "tokens_seen": 2135425024
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035636334456748517,
+      "loss": 0.063,
+      "theoretical_loss": 3.440255491359039,
+      "tokens_seen": 2135687168
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035628310062590273,
+      "loss": 0.0671,
+      "theoretical_loss": 3.440221026193852,
+      "tokens_seen": 2135949312
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035620285668432034,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4401865664425024,
+      "tokens_seen": 2136211456
+    },
+    {
+      "epoch": 0.65,
+      "objective/train/advantage_avg": 0.0004655689117498696,
+      "objective/train/docs_used": 778288,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3205798864364624,
+      "objective/train/original_loss": 1.3205798864364624,
+      "objective/train/theoretical_loss": 3.440152112103476,
+      "objective/train/tokens_used": 2156933600,
+      "objective/train/value_avg": -0.00875091552734375,
+      "objective/train/value_loss": 0.00023888830037321895,
+      "objective/train/value_max": -5.918741226196289e-05,
+      "objective/train/value_min": -0.34912109375,
+      "objective/train/value_reward_corr": 0.666328584912689,
+      "objective/train/value_std": 0.01355743408203125,
+      "objective/train/weight_avg": 1.0005714893341064,
+      "objective/train/weighted_lm_loss": 1.3204231262207031,
+      "objective/train/weights_max": 1.0932023525238037,
+      "objective/train/weights_min": 0.37003564834594727,
+      "theoretical_loss": 3.440152112103476,
+      "tokens_seen": 2136473600
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035612261274273795,
+      "loss": 0.066,
+      "theoretical_loss": 3.440152112103476,
+      "tokens_seen": 2136473600
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003560423688011555,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4401176631752577,
+      "tokens_seen": 2136735744
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003559621248595731,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4400832196563353,
+      "tokens_seen": 2136997888
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003558818809179907,
+      "loss": 0.0667,
+      "theoretical_loss": 3.440048781545195,
+      "tokens_seen": 2137260032
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003558016369764083,
+      "loss": 0.0691,
+      "theoretical_loss": 3.440014348840325,
+      "tokens_seen": 2137522176
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035572139303482585,
+      "loss": 0.0674,
+      "theoretical_loss": 3.439979921540214,
+      "tokens_seen": 2137784320
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035564114909324347,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4399454996433514,
+      "tokens_seen": 2138046464
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003555609051516611,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4399110831482256,
+      "tokens_seen": 2138308608
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035548066121007864,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4398766720533276,
+      "tokens_seen": 2138570752
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035540041726849625,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4398422663571484,
+      "tokens_seen": 2138832896
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003553201733269138,
+      "loss": 0.0662,
+      "theoretical_loss": 3.439807866058179,
+      "tokens_seen": 2139095040
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003552399293853314,
+      "loss": 0.0697,
+      "theoretical_loss": 3.439773471154912,
+      "tokens_seen": 2139357184
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.000355159685443749,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4397390816458397,
+      "tokens_seen": 2139619328
+    },
+    {
+      "epoch": 0.65,
+      "objective/train/advantage_avg": 0.0012974967248737812,
+      "objective/train/docs_used": 779442,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3170843124389648,
+      "objective/train/original_loss": 1.3170843124389648,
+      "objective/train/theoretical_loss": 3.439721888913656,
+      "objective/train/tokens_used": 2160210400,
+      "objective/train/value_avg": -0.00569915771484375,
+      "objective/train/value_loss": 0.00020559185941237956,
+      "objective/train/value_max": -3.1948089599609375e-05,
+      "objective/train/value_min": -0.98291015625,
+      "objective/train/value_reward_corr": 0.6640440608056907,
+      "objective/train/value_std": 0.011474609375,
+      "objective/train/weight_avg": 1.0013810396194458,
+      "objective/train/weighted_lm_loss": 1.3188480138778687,
+      "objective/train/weights_max": 1.2265723943710327,
+      "objective/train/weights_min": 0.25205379724502563,
+      "theoretical_loss": 3.439721888913656,
+      "tokens_seen": 2139750400
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003550794415021666,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4397046975294554,
+      "tokens_seen": 2139881472
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003549991975605842,
+      "loss": 0.069,
+      "theoretical_loss": 3.4396703188042537,
+      "tokens_seen": 2140143616
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035491895361900176,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4396359454687278,
+      "tokens_seen": 2140405760
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003548387096774194,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4396015775213735,
+      "tokens_seen": 2140667904
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.000354758465735837,
+      "loss": 0.0663,
+      "theoretical_loss": 3.439567214960687,
+      "tokens_seen": 2140930048
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003546782217942545,
+      "loss": 0.0646,
+      "theoretical_loss": 3.439532857785164,
+      "tokens_seen": 2141192192
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003545979778526721,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4394985059933014,
+      "tokens_seen": 2141454336
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003545177339110897,
+      "loss": 0.0668,
+      "theoretical_loss": 3.439464159583597,
+      "tokens_seen": 2141716480
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035443748996950733,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4394298185545487,
+      "tokens_seen": 2141978624
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003543572460279249,
+      "loss": 0.0682,
+      "theoretical_loss": 3.439395482904655,
+      "tokens_seen": 2142240768
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003542770020863425,
+      "loss": 0.0648,
+      "theoretical_loss": 3.439361152632416,
+      "tokens_seen": 2142502912
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003541967581447601,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4393268277363305,
+      "tokens_seen": 2142765056
+    },
+    {
+      "epoch": 0.65,
+      "objective/train/advantage_avg": 0.001232887152582407,
+      "objective/train/docs_used": 780500,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1809592247009277,
+      "objective/train/original_loss": 1.1809592247009277,
+      "objective/train/theoretical_loss": 3.4392925082149004,
+      "objective/train/tokens_used": 2163487200,
+      "objective/train/value_avg": -0.006099700927734375,
+      "objective/train/value_loss": 0.00020563544239848852,
+      "objective/train/value_max": -4.684925079345703e-05,
+      "objective/train/value_min": -0.39013671875,
+      "objective/train/value_reward_corr": 0.6734060620060285,
+      "objective/train/value_std": 0.01251983642578125,
+      "objective/train/weight_avg": 1.0013197660446167,
+      "objective/train/weighted_lm_loss": 1.1824616193771362,
+      "objective/train/weights_max": 1.163423776626587,
+      "objective/train/weights_min": 0.36951375007629395,
+      "theoretical_loss": 3.4392925082149004,
+      "tokens_seen": 2143027200
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003541165142031776,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4392925082149004,
+      "tokens_seen": 2143027200
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035403627026159523,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4392581940666256,
+      "tokens_seen": 2143289344
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035395602632001284,
+      "loss": 0.0647,
+      "theoretical_loss": 3.4392238852900086,
+      "tokens_seen": 2143551488
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035387578237843046,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4391895818835514,
+      "tokens_seen": 2143813632
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.000353795538436848,
+      "loss": 0.0638,
+      "theoretical_loss": 3.439155283845757,
+      "tokens_seen": 2144075776
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035371529449526563,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4391209911751286,
+      "tokens_seen": 2144337920
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035363505055368324,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4390867038701716,
+      "tokens_seen": 2144600064
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003535548066121008,
+      "loss": 0.0669,
+      "theoretical_loss": 3.439052421929389,
+      "tokens_seen": 2144862208
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035347456267051836,
+      "loss": 0.0643,
+      "theoretical_loss": 3.439018145351287,
+      "tokens_seen": 2145124352
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035339431872893597,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4389838741343715,
+      "tokens_seen": 2145386496
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003533140747873536,
+      "loss": 0.066,
+      "theoretical_loss": 3.438949608277149,
+      "tokens_seen": 2145648640
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035323383084577114,
+      "loss": 0.0649,
+      "theoretical_loss": 3.438915347778127,
+      "tokens_seen": 2145910784
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035315358690418875,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4388810926358127,
+      "tokens_seen": 2146172928
+    },
+    {
+      "epoch": 0.65,
+      "objective/train/advantage_avg": 0.0006946653011254966,
+      "objective/train/docs_used": 781712,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.253171682357788,
+      "objective/train/original_loss": 1.253171443939209,
+      "objective/train/theoretical_loss": 3.4388639670729546,
+      "objective/train/tokens_used": 2166764000,
+      "objective/train/value_avg": -0.008758544921875,
+      "objective/train/value_loss": 0.0003154955920763314,
+      "objective/train/value_max": -6.157159805297852e-05,
+      "objective/train/value_min": -0.465087890625,
+      "objective/train/value_reward_corr": 0.6826854809281973,
+      "objective/train/value_std": 0.017791748046875,
+      "objective/train/weight_avg": 1.0008307695388794,
+      "objective/train/weighted_lm_loss": 1.254067063331604,
+      "objective/train/weights_max": 1.5921541452407837,
+      "objective/train/weights_min": 0.3686527907848358,
+      "theoretical_loss": 3.4388639670729546,
+      "tokens_seen": 2146304000
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035307334296260637,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4388468428487142,
+      "tokens_seen": 2146435072
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003529930990210239,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4388125984153413,
+      "tokens_seen": 2146697216
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003529128550794415,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4387783593342025,
+      "tokens_seen": 2146959360
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003528326111378591,
+      "loss": 0.07,
+      "theoretical_loss": 3.438744125603809,
+      "tokens_seen": 2147221504
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035275236719627665,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4387098972226706,
+      "tokens_seen": 2147483648
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035267212325469427,
+      "loss": 0.0623,
+      "theoretical_loss": 3.4386756741892994,
+      "tokens_seen": 2147745792
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003525918793131119,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4386414565022063,
+      "tokens_seen": 2148007936
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003525116353715295,
+      "loss": 0.0668,
+      "theoretical_loss": 3.438607244159905,
+      "tokens_seen": 2148270080
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035243139142994705,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4385730371609076,
+      "tokens_seen": 2148532224
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003523511474883646,
+      "loss": 0.0673,
+      "theoretical_loss": 3.438538835503728,
+      "tokens_seen": 2148794368
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003522709035467822,
+      "loss": 0.0678,
+      "theoretical_loss": 3.438504639186881,
+      "tokens_seen": 2149056512
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003521906596051998,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4384704482088813,
+      "tokens_seen": 2149318656
+    },
+    {
+      "epoch": 0.65,
+      "objective/train/advantage_avg": 0.00044662863365374506,
+      "objective/train/docs_used": 782928,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1585477590560913,
+      "objective/train/original_loss": 1.1585477590560913,
+      "objective/train/theoretical_loss": 3.4384362625682434,
+      "objective/train/tokens_used": 2170040800,
+      "objective/train/value_avg": -0.005054473876953125,
+      "objective/train/value_loss": 9.015284013003111e-05,
+      "objective/train/value_max": -3.11732292175293e-05,
+      "objective/train/value_min": -0.219970703125,
+      "objective/train/value_reward_corr": 0.7191277656844346,
+      "objective/train/value_std": 0.00902557373046875,
+      "objective/train/weight_avg": 1.0004909038543701,
+      "objective/train/weighted_lm_loss": 1.1594600677490234,
+      "objective/train/weights_max": 1.1269409656524658,
+      "objective/train/weights_min": 0.8207743763923645,
+      "theoretical_loss": 3.4384362625682434,
+      "tokens_seen": 2149580800
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003521104156636174,
+      "loss": 0.0636,
+      "theoretical_loss": 3.4384362625682434,
+      "tokens_seen": 2149580800
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.000352030171722035,
+      "loss": 0.0657,
+      "theoretical_loss": 3.438402082263485,
+      "tokens_seen": 2149842944
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003519499277804526,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4383679072931215,
+      "tokens_seen": 2150105088
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003518696838388702,
+      "loss": 0.0655,
+      "theoretical_loss": 3.438333737655671,
+      "tokens_seen": 2150367232
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003517894398972878,
+      "loss": 0.0687,
+      "theoretical_loss": 3.43829957334965,
+      "tokens_seen": 2150629376
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035170919595570535,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4382654143735785,
+      "tokens_seen": 2150891520
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003516289520141229,
+      "loss": 0.0678,
+      "theoretical_loss": 3.438231260725975,
+      "tokens_seen": 2151153664
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003515487080725405,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4381971124053594,
+      "tokens_seen": 2151415808
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035146846413095813,
+      "loss": 0.0682,
+      "theoretical_loss": 3.438162969410251,
+      "tokens_seen": 2151677952
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035138822018937574,
+      "loss": 0.0675,
+      "theoretical_loss": 3.438128831739171,
+      "tokens_seen": 2151940096
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003513079762477933,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4380946993906414,
+      "tokens_seen": 2152202240
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003512277323062109,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4380605723631836,
+      "tokens_seen": 2152464384
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035114748836462847,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4380264506553204,
+      "tokens_seen": 2152726528
+    },
+    {
+      "epoch": 0.65,
+      "objective/train/advantage_avg": 0.000863985565956682,
+      "objective/train/docs_used": 784071,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3599311113357544,
+      "objective/train/original_loss": 1.3599311113357544,
+      "objective/train/theoretical_loss": 3.4380093917957755,
+      "objective/train/tokens_used": 2173317600,
+      "objective/train/value_avg": -0.0035495758056640625,
+      "objective/train/value_loss": 7.935243775136769e-05,
+      "objective/train/value_max": -5.7816505432128906e-05,
+      "objective/train/value_min": -0.1246337890625,
+      "objective/train/value_reward_corr": 0.49532153914812793,
+      "objective/train/value_std": 0.0050201416015625,
+      "objective/train/weight_avg": 1.0008996725082397,
+      "objective/train/weighted_lm_loss": 1.3618212938308716,
+      "objective/train/weights_max": 1.1140156984329224,
+      "objective/train/weights_min": 0.37567365169525146,
+      "theoretical_loss": 3.4380093917957755,
+      "tokens_seen": 2152857600
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035106724442304603,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4379923342655747,
+      "tokens_seen": 2152988672
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035098700048146364,
+      "loss": 0.0671,
+      "theoretical_loss": 3.437958223192471,
+      "tokens_seen": 2153250816
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035090675653988126,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4379241174345325,
+      "tokens_seen": 2153512960
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035082651259829887,
+      "loss": 0.0635,
+      "theoretical_loss": 3.437890016990285,
+      "tokens_seen": 2153775104
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003507462686567164,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4378559218582536,
+      "tokens_seen": 2154037248
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035066602471513404,
+      "loss": 0.065,
+      "theoretical_loss": 3.4378218320369647,
+      "tokens_seen": 2154299392
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003505857807735516,
+      "loss": 0.0653,
+      "theoretical_loss": 3.437787747524945,
+      "tokens_seen": 2154561536
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035050553683196916,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4377536683207217,
+      "tokens_seen": 2154823680
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035042529289038677,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4377195944228225,
+      "tokens_seen": 2155085824
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003503450489488044,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4376855258297763,
+      "tokens_seen": 2155347968
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035026480500722194,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4376514625401113,
+      "tokens_seen": 2155610112
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035018456106563955,
+      "loss": 0.0665,
+      "theoretical_loss": 3.437617404552358,
+      "tokens_seen": 2155872256
+    },
+    {
+      "epoch": 0.65,
+      "objective/train/advantage_avg": 0.0005617592250928283,
+      "objective/train/docs_used": 785270,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3288520574569702,
+      "objective/train/original_loss": 1.3288519382476807,
+      "objective/train/theoretical_loss": 3.4375833518650465,
+      "objective/train/tokens_used": 2176594400,
+      "objective/train/value_avg": -0.007755279541015625,
+      "objective/train/value_loss": 0.0002667588123586029,
+      "objective/train/value_max": -4.07099723815918e-05,
+      "objective/train/value_min": -0.849609375,
+      "objective/train/value_reward_corr": 0.736639202225347,
+      "objective/train/value_std": 0.018310546875,
+      "objective/train/weight_avg": 1.0006906986236572,
+      "objective/train/weighted_lm_loss": 1.3300721645355225,
+      "objective/train/weights_max": 2.143956422805786,
+      "objective/train/weights_min": 0.5176886916160583,
+      "theoretical_loss": 3.4375833518650465,
+      "tokens_seen": 2156134400
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00035010431712405717,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4375833518650465,
+      "tokens_seen": 2156134400
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003500240731824748,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4375493044767076,
+      "tokens_seen": 2156396544
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003499438292408923,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4375152623858725,
+      "tokens_seen": 2156658688
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003498635852993099,
+      "loss": 0.0664,
+      "theoretical_loss": 3.437481225591073,
+      "tokens_seen": 2156920832
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003497833413577275,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4374471940908418,
+      "tokens_seen": 2157182976
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00034970309741614507,
+      "loss": 0.0647,
+      "theoretical_loss": 3.4374131678837125,
+      "tokens_seen": 2157445120
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003496228534745627,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4373791469682184,
+      "tokens_seen": 2157707264
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003495426095329803,
+      "loss": 0.0662,
+      "theoretical_loss": 3.437345131342894,
+      "tokens_seen": 2157969408
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003494623655913979,
+      "loss": 0.0664,
+      "theoretical_loss": 3.437311121006274,
+      "tokens_seen": 2158231552
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003493821216498154,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4372771159568942,
+      "tokens_seen": 2158493696
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.000349301877708233,
+      "loss": 0.0678,
+      "theoretical_loss": 3.43724311619329,
+      "tokens_seen": 2158755840
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00034922163376665063,
+      "loss": 0.066,
+      "theoretical_loss": 3.437209121713999,
+      "tokens_seen": 2159017984
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003491413898250682,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4371751325175586,
+      "tokens_seen": 2159280128
+    },
+    {
+      "epoch": 0.65,
+      "objective/train/advantage_avg": 0.00039554800605401397,
+      "objective/train/docs_used": 786428,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3268866539001465,
+      "objective/train/original_loss": 1.3268866539001465,
+      "objective/train/theoretical_loss": 3.4371581398999496,
+      "objective/train/tokens_used": 2179871200,
+      "objective/train/value_avg": -0.011016845703125,
+      "objective/train/value_loss": 0.000277909217402339,
+      "objective/train/value_max": -4.13060188293457e-05,
+      "objective/train/value_min": -0.7626953125,
+      "objective/train/value_reward_corr": 0.8255372104578527,
+      "objective/train/value_std": 0.0236968994140625,
+      "objective/train/weight_avg": 1.0005244016647339,
+      "objective/train/weighted_lm_loss": 1.3273314237594604,
+      "objective/train/weights_max": 1.3975777626037598,
+      "objective/train/weights_min": 0.40282997488975525,
+      "theoretical_loss": 3.4371581398999496,
+      "tokens_seen": 2159411200
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003490611458834858,
+      "loss": 0.0656,
+      "theoretical_loss": 3.437141148602505,
+      "tokens_seen": 2159542272
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003489809019419034,
+      "loss": 0.0643,
+      "theoretical_loss": 3.437107169967378,
+      "tokens_seen": 2159804416
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00034890065800032103,
+      "loss": 0.0671,
+      "theoretical_loss": 3.437073196610716,
+      "tokens_seen": 2160066560
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003488204140587386,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4370392285310594,
+      "tokens_seen": 2160328704
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00034874017011715615,
+      "loss": 0.0641,
+      "theoretical_loss": 3.437005265726947,
+      "tokens_seen": 2160590848
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00034865992617557376,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4369713081969206,
+      "tokens_seen": 2160852992
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.0003485796822339913,
+      "loss": 0.068,
+      "theoretical_loss": 3.436937355939521,
+      "tokens_seen": 2161115136
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 0.00034849943829240893,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4369034089532904,
+      "tokens_seen": 2161377280
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034841919435082654,
+      "loss": 0.0654,
+      "theoretical_loss": 3.436869467236771,
+      "tokens_seen": 2161639424
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003483389504092441,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4368355307885063,
+      "tokens_seen": 2161901568
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003482587064676617,
+      "loss": 0.0667,
+      "theoretical_loss": 3.436801599607039,
+      "tokens_seen": 2162163712
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034817846252607927,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4367676736909143,
+      "tokens_seen": 2162425856
+    },
+    {
+      "epoch": 0.66,
+      "objective/train/advantage_avg": 0.001019138377159834,
+      "objective/train/docs_used": 787745,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4617555141448975,
+      "objective/train/original_loss": 1.461755633354187,
+      "objective/train/theoretical_loss": 3.4367337530386766,
+      "objective/train/tokens_used": 2183148000,
+      "objective/train/value_avg": -0.00838470458984375,
+      "objective/train/value_loss": 0.00023759658506605774,
+      "objective/train/value_max": -3.4809112548828125e-05,
+      "objective/train/value_min": -0.732421875,
+      "objective/train/value_reward_corr": 0.7778528006169423,
+      "objective/train/value_std": 0.0183563232421875,
+      "objective/train/weight_avg": 1.0011314153671265,
+      "objective/train/weighted_lm_loss": 1.4620612859725952,
+      "objective/train/weights_max": 1.4613994359970093,
+      "objective/train/weights_min": 0.5150108933448792,
+      "theoretical_loss": 3.4367337530386766,
+      "tokens_seen": 2162688000
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003480982185844969,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4367337530386766,
+      "tokens_seen": 2162688000
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034801797464291444,
+      "loss": 0.068,
+      "theoretical_loss": 3.436699837648871,
+      "tokens_seen": 2162950144
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034793773070133206,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4366659275200444,
+      "tokens_seen": 2163212288
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034785748675974967,
+      "loss": 0.0677,
+      "theoretical_loss": 3.436632022650742,
+      "tokens_seen": 2163474432
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003477772428181672,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4365981230395115,
+      "tokens_seen": 2163736576
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034769699887658484,
+      "loss": 0.0652,
+      "theoretical_loss": 3.436564228684901,
+      "tokens_seen": 2163998720
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003476167549350024,
+      "loss": 0.066,
+      "theoretical_loss": 3.436530339585458,
+      "tokens_seen": 2164260864
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034753651099342,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4364964557397317,
+      "tokens_seen": 2164523008
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034745626705183757,
+      "loss": 0.0663,
+      "theoretical_loss": 3.436462577146272,
+      "tokens_seen": 2164785152
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003473760231102552,
+      "loss": 0.066,
+      "theoretical_loss": 3.4364287038036276,
+      "tokens_seen": 2165047296
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003472957791686728,
+      "loss": 0.067,
+      "theoretical_loss": 3.4363948357103506,
+      "tokens_seen": 2165309440
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034721553522709035,
+      "loss": 0.0668,
+      "theoretical_loss": 3.436360972864991,
+      "tokens_seen": 2165571584
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034713529128550796,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4363271152661006,
+      "tokens_seen": 2165833728
+    },
+    {
+      "epoch": 0.66,
+      "objective/train/advantage_avg": 0.0009586879750713706,
+      "objective/train/docs_used": 788845,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3089131116867065,
+      "objective/train/original_loss": 1.308912992477417,
+      "objective/train/theoretical_loss": 3.4363101884336293,
+      "objective/train/tokens_used": 2186424800,
+      "objective/train/value_avg": -0.0082855224609375,
+      "objective/train/value_loss": 0.0004079457721672952,
+      "objective/train/value_max": -1.9252300262451172e-05,
+      "objective/train/value_min": -0.744140625,
+      "objective/train/value_reward_corr": 0.7048424748964199,
+      "objective/train/value_std": 0.0200958251953125,
+      "objective/train/weight_avg": 1.0011377334594727,
+      "objective/train/weighted_lm_loss": 1.3098411560058594,
+      "objective/train/weights_max": 1.4547736644744873,
+      "objective/train/weights_min": 0.3825574517250061,
+      "theoretical_loss": 3.4363101884336293,
+      "tokens_seen": 2165964800
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003470550473439256,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4362932629122325,
+      "tokens_seen": 2166095872
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034697480340234314,
+      "loss": 0.0728,
+      "theoretical_loss": 3.436259415801939,
+      "tokens_seen": 2166358016
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003468945594607607,
+      "loss": 0.0655,
+      "theoretical_loss": 3.436225573933773,
+      "tokens_seen": 2166620160
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003468143155191783,
+      "loss": 0.0668,
+      "theoretical_loss": 3.43619173730629,
+      "tokens_seen": 2166882304
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003467340715775959,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4361579059180425,
+      "tokens_seen": 2167144448
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003466538276360135,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4361240797675876,
+      "tokens_seen": 2167406592
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003465735836944311,
+      "loss": 0.0691,
+      "theoretical_loss": 3.43609025885348,
+      "tokens_seen": 2167668736
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003464933397528487,
+      "loss": 0.0663,
+      "theoretical_loss": 3.436056443174276,
+      "tokens_seen": 2167930880
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003464130958112662,
+      "loss": 0.0685,
+      "theoretical_loss": 3.436022632728533,
+      "tokens_seen": 2168193024
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003463328518696838,
+      "loss": 0.07,
+      "theoretical_loss": 3.4359888275148083,
+      "tokens_seen": 2168455168
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034625260792810143,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4359550275316595,
+      "tokens_seen": 2168717312
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034617236398651904,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4359212327776456,
+      "tokens_seen": 2168979456
+    },
+    {
+      "epoch": 0.66,
+      "objective/train/advantage_avg": -0.00022465267102234066,
+      "objective/train/docs_used": 790125,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5220752954483032,
+      "objective/train/original_loss": 1.5220751762390137,
+      "objective/train/theoretical_loss": 3.4358874432513256,
+      "objective/train/tokens_used": 2189701600,
+      "objective/train/value_avg": -0.01103973388671875,
+      "objective/train/value_loss": 0.0004339893057476729,
+      "objective/train/value_max": -4.649162292480469e-05,
+      "objective/train/value_min": -0.76220703125,
+      "objective/train/value_reward_corr": 0.8229982443682127,
+      "objective/train/value_std": 0.0262908935546875,
+      "objective/train/weight_avg": 0.9999734163284302,
+      "objective/train/weighted_lm_loss": 1.522541880607605,
+      "objective/train/weights_max": 1.6124958992004395,
+      "objective/train/weights_min": 0.39407527446746826,
+      "theoretical_loss": 3.4358874432513256,
+      "tokens_seen": 2169241600
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003460921200449366,
+      "loss": 0.067,
+      "theoretical_loss": 3.4358874432513256,
+      "tokens_seen": 2169241600
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003460118761033542,
+      "loss": 0.0658,
+      "theoretical_loss": 3.435853658951259,
+      "tokens_seen": 2169503744
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034593163216177183,
+      "loss": 0.064,
+      "theoretical_loss": 3.4358198798760067,
+      "tokens_seen": 2169765888
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034585138822018933,
+      "loss": 0.066,
+      "theoretical_loss": 3.4357861060241293,
+      "tokens_seen": 2170028032
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034577114427860695,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4357523373941876,
+      "tokens_seen": 2170290176
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034569090033702456,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4357185739847447,
+      "tokens_seen": 2170552320
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034561065639544217,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4356848157943625,
+      "tokens_seen": 2170814464
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034553041245385973,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4356510628216044,
+      "tokens_seen": 2171076608
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034545016851227734,
+      "loss": 0.0696,
+      "theoretical_loss": 3.435617315065034,
+      "tokens_seen": 2171338752
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034536992457069495,
+      "loss": 0.0678,
+      "theoretical_loss": 3.435583572523216,
+      "tokens_seen": 2171600896
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003452896806291125,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4355498351947142,
+      "tokens_seen": 2171863040
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034520943668753007,
+      "loss": 0.0668,
+      "theoretical_loss": 3.435516103078095,
+      "tokens_seen": 2172125184
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003451291927459477,
+      "loss": 0.0688,
+      "theoretical_loss": 3.435482376171924,
+      "tokens_seen": 2172387328
+    },
+    {
+      "epoch": 0.66,
+      "objective/train/advantage_avg": 4.255705061950721e-05,
+      "objective/train/docs_used": 791273,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3144559860229492,
+      "objective/train/original_loss": 1.3144559860229492,
+      "objective/train/theoretical_loss": 3.4354655146723085,
+      "objective/train/tokens_used": 2192978400,
+      "objective/train/value_avg": -0.007049560546875,
+      "objective/train/value_loss": 0.00017210084479302168,
+      "objective/train/value_max": -2.777576446533203e-05,
+      "objective/train/value_min": -0.5546875,
+      "objective/train/value_reward_corr": 0.7298873457361592,
+      "objective/train/value_std": 0.01337432861328125,
+      "objective/train/weight_avg": 1.0001221895217896,
+      "objective/train/weighted_lm_loss": 1.313899278640747,
+      "objective/train/weights_max": 1.1279730796813965,
+      "objective/train/weights_min": 0.36996087431907654,
+      "theoretical_loss": 3.4354655146723085,
+      "tokens_seen": 2172518400
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003450489488043653,
+      "loss": 0.0654,
+      "theoretical_loss": 3.435448654474768,
+      "tokens_seen": 2172649472
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034496870486278285,
+      "loss": 0.0674,
+      "theoretical_loss": 3.435414937985194,
+      "tokens_seen": 2172911616
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034488846092120047,
+      "loss": 0.0688,
+      "theoretical_loss": 3.43538122670177,
+      "tokens_seen": 2173173760
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003448082169796181,
+      "loss": 0.0691,
+      "theoretical_loss": 3.435347520623063,
+      "tokens_seen": 2173435904
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034472797303803564,
+      "loss": 0.0677,
+      "theoretical_loss": 3.435313819747644,
+      "tokens_seen": 2173698048
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003446477290964532,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4352801240740805,
+      "tokens_seen": 2173960192
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003445674851548708,
+      "loss": 0.0643,
+      "theoretical_loss": 3.435246433600943,
+      "tokens_seen": 2174222336
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034448724121328837,
+      "loss": 0.0645,
+      "theoretical_loss": 3.4352127483268022,
+      "tokens_seen": 2174484480
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.000344406997271706,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4351790682502297,
+      "tokens_seen": 2174746624
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003443267533301236,
+      "loss": 0.0675,
+      "theoretical_loss": 3.435145393369796,
+      "tokens_seen": 2175008768
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003442465093885412,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4351117236840745,
+      "tokens_seen": 2175270912
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034416626544695876,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4350780591916372,
+      "tokens_seen": 2175533056
+    },
+    {
+      "epoch": 0.66,
+      "objective/train/advantage_avg": 0.0008052947814576328,
+      "objective/train/docs_used": 792571,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2874644994735718,
+      "objective/train/original_loss": 1.2874643802642822,
+      "objective/train/theoretical_loss": 3.4350443998910576,
+      "objective/train/tokens_used": 2196255200,
+      "objective/train/value_avg": -0.009521484375,
+      "objective/train/value_loss": 0.0003565671213436872,
+      "objective/train/value_max": -6.812810897827148e-05,
+      "objective/train/value_min": -0.54638671875,
+      "objective/train/value_reward_corr": 0.6494977411090334,
+      "objective/train/value_std": 0.018707275390625,
+      "objective/train/weight_avg": 1.0009691715240479,
+      "objective/train/weighted_lm_loss": 1.288053035736084,
+      "objective/train/weights_max": 1.560215950012207,
+      "objective/train/weights_min": 0.3739407956600189,
+      "theoretical_loss": 3.4350443998910576,
+      "tokens_seen": 2175795200
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003440860215053764,
+      "loss": 0.065,
+      "theoretical_loss": 3.4350443998910576,
+      "tokens_seen": 2175795200
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034400577756379393,
+      "loss": 0.0674,
+      "theoretical_loss": 3.435010745780909,
+      "tokens_seen": 2176057344
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003439255336222115,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4349770968597677,
+      "tokens_seen": 2176319488
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003438452896806291,
+      "loss": 0.065,
+      "theoretical_loss": 3.4349434531262073,
+      "tokens_seen": 2176581632
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003437650457390467,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4349098145788033,
+      "tokens_seen": 2176843776
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034368480179746433,
+      "loss": 0.0673,
+      "theoretical_loss": 3.434876181216133,
+      "tokens_seen": 2177105920
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003436045578558819,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4348425530367717,
+      "tokens_seen": 2177368064
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003435243139142995,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4348089300392974,
+      "tokens_seen": 2177630208
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034344406997271706,
+      "loss": 0.066,
+      "theoretical_loss": 3.434775312222288,
+      "tokens_seen": 2177892352
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003433638260311346,
+      "loss": 0.0676,
+      "theoretical_loss": 3.434741699584322,
+      "tokens_seen": 2178154496
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034328358208955223,
+      "loss": 0.0677,
+      "theoretical_loss": 3.434708092123978,
+      "tokens_seen": 2178416640
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034320333814796984,
+      "loss": 0.0674,
+      "theoretical_loss": 3.434674489839836,
+      "tokens_seen": 2178678784
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034312309420638746,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4346408927304757,
+      "tokens_seen": 2178940928
+    },
+    {
+      "epoch": 0.66,
+      "objective/train/advantage_avg": -0.001850332017056644,
+      "objective/train/docs_used": 793725,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1691232919692993,
+      "objective/train/original_loss": 1.1691234111785889,
+      "objective/train/theoretical_loss": 3.434624096115895,
+      "objective/train/tokens_used": 2199532000,
+      "objective/train/value_avg": -0.00922393798828125,
+      "objective/train/value_loss": 0.00025595849729143083,
+      "objective/train/value_max": -7.426738739013672e-05,
+      "objective/train/value_min": -0.34521484375,
+      "objective/train/value_reward_corr": 0.850702016525787,
+      "objective/train/value_std": 0.0190887451171875,
+      "objective/train/weight_avg": 0.9982719421386719,
+      "objective/train/weighted_lm_loss": 1.1677327156066895,
+      "objective/train/weights_max": 1.2134528160095215,
+      "objective/train/weights_min": 0.41253384947776794,
+      "theoretical_loss": 3.434624096115895,
+      "tokens_seen": 2179072000
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.000343042850264805,
+      "loss": 0.067,
+      "theoretical_loss": 3.434607300794478,
+      "tokens_seen": 2179203072
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034296260632322263,
+      "loss": 0.0645,
+      "theoretical_loss": 3.434573714030424,
+      "tokens_seen": 2179465216
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003428823623816402,
+      "loss": 0.0648,
+      "theoretical_loss": 3.4345401324368954,
+      "tokens_seen": 2179727360
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034280211844005774,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4345065560124746,
+      "tokens_seen": 2179989504
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034272187449847536,
+      "loss": 0.0658,
+      "theoretical_loss": 3.434472984755745,
+      "tokens_seen": 2180251648
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034264163055689297,
+      "loss": 0.0674,
+      "theoretical_loss": 3.434439418665289,
+      "tokens_seen": 2180513792
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034256138661531053,
+      "loss": 0.064,
+      "theoretical_loss": 3.434405857739691,
+      "tokens_seen": 2180775936
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034248114267372814,
+      "loss": 0.067,
+      "theoretical_loss": 3.434372301977536,
+      "tokens_seen": 2181038080
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034240089873214575,
+      "loss": 0.067,
+      "theoretical_loss": 3.4343387513774095,
+      "tokens_seen": 2181300224
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034232065479056337,
+      "loss": 0.0696,
+      "theoretical_loss": 3.434305205937896,
+      "tokens_seen": 2181562368
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034224041084898087,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4342716656575822,
+      "tokens_seen": 2181824512
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003421601669073985,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4342381305350553,
+      "tokens_seen": 2182086656
+    },
+    {
+      "epoch": 0.66,
+      "objective/train/advantage_avg": 0.0005700654583051801,
+      "objective/train/docs_used": 794520,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4715244770050049,
+      "objective/train/original_loss": 1.4715242385864258,
+      "objective/train/theoretical_loss": 3.434204600568902,
+      "objective/train/tokens_used": 2202808800,
+      "objective/train/value_avg": -0.006561279296875,
+      "objective/train/value_loss": 0.00040286124567501247,
+      "objective/train/value_max": -3.3974647521972656e-05,
+      "objective/train/value_min": -0.49853515625,
+      "objective/train/value_reward_corr": 0.6038308390405315,
+      "objective/train/value_std": 0.0157318115234375,
+      "objective/train/weight_avg": 1.000749945640564,
+      "objective/train/weighted_lm_loss": 1.47190523147583,
+      "objective/train/weights_max": 1.613708734512329,
+      "objective/train/weights_min": 0.39268267154693604,
+      "theoretical_loss": 3.434204600568902,
+      "tokens_seen": 2182348800
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003420799229658161,
+      "loss": 0.0697,
+      "theoretical_loss": 3.434204600568902,
+      "tokens_seen": 2182348800
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034199967902423365,
+      "loss": 0.0699,
+      "theoretical_loss": 3.434171075757711,
+      "tokens_seen": 2182610944
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034191943508265127,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4341375561000698,
+      "tokens_seen": 2182873088
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003418391911410689,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4341040415945683,
+      "tokens_seen": 2183135232
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003417589471994865,
+      "loss": 0.0672,
+      "theoretical_loss": 3.434070532239796,
+      "tokens_seen": 2183397376
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.000341678703257904,
+      "loss": 0.066,
+      "theoretical_loss": 3.4340370280343424,
+      "tokens_seen": 2183659520
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003415984593163216,
+      "loss": 0.0651,
+      "theoretical_loss": 3.434003528976799,
+      "tokens_seen": 2183921664
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003415182153747392,
+      "loss": 0.0663,
+      "theoretical_loss": 3.433970035065756,
+      "tokens_seen": 2184183808
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003414379714331568,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4339365462998064,
+      "tokens_seen": 2184445952
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003413577274915744,
+      "loss": 0.065,
+      "theoretical_loss": 3.4339030626775413,
+      "tokens_seen": 2184708096
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.000341277483549992,
+      "loss": 0.0653,
+      "theoretical_loss": 3.433869584197555,
+      "tokens_seen": 2184970240
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003411972396084096,
+      "loss": 0.0665,
+      "theoretical_loss": 3.43383611085844,
+      "tokens_seen": 2185232384
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003411169956668271,
+      "loss": 0.0661,
+      "theoretical_loss": 3.43380264265879,
+      "tokens_seen": 2185494528
+    },
+    {
+      "epoch": 0.66,
+      "objective/train/advantage_avg": 0.0009123857016675174,
+      "objective/train/docs_used": 795634,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2502954006195068,
+      "objective/train/original_loss": 1.2502954006195068,
+      "objective/train/theoretical_loss": 3.4337859104858257,
+      "objective/train/tokens_used": 2206085600,
+      "objective/train/value_avg": -0.006984710693359375,
+      "objective/train/value_loss": 8.377248741453514e-05,
+      "objective/train/value_max": -4.649162292480469e-05,
+      "objective/train/value_min": -0.2447509765625,
+      "objective/train/value_reward_corr": 0.7441979043018272,
+      "objective/train/value_std": 0.01093292236328125,
+      "objective/train/weight_avg": 1.0009539127349854,
+      "objective/train/weighted_lm_loss": 1.2514960765838623,
+      "objective/train/weights_max": 1.122109055519104,
+      "objective/train/weights_min": 0.8294188380241394,
+      "theoretical_loss": 3.4337859104858257,
+      "tokens_seen": 2185625600
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034103675172524473,
+      "loss": 0.0652,
+      "theoretical_loss": 3.433769179597201,
+      "tokens_seen": 2185756672
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034095650778366235,
+      "loss": 0.0669,
+      "theoretical_loss": 3.433735721672267,
+      "tokens_seen": 2186018816
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003408762638420799,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4337022688825836,
+      "tokens_seen": 2186280960
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003407960199004975,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4336688212267474,
+      "tokens_seen": 2186543104
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034071577595891513,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4336353787033556,
+      "tokens_seen": 2186805248
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034063553201733274,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4336019413110046,
+      "tokens_seen": 2187067392
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003405552880757503,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4335685090482926,
+      "tokens_seen": 2187329536
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034047504413416786,
+      "loss": 0.0702,
+      "theoretical_loss": 3.4335350819138184,
+      "tokens_seen": 2187591680
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034039480019258547,
+      "loss": 0.0663,
+      "theoretical_loss": 3.433501659906181,
+      "tokens_seen": 2187853824
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034031455625100303,
+      "loss": 0.0655,
+      "theoretical_loss": 3.433468243023979,
+      "tokens_seen": 2188115968
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034023431230942064,
+      "loss": 0.0658,
+      "theoretical_loss": 3.433434831265814,
+      "tokens_seen": 2188378112
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00034015406836783826,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4334014246302855,
+      "tokens_seen": 2188640256
+    },
+    {
+      "epoch": 0.66,
+      "objective/train/advantage_avg": 0.0003115667204838246,
+      "objective/train/docs_used": 796788,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.321669340133667,
+      "objective/train/original_loss": 1.321669340133667,
+      "objective/train/theoretical_loss": 3.4333680231159946,
+      "objective/train/tokens_used": 2209362400,
+      "objective/train/value_avg": -0.0115966796875,
+      "objective/train/value_loss": 0.0005159495631232858,
+      "objective/train/value_max": -4.947185516357422e-05,
+      "objective/train/value_min": -0.75146484375,
+      "objective/train/value_reward_corr": 0.7070658662793439,
+      "objective/train/value_std": 0.0208740234375,
+      "objective/train/weight_avg": 1.0005357265472412,
+      "objective/train/weighted_lm_loss": 1.3218821287155151,
+      "objective/train/weights_max": 1.463372826576233,
+      "objective/train/weights_min": 0.370111882686615,
+      "theoretical_loss": 3.4333680231159946,
+      "tokens_seen": 2188902400
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003400738244262558,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4333680231159946,
+      "tokens_seen": 2188902400
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003399935804846734,
+      "loss": 0.066,
+      "theoretical_loss": 3.433334626721544,
+      "tokens_seen": 2189164544
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.000339913336543091,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4333012354455352,
+      "tokens_seen": 2189426688
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003398330926015086,
+      "loss": 0.067,
+      "theoretical_loss": 3.433267849286571,
+      "tokens_seen": 2189688832
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00033975284865992616,
+      "loss": 0.072,
+      "theoretical_loss": 3.4332344682432554,
+      "tokens_seen": 2189950976
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00033967260471834377,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4332010923141913,
+      "tokens_seen": 2190213120
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003395923607767614,
+      "loss": 0.0646,
+      "theoretical_loss": 3.4331677214979845,
+      "tokens_seen": 2190475264
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00033951211683517894,
+      "loss": 0.069,
+      "theoretical_loss": 3.4331343557932392,
+      "tokens_seen": 2190737408
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00033943187289359655,
+      "loss": 0.0655,
+      "theoretical_loss": 3.433100995198561,
+      "tokens_seen": 2190999552
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00033935162895201417,
+      "loss": 0.0672,
+      "theoretical_loss": 3.433067639712556,
+      "tokens_seen": 2191261696
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003392713850104317,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4330342893338313,
+      "tokens_seen": 2191523840
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003391911410688493,
+      "loss": 0.0637,
+      "theoretical_loss": 3.4330009440609937,
+      "tokens_seen": 2191785984
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003391108971272669,
+      "loss": 0.066,
+      "theoretical_loss": 3.432967603892651,
+      "tokens_seen": 2192048128
+    },
+    {
+      "epoch": 0.66,
+      "objective/train/advantage_avg": 0.00028000472229905427,
+      "objective/train/docs_used": 798413,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3710542917251587,
+      "objective/train/original_loss": 1.3710544109344482,
+      "objective/train/theoretical_loss": 3.4329509357222303,
+      "objective/train/tokens_used": 2212639200,
+      "objective/train/value_avg": -0.00600433349609375,
+      "objective/train/value_loss": 0.0005311193526722491,
+      "objective/train/value_max": -5.02467155456543e-05,
+      "objective/train/value_min": -0.87890625,
+      "objective/train/value_reward_corr": 0.6472117724278922,
+      "objective/train/value_std": 0.01491546630859375,
+      "objective/train/weight_avg": 1.0005016326904297,
+      "objective/train/weighted_lm_loss": 1.3713548183441162,
+      "objective/train/weights_max": 2.0175466537475586,
+      "objective/train/weights_min": 0.3728155791759491,
+      "theoretical_loss": 3.4329509357222303,
+      "tokens_seen": 2192179200
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003390306531856845,
+      "loss": 0.0695,
+      "theoretical_loss": 3.432934268827412,
+      "tokens_seen": 2192310272
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00033895040924410207,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4329009388638845,
+      "tokens_seen": 2192572416
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003388701653025197,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4328676140006786,
+      "tokens_seen": 2192834560
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003387899213609373,
+      "loss": 0.068,
+      "theoretical_loss": 3.4328342942364043,
+      "tokens_seen": 2193096704
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00033870967741935485,
+      "loss": 0.0648,
+      "theoretical_loss": 3.432800979569672,
+      "tokens_seen": 2193358848
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003386294334777724,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4327676699990928,
+      "tokens_seen": 2193620992
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00033854918953619,
+      "loss": 0.0674,
+      "theoretical_loss": 3.432734365523278,
+      "tokens_seen": 2193883136
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.00033846894559460763,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4327010661408397,
+      "tokens_seen": 2194145280
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 0.0003383887016530252,
+      "loss": 0.067,
+      "theoretical_loss": 3.432667771850391,
+      "tokens_seen": 2194407424
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003383084577114428,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4326344826505446,
+      "tokens_seen": 2194669568
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003382282137698604,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4326011985399147,
+      "tokens_seen": 2194931712
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003381479698282779,
+      "loss": 0.068,
+      "theoretical_loss": 3.4325679195171155,
+      "tokens_seen": 2195193856
+    },
+    {
+      "epoch": 0.67,
+      "objective/train/advantage_avg": -0.00021135738643351942,
+      "objective/train/docs_used": 799213,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.386327862739563,
+      "objective/train/original_loss": 1.3863276243209839,
+      "objective/train/theoretical_loss": 3.4325346455807613,
+      "objective/train/tokens_used": 2215916000,
+      "objective/train/value_avg": -0.006687164306640625,
+      "objective/train/value_loss": 0.00016143263201229274,
+      "objective/train/value_max": -3.3736228942871094e-05,
+      "objective/train/value_min": -0.63623046875,
+      "objective/train/value_reward_corr": 0.798931972290373,
+      "objective/train/value_std": 0.0166168212890625,
+      "objective/train/weight_avg": 0.9998688697814941,
+      "objective/train/weighted_lm_loss": 1.3864178657531738,
+      "objective/train/weights_max": 1.5625030994415283,
+      "objective/train/weights_min": 0.6712589859962463,
+      "theoretical_loss": 3.4325346455807613,
+      "tokens_seen": 2195456000
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033806772588669553,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4325346455807613,
+      "tokens_seen": 2195456000
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033798748194511315,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4325013767294683,
+      "tokens_seen": 2195718144
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033790723800353076,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4324681129618515,
+      "tokens_seen": 2195980288
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003378269940619483,
+      "loss": 0.067,
+      "theoretical_loss": 3.4324348542765284,
+      "tokens_seen": 2196242432
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033774675012036593,
+      "loss": 0.07,
+      "theoretical_loss": 3.432401600672115,
+      "tokens_seen": 2196504576
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033766650617878354,
+      "loss": 0.067,
+      "theoretical_loss": 3.43236835214723,
+      "tokens_seen": 2196766720
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003375862622372011,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4323351087004905,
+      "tokens_seen": 2197028864
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033750601829561866,
+      "loss": 0.0646,
+      "theoretical_loss": 3.4323018703305155,
+      "tokens_seen": 2197291008
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033742577435403627,
+      "loss": 0.0655,
+      "theoretical_loss": 3.432268637035924,
+      "tokens_seen": 2197553152
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003373455304124539,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4322354088153357,
+      "tokens_seen": 2197815296
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033726528647087144,
+      "loss": 0.068,
+      "theoretical_loss": 3.432202185667371,
+      "tokens_seen": 2198077440
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033718504252928906,
+      "loss": 0.0666,
+      "theoretical_loss": 3.432168967590651,
+      "tokens_seen": 2198339584
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033710479858770667,
+      "loss": 0.0673,
+      "theoretical_loss": 3.432135754583796,
+      "tokens_seen": 2198601728
+    },
+    {
+      "epoch": 0.67,
+      "objective/train/advantage_avg": 0.001131325727328658,
+      "objective/train/docs_used": 800708,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2420262098312378,
+      "objective/train/original_loss": 1.2420260906219482,
+      "objective/train/theoretical_loss": 3.432119149981138,
+      "objective/train/tokens_used": 2219192800,
+      "objective/train/value_avg": -0.01074981689453125,
+      "objective/train/value_loss": 0.0003463710017967969,
+      "objective/train/value_max": -6.300210952758789e-05,
+      "objective/train/value_min": -0.396728515625,
+      "objective/train/value_reward_corr": 0.6319513812312934,
+      "objective/train/value_std": 0.01788330078125,
+      "objective/train/weight_avg": 1.001285433769226,
+      "objective/train/weighted_lm_loss": 1.242701530456543,
+      "objective/train/weights_max": 1.4869521856307983,
+      "objective/train/weights_min": 0.3684363067150116,
+      "theoretical_loss": 3.432119149981138,
+      "tokens_seen": 2198732800
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003370245546461242,
+      "loss": 0.0672,
+      "theoretical_loss": 3.432102546645429,
+      "tokens_seen": 2198863872
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003369443107045418,
+      "loss": 0.0678,
+      "theoretical_loss": 3.432069343774172,
+      "tokens_seen": 2199126016
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003368640667629594,
+      "loss": 0.0678,
+      "theoretical_loss": 3.432036145968648,
+      "tokens_seen": 2199388160
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.000336783822821377,
+      "loss": 0.0683,
+      "theoretical_loss": 3.43200295322748,
+      "tokens_seen": 2199650304
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033670357887979457,
+      "loss": 0.0648,
+      "theoretical_loss": 3.431969765549293,
+      "tokens_seen": 2199912448
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003366233349382122,
+      "loss": 0.0661,
+      "theoretical_loss": 3.43193658293271,
+      "tokens_seen": 2200174592
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003365430909966298,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4319034053763575,
+      "tokens_seen": 2200436736
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033646284705504735,
+      "loss": 0.0656,
+      "theoretical_loss": 3.431870232878861,
+      "tokens_seen": 2200698880
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003363826031134649,
+      "loss": 0.0672,
+      "theoretical_loss": 3.431837065438846,
+      "tokens_seen": 2200961024
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003363023591718825,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4318039030549397,
+      "tokens_seen": 2201223168
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003362221152303001,
+      "loss": 0.0663,
+      "theoretical_loss": 3.431770745725769,
+      "tokens_seen": 2201485312
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003361418712887177,
+      "loss": 0.0643,
+      "theoretical_loss": 3.431737593449962,
+      "tokens_seen": 2201747456
+    },
+    {
+      "epoch": 0.67,
+      "objective/train/advantage_avg": 0.00045791416778229177,
+      "objective/train/docs_used": 801578,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3965429067611694,
+      "objective/train/original_loss": 1.396543025970459,
+      "objective/train/theoretical_loss": 3.4317044462261466,
+      "objective/train/tokens_used": 2222469600,
+      "objective/train/value_avg": -0.00762939453125,
+      "objective/train/value_loss": 0.00016113599122036248,
+      "objective/train/value_max": -5.608797073364258e-05,
+      "objective/train/value_min": -0.5087890625,
+      "objective/train/value_reward_corr": 0.7312199023293408,
+      "objective/train/value_std": 0.01407623291015625,
+      "objective/train/weight_avg": 1.0005367994308472,
+      "objective/train/weighted_lm_loss": 1.3970656394958496,
+      "objective/train/weights_max": 1.3251081705093384,
+      "objective/train/weights_min": 0.5159518122673035,
+      "theoretical_loss": 3.4317044462261466,
+      "tokens_seen": 2202009600
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003360616273471353,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4317044462261466,
+      "tokens_seen": 2202009600
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003359813834055529,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4316713040529523,
+      "tokens_seen": 2202271744
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003359011394639705,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4316381669290084,
+      "tokens_seen": 2202533888
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003358208955223881,
+      "loss": 0.0655,
+      "theoretical_loss": 3.431605034852944,
+      "tokens_seen": 2202796032
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033574065158080565,
+      "loss": 0.0701,
+      "theoretical_loss": 3.43157190782339,
+      "tokens_seen": 2203058176
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003356604076392232,
+      "loss": 0.065,
+      "theoretical_loss": 3.431538785838978,
+      "tokens_seen": 2203320320
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003355801636976408,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4315056688983385,
+      "tokens_seen": 2203582464
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033549991975605843,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4314725570001046,
+      "tokens_seen": 2203844608
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033541967581447604,
+      "loss": 0.0661,
+      "theoretical_loss": 3.431439450142908,
+      "tokens_seen": 2204106752
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003353394318728936,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4314063483253823,
+      "tokens_seen": 2204368896
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003352591879313112,
+      "loss": 0.0656,
+      "theoretical_loss": 3.431373251546161,
+      "tokens_seen": 2204631040
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003351789439897288,
+      "loss": 0.0691,
+      "theoretical_loss": 3.431340159803878,
+      "tokens_seen": 2204893184
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033509870004814633,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4313070730971686,
+      "tokens_seen": 2205155328
+    },
+    {
+      "epoch": 0.67,
+      "objective/train/advantage_avg": 0.00044071170850656927,
+      "objective/train/docs_used": 802757,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3551077842712402,
+      "objective/train/original_loss": 1.3551077842712402,
+      "objective/train/theoretical_loss": 3.4312905316317277,
+      "objective/train/tokens_used": 2225746400,
+      "objective/train/value_avg": -0.006072998046875,
+      "objective/train/value_loss": 0.00028490403201431036,
+      "objective/train/value_max": -2.4497509002685547e-05,
+      "objective/train/value_min": -0.94873046875,
+      "objective/train/value_reward_corr": 0.7518883224176258,
+      "objective/train/value_std": 0.01715087890625,
+      "objective/train/weight_avg": 1.000563144683838,
+      "objective/train/weighted_lm_loss": 1.3555856943130493,
+      "objective/train/weights_max": 1.4964208602905273,
+      "objective/train/weights_min": 0.3752783536911011,
+      "theoretical_loss": 3.4312905316317277,
+      "tokens_seen": 2205286400
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033501845610656395,
+      "loss": 0.069,
+      "theoretical_loss": 3.431273991424668,
+      "tokens_seen": 2205417472
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033493821216498156,
+      "loss": 0.068,
+      "theoretical_loss": 3.431240914785012,
+      "tokens_seen": 2205679616
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033485796822339917,
+      "loss": 0.0654,
+      "theoretical_loss": 3.431207843176836,
+      "tokens_seen": 2205941760
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033477772428181673,
+      "loss": 0.0678,
+      "theoretical_loss": 3.431174776598778,
+      "tokens_seen": 2206203904
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033469748034023434,
+      "loss": 0.0676,
+      "theoretical_loss": 3.431141715049475,
+      "tokens_seen": 2206466048
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033461723639865195,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4311086585275645,
+      "tokens_seen": 2206728192
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033453699245706946,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4310756070316857,
+      "tokens_seen": 2206990336
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033445674851548707,
+      "loss": 0.0674,
+      "theoretical_loss": 3.431042560560477,
+      "tokens_seen": 2207252480
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003343765045739047,
+      "loss": 0.069,
+      "theoretical_loss": 3.431009519112578,
+      "tokens_seen": 2207514624
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033429626063232224,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4309764826866287,
+      "tokens_seen": 2207776768
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033421601669073985,
+      "loss": 0.0675,
+      "theoretical_loss": 3.43094345128127,
+      "tokens_seen": 2208038912
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033413577274915747,
+      "loss": 0.0699,
+      "theoretical_loss": 3.430910424895143,
+      "tokens_seen": 2208301056
+    },
+    {
+      "epoch": 0.67,
+      "objective/train/advantage_avg": 0.0003184771630913019,
+      "objective/train/docs_used": 804023,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4095585346221924,
+      "objective/train/original_loss": 1.4095585346221924,
+      "objective/train/theoretical_loss": 3.4308774035268885,
+      "objective/train/tokens_used": 2229023200,
+      "objective/train/value_avg": -0.00682830810546875,
+      "objective/train/value_loss": 0.0001328888174612075,
+      "objective/train/value_max": -4.3332576751708984e-05,
+      "objective/train/value_min": -0.2161865234375,
+      "objective/train/value_reward_corr": 0.7068750377978834,
+      "objective/train/value_std": 0.01126861572265625,
+      "objective/train/weight_avg": 1.0003799200057983,
+      "objective/train/weighted_lm_loss": 1.4102829694747925,
+      "objective/train/weights_max": 1.1803033351898193,
+      "objective/train/weights_min": 0.36821746826171875,
+      "theoretical_loss": 3.4308774035268885,
+      "tokens_seen": 2208563200
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003340555288075751,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4308774035268885,
+      "tokens_seen": 2208563200
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003339752848659926,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4308443871751497,
+      "tokens_seen": 2208825344
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003338950409244102,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4308113758385685,
+      "tokens_seen": 2209087488
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003338147969828278,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4307783695157887,
+      "tokens_seen": 2209349632
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033373455304124537,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4307453682054536,
+      "tokens_seen": 2209611776
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.000333654309099663,
+      "loss": 0.0674,
+      "theoretical_loss": 3.430712371906208,
+      "tokens_seen": 2209873920
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003335740651580806,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4306793806166955,
+      "tokens_seen": 2210136064
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003334938212164982,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4306463943355627,
+      "tokens_seen": 2210398208
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003334135772749157,
+      "loss": 0.0672,
+      "theoretical_loss": 3.430613413061455,
+      "tokens_seen": 2210660352
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003333333333333333,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4305804367930186,
+      "tokens_seen": 2210922496
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033325308939175093,
+      "loss": 0.0639,
+      "theoretical_loss": 3.4305474655289006,
+      "tokens_seen": 2211184640
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003331728454501685,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4305144992677485,
+      "tokens_seen": 2211446784
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003330926015085861,
+      "loss": 0.066,
+      "theoretical_loss": 3.4304815380082103,
+      "tokens_seen": 2211708928
+    },
+    {
+      "epoch": 0.67,
+      "objective/train/advantage_avg": -0.00014401759835891426,
+      "objective/train/docs_used": 805208,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3896081447601318,
+      "objective/train/original_loss": 1.3896081447601318,
+      "objective/train/theoretical_loss": 3.4304650592536237,
+      "objective/train/tokens_used": 2232300000,
+      "objective/train/value_avg": -0.00646209716796875,
+      "objective/train/value_loss": 0.00011668778461171314,
+      "objective/train/value_max": -2.7120113372802734e-05,
+      "objective/train/value_min": -0.60205078125,
+      "objective/train/value_reward_corr": 0.7311364752772134,
+      "objective/train/value_std": 0.01180267333984375,
+      "objective/train/weight_avg": 0.999912440776825,
+      "objective/train/weighted_lm_loss": 1.389186978340149,
+      "objective/train/weights_max": 1.1917989253997803,
+      "objective/train/weights_min": 0.534700870513916,
+      "theoretical_loss": 3.4304650592536237,
+      "tokens_seen": 2211840000
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003330123575670037,
+      "loss": 0.065,
+      "theoretical_loss": 3.430448581748934,
+      "tokens_seen": 2211971072
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033293211362542133,
+      "loss": 0.0692,
+      "theoretical_loss": 3.430415630488569,
+      "tokens_seen": 2212233216
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003328518696838389,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4303826842257648,
+      "tokens_seen": 2212495360
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033277162574225645,
+      "loss": 0.0677,
+      "theoretical_loss": 3.430349742959171,
+      "tokens_seen": 2212757504
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033269138180067406,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4303168066874385,
+      "tokens_seen": 2213019648
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003326111378590916,
+      "loss": 0.0671,
+      "theoretical_loss": 3.430283875409219,
+      "tokens_seen": 2213281792
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033253089391750923,
+      "loss": 0.065,
+      "theoretical_loss": 3.430250949123163,
+      "tokens_seen": 2213543936
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033245064997592684,
+      "loss": 0.0669,
+      "theoretical_loss": 3.430218027827924,
+      "tokens_seen": 2213806080
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033237040603434446,
+      "loss": 0.068,
+      "theoretical_loss": 3.430185111522153,
+      "tokens_seen": 2214068224
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.000332290162092762,
+      "loss": 0.0642,
+      "theoretical_loss": 3.4301522002045046,
+      "tokens_seen": 2214330368
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003322099181511796,
+      "loss": 0.0692,
+      "theoretical_loss": 3.430119293873632,
+      "tokens_seen": 2214592512
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003321296742095972,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4300863925281893,
+      "tokens_seen": 2214854656
+    },
+    {
+      "epoch": 0.67,
+      "objective/train/advantage_avg": 0.000626943598035723,
+      "objective/train/docs_used": 806281,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2398408651351929,
+      "objective/train/original_loss": 1.2398409843444824,
+      "objective/train/theoretical_loss": 3.430053496166831,
+      "objective/train/tokens_used": 2235576800,
+      "objective/train/value_avg": -0.0085296630859375,
+      "objective/train/value_loss": 0.00041652534855529666,
+      "objective/train/value_max": -4.297494888305664e-05,
+      "objective/train/value_min": -0.98779296875,
+      "objective/train/value_reward_corr": 0.6897977803093197,
+      "objective/train/value_std": 0.021728515625,
+      "objective/train/weight_avg": 1.0008251667022705,
+      "objective/train/weighted_lm_loss": 1.240440845489502,
+      "objective/train/weights_max": 1.96941077709198,
+      "objective/train/weights_min": 0.3877526521682739,
+      "theoretical_loss": 3.430053496166831,
+      "tokens_seen": 2215116800
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033204943026801474,
+      "loss": 0.0661,
+      "theoretical_loss": 3.430053496166831,
+      "tokens_seen": 2215116800
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033196918632643236,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4300206047882136,
+      "tokens_seen": 2215378944
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033188894238484997,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4299877183909917,
+      "tokens_seen": 2215641088
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033180869844326753,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4299548369738218,
+      "tokens_seen": 2215903232
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033172845450168514,
+      "loss": 0.0647,
+      "theoretical_loss": 3.429921960535361,
+      "tokens_seen": 2216165376
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003316482105601027,
+      "loss": 0.0666,
+      "theoretical_loss": 3.429889089074267,
+      "tokens_seen": 2216427520
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003315679666185203,
+      "loss": 0.0636,
+      "theoretical_loss": 3.429856222589197,
+      "tokens_seen": 2216689664
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033148772267693787,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4298233610788094,
+      "tokens_seen": 2216951808
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003314074787353555,
+      "loss": 0.067,
+      "theoretical_loss": 3.429790504541764,
+      "tokens_seen": 2217213952
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003313272347937731,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4297576529767193,
+      "tokens_seen": 2217476096
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033124699085219065,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4297248063823362,
+      "tokens_seen": 2217738240
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033116674691060827,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4296919647572746,
+      "tokens_seen": 2218000384
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003310865029690259,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4296591281001954,
+      "tokens_seen": 2218262528
+    },
+    {
+      "epoch": 0.67,
+      "objective/train/advantage_avg": -0.001073150895535946,
+      "objective/train/docs_used": 807316,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.303073525428772,
+      "objective/train/original_loss": 1.303073525428772,
+      "objective/train/theoretical_loss": 3.429642711634231,
+      "objective/train/tokens_used": 2238853600,
+      "objective/train/value_avg": -0.0086517333984375,
+      "objective/train/value_loss": 0.0002828251163009554,
+      "objective/train/value_max": -2.2113323211669922e-05,
+      "objective/train/value_min": -0.83544921875,
+      "objective/train/value_reward_corr": 0.7588367060954335,
+      "objective/train/value_std": 0.0181427001953125,
+      "objective/train/weight_avg": 0.99905925989151,
+      "objective/train/weighted_lm_loss": 1.3010692596435547,
+      "objective/train/weights_max": 1.2279506921768188,
+      "objective/train/weights_min": 0.37165406346321106,
+      "theoretical_loss": 3.429642711634231,
+      "tokens_seen": 2218393600
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033100625902744344,
+      "loss": 0.0644,
+      "theoretical_loss": 3.42962629640976,
+      "tokens_seen": 2218524672
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.000330926015085861,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4295934696846313,
+      "tokens_seen": 2218786816
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003308457711442786,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4295606479234713,
+      "tokens_seen": 2219048960
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003307655272026962,
+      "loss": 0.066,
+      "theoretical_loss": 3.429527831124943,
+      "tokens_seen": 2219311104
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003306852832611138,
+      "loss": 0.0673,
+      "theoretical_loss": 3.42949501928771,
+      "tokens_seen": 2219573248
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003306050393195314,
+      "loss": 0.065,
+      "theoretical_loss": 3.4294622124104364,
+      "tokens_seen": 2219835392
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.000330524795377949,
+      "loss": 0.0639,
+      "theoretical_loss": 3.429429410491787,
+      "tokens_seen": 2220097536
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033044455143636656,
+      "loss": 0.0658,
+      "theoretical_loss": 3.429396613530427,
+      "tokens_seen": 2220359680
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003303643074947841,
+      "loss": 0.0667,
+      "theoretical_loss": 3.429363821525022,
+      "tokens_seen": 2220621824
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033028406355320173,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4293310344742385,
+      "tokens_seen": 2220883968
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00033020381961161935,
+      "loss": 0.0649,
+      "theoretical_loss": 3.429298252376743,
+      "tokens_seen": 2221146112
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003301235756700369,
+      "loss": 0.0638,
+      "theoretical_loss": 3.429265475231202,
+      "tokens_seen": 2221408256
+    },
+    {
+      "epoch": 0.67,
+      "objective/train/advantage_avg": 0.001115342602133751,
+      "objective/train/docs_used": 808517,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3400652408599854,
+      "objective/train/original_loss": 1.3400652408599854,
+      "objective/train/theoretical_loss": 3.429232703036284,
+      "objective/train/tokens_used": 2242130400,
+      "objective/train/value_avg": -0.010284423828125,
+      "objective/train/value_loss": 0.00031161116203293204,
+      "objective/train/value_max": -3.6776065826416016e-05,
+      "objective/train/value_min": -0.7216796875,
+      "objective/train/value_reward_corr": 0.7999857711414065,
+      "objective/train/value_std": 0.0249176025390625,
+      "objective/train/weight_avg": 1.0012611150741577,
+      "objective/train/weighted_lm_loss": 1.3421872854232788,
+      "objective/train/weights_max": 1.4046753644943237,
+      "objective/train/weights_min": 0.4194653332233429,
+      "theoretical_loss": 3.429232703036284,
+      "tokens_seen": 2221670400
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003300433317284545,
+      "loss": 0.065,
+      "theoretical_loss": 3.429232703036284,
+      "tokens_seen": 2221670400
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032996308778687213,
+      "loss": 0.067,
+      "theoretical_loss": 3.429199935790657,
+      "tokens_seen": 2221932544
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003298828438452897,
+      "loss": 0.069,
+      "theoretical_loss": 3.4291671734929903,
+      "tokens_seen": 2222194688
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032980259990370725,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4291344161419524,
+      "tokens_seen": 2222456832
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032972235596212486,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4291016637362137,
+      "tokens_seen": 2222718976
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032964211202054247,
+      "loss": 0.0657,
+      "theoretical_loss": 3.429068916274444,
+      "tokens_seen": 2222981120
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032956186807896003,
+      "loss": 0.0653,
+      "theoretical_loss": 3.429036173755314,
+      "tokens_seen": 2223243264
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032948162413737764,
+      "loss": 0.0692,
+      "theoretical_loss": 3.429003436177496,
+      "tokens_seen": 2223505408
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032940138019579526,
+      "loss": 0.067,
+      "theoretical_loss": 3.4289707035396613,
+      "tokens_seen": 2223767552
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003293211362542128,
+      "loss": 0.0698,
+      "theoretical_loss": 3.428937975840482,
+      "tokens_seen": 2224029696
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032924089231263037,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4289052530786313,
+      "tokens_seen": 2224291840
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.000329160648371048,
+      "loss": 0.068,
+      "theoretical_loss": 3.4288725352527827,
+      "tokens_seen": 2224553984
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003290804044294656,
+      "loss": 0.0662,
+      "theoretical_loss": 3.42883982236161,
+      "tokens_seen": 2224816128
+    },
+    {
+      "epoch": 0.67,
+      "objective/train/advantage_avg": 0.00034239990054629743,
+      "objective/train/docs_used": 809850,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4501973390579224,
+      "objective/train/original_loss": 1.450197458267212,
+      "objective/train/theoretical_loss": 3.4288234677661125,
+      "objective/train/tokens_used": 2245407200,
+      "objective/train/value_avg": -0.005359649658203125,
+      "objective/train/value_loss": 0.00012629618868231773,
+      "objective/train/value_max": -3.8504600524902344e-05,
+      "objective/train/value_min": -0.23193359375,
+      "objective/train/value_reward_corr": 0.6468167905238793,
+      "objective/train/value_std": 0.00925445556640625,
+      "objective/train/weight_avg": 1.0004007816314697,
+      "objective/train/weighted_lm_loss": 1.4504789113998413,
+      "objective/train/weights_max": 1.1305923461914062,
+      "objective/train/weights_min": 0.3682044744491577,
+      "theoretical_loss": 3.4288234677661125,
+      "tokens_seen": 2224947200
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032900016048788316,
+      "loss": 0.0648,
+      "theoretical_loss": 3.428807114403787,
+      "tokens_seen": 2225078272
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032891991654630077,
+      "loss": 0.0626,
+      "theoretical_loss": 3.428774411377989,
+      "tokens_seen": 2225340416
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003288396726047184,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4287417132828923,
+      "tokens_seen": 2225602560
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.00032875942866313594,
+      "loss": 0.0656,
+      "theoretical_loss": 3.428709020117172,
+      "tokens_seen": 2225864704
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003286791847215535,
+      "loss": 0.0646,
+      "theoretical_loss": 3.4286763318795046,
+      "tokens_seen": 2226126848
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003285989407799711,
+      "loss": 0.0639,
+      "theoretical_loss": 3.428643648568567,
+      "tokens_seen": 2226388992
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003285186968383887,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4286109701830374,
+      "tokens_seen": 2226651136
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003284384528968063,
+      "loss": 0.0677,
+      "theoretical_loss": 3.428578296721593,
+      "tokens_seen": 2226913280
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 0.0003283582089552239,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4285456281829125,
+      "tokens_seen": 2227175424
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003282779650136415,
+      "loss": 0.0694,
+      "theoretical_loss": 3.428512964565675,
+      "tokens_seen": 2227437568
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032819772107205907,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4284803058685602,
+      "tokens_seen": 2227699712
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003281174771304767,
+      "loss": 0.0657,
+      "theoretical_loss": 3.428447652090248,
+      "tokens_seen": 2227961856
+    },
+    {
+      "epoch": 0.68,
+      "objective/train/advantage_avg": 0.0006519161979667842,
+      "objective/train/docs_used": 811140,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.225870966911316,
+      "objective/train/original_loss": 1.2258708477020264,
+      "objective/train/theoretical_loss": 3.4284150032294187,
+      "objective/train/tokens_used": 2248684000,
+      "objective/train/value_avg": -0.00849151611328125,
+      "objective/train/value_loss": 0.0002140124561265111,
+      "objective/train/value_max": -2.9087066650390625e-05,
+      "objective/train/value_min": -0.2666015625,
+      "objective/train/value_reward_corr": 0.770223641316452,
+      "objective/train/value_std": 0.01739501953125,
+      "objective/train/weight_avg": 1.0007503032684326,
+      "objective/train/weighted_lm_loss": 1.2265666723251343,
+      "objective/train/weights_max": 1.1595954895019531,
+      "objective/train/weights_min": 0.3700695335865021,
+      "theoretical_loss": 3.4284150032294187,
+      "tokens_seen": 2228224000
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032803723318889424,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4284150032294187,
+      "tokens_seen": 2228224000
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003279569892473118,
+      "loss": 0.0666,
+      "theoretical_loss": 3.428382359284754,
+      "tokens_seen": 2228486144
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003278767453057294,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4283497202549347,
+      "tokens_seen": 2228748288
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.000327796501364147,
+      "loss": 0.069,
+      "theoretical_loss": 3.4283170861386436,
+      "tokens_seen": 2229010432
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032771625742256463,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4282844569345623,
+      "tokens_seen": 2229272576
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003276360134809822,
+      "loss": 0.0675,
+      "theoretical_loss": 3.428251832641375,
+      "tokens_seen": 2229534720
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003275557695393998,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4282192132577647,
+      "tokens_seen": 2229796864
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032747552559781736,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4281865987824154,
+      "tokens_seen": 2230059008
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003273952816562349,
+      "loss": 0.0661,
+      "theoretical_loss": 3.428153989214012,
+      "tokens_seen": 2230321152
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032731503771465253,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4281213845512397,
+      "tokens_seen": 2230583296
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032723479377307015,
+      "loss": 0.0642,
+      "theoretical_loss": 3.428088784792784,
+      "tokens_seen": 2230845440
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032715454983148776,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4280561899373305,
+      "tokens_seen": 2231107584
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003270743058899053,
+      "loss": 0.0673,
+      "theoretical_loss": 3.428023599983567,
+      "tokens_seen": 2231369728
+    },
+    {
+      "epoch": 0.68,
+      "objective/train/advantage_avg": 0.00044635971426032484,
+      "objective/train/docs_used": 812344,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1689112186431885,
+      "objective/train/original_loss": 1.1689109802246094,
+      "objective/train/theoretical_loss": 3.4280073068444086,
+      "objective/train/tokens_used": 2251960800,
+      "objective/train/value_avg": -0.007732391357421875,
+      "objective/train/value_loss": 0.0002456700021866709,
+      "objective/train/value_max": -5.435943603515625e-05,
+      "objective/train/value_min": -0.2529296875,
+      "objective/train/value_reward_corr": 0.6849964703725564,
+      "objective/train/value_std": 0.01465606689453125,
+      "objective/train/weight_avg": 1.0005552768707275,
+      "objective/train/weighted_lm_loss": 1.1684818267822266,
+      "objective/train/weights_max": 1.287792682647705,
+      "objective/train/weights_min": 0.36833932995796204,
+      "theoretical_loss": 3.4280073068444086,
+      "tokens_seen": 2231500800
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032699406194832293,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4279910149301798,
+      "tokens_seen": 2231631872
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003269138180067405,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4279584347758565,
+      "tokens_seen": 2231894016
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032683357406515805,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4279258595192856,
+      "tokens_seen": 2232156160
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032675333012357566,
+      "loss": 0.0648,
+      "theoretical_loss": 3.427893289159156,
+      "tokens_seen": 2232418304
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032667308618199327,
+      "loss": 0.0671,
+      "theoretical_loss": 3.427860723694156,
+      "tokens_seen": 2232680448
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003265928422404109,
+      "loss": 0.0677,
+      "theoretical_loss": 3.427828163122976,
+      "tokens_seen": 2232942592
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032651259829882844,
+      "loss": 0.0671,
+      "theoretical_loss": 3.427795607444306,
+      "tokens_seen": 2233204736
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032643235435724606,
+      "loss": 0.069,
+      "theoretical_loss": 3.4277630566568367,
+      "tokens_seen": 2233466880
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032635211041566367,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4277305107592593,
+      "tokens_seen": 2233729024
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032627186647408117,
+      "loss": 0.0682,
+      "theoretical_loss": 3.427697969750265,
+      "tokens_seen": 2233991168
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003261916225324988,
+      "loss": 0.0684,
+      "theoretical_loss": 3.427665433628547,
+      "tokens_seen": 2234253312
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003261113785909164,
+      "loss": 0.0678,
+      "theoretical_loss": 3.427632902392797,
+      "tokens_seen": 2234515456
+    },
+    {
+      "epoch": 0.68,
+      "objective/train/advantage_avg": 0.0008528545149601996,
+      "objective/train/docs_used": 813661,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3385423421859741,
+      "objective/train/original_loss": 1.3385424613952637,
+      "objective/train/theoretical_loss": 3.427600376041709,
+      "objective/train/tokens_used": 2255237600,
+      "objective/train/value_avg": -0.0046844482421875,
+      "objective/train/value_loss": 7.787495269440114e-05,
+      "objective/train/value_max": -3.218650817871094e-05,
+      "objective/train/value_min": -0.242919921875,
+      "objective/train/value_reward_corr": 0.7000210827951145,
+      "objective/train/value_std": 0.00942230224609375,
+      "objective/train/weight_avg": 1.0008913278579712,
+      "objective/train/weighted_lm_loss": 1.3399440050125122,
+      "objective/train/weights_max": 1.1798712015151978,
+      "objective/train/weights_min": 0.8197346925735474,
+      "theoretical_loss": 3.427600376041709,
+      "tokens_seen": 2234777600
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032603113464933396,
+      "loss": 0.0671,
+      "theoretical_loss": 3.427600376041709,
+      "tokens_seen": 2234777600
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032595089070775157,
+      "loss": 0.0647,
+      "theoretical_loss": 3.4275678545739763,
+      "tokens_seen": 2235039744
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003258706467661692,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4275353379882927,
+      "tokens_seen": 2235301888
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003257904028245868,
+      "loss": 0.0669,
+      "theoretical_loss": 3.427502826283354,
+      "tokens_seen": 2235564032
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003257101588830043,
+      "loss": 0.0668,
+      "theoretical_loss": 3.427470319457854,
+      "tokens_seen": 2235826176
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003256299149414219,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4274378175104894,
+      "tokens_seen": 2236088320
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003255496709998395,
+      "loss": 0.0659,
+      "theoretical_loss": 3.427405320439956,
+      "tokens_seen": 2236350464
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003254694270582571,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4273728282449514,
+      "tokens_seen": 2236612608
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003253891831166747,
+      "loss": 0.066,
+      "theoretical_loss": 3.4273403409241716,
+      "tokens_seen": 2236874752
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003253089391750923,
+      "loss": 0.0667,
+      "theoretical_loss": 3.427307858476315,
+      "tokens_seen": 2237136896
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003252286952335099,
+      "loss": 0.069,
+      "theoretical_loss": 3.4272753809000793,
+      "tokens_seen": 2237399040
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003251484512919275,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4272429081941636,
+      "tokens_seen": 2237661184
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032506820735034504,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4272104403572667,
+      "tokens_seen": 2237923328
+    },
+    {
+      "epoch": 0.68,
+      "objective/train/advantage_avg": 0.0004680883139371872,
+      "objective/train/docs_used": 814894,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2324265241622925,
+      "objective/train/original_loss": 1.232426404953003,
+      "objective/train/theoretical_loss": 3.4271942082642948,
+      "objective/train/tokens_used": 2258514400,
+      "objective/train/value_avg": -0.0095062255859375,
+      "objective/train/value_loss": 0.0003552815178409219,
+      "objective/train/value_max": -3.6776065826416016e-05,
+      "objective/train/value_min": -0.6796875,
+      "objective/train/value_reward_corr": 0.7258998078479855,
+      "objective/train/value_std": 0.0178985595703125,
+      "objective/train/weight_avg": 1.0006299018859863,
+      "objective/train/weighted_lm_loss": 1.2327933311462402,
+      "objective/train/weights_max": 1.376213550567627,
+      "objective/train/weights_min": 0.3958108723163605,
+      "theoretical_loss": 3.4271942082642948,
+      "tokens_seen": 2238054400
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032498796340876265,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4271779773880895,
+      "tokens_seen": 2238185472
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003249077194671802,
+      "loss": 0.0657,
+      "theoretical_loss": 3.427145519285331,
+      "tokens_seen": 2238447616
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003248274755255978,
+      "loss": 0.0694,
+      "theoretical_loss": 3.427113066047692,
+      "tokens_seen": 2238709760
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032474723158401543,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4270806176738744,
+      "tokens_seen": 2238971904
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032466698764243304,
+      "loss": 0.067,
+      "theoretical_loss": 3.4270481741625796,
+      "tokens_seen": 2239234048
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003245867437008506,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4270157355125095,
+      "tokens_seen": 2239496192
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032450649975926816,
+      "loss": 0.0687,
+      "theoretical_loss": 3.426983301722367,
+      "tokens_seen": 2239758336
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003244262558176858,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4269508727908553,
+      "tokens_seen": 2240020480
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032434601187610333,
+      "loss": 0.0668,
+      "theoretical_loss": 3.426918448716678,
+      "tokens_seen": 2240282624
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032426576793452095,
+      "loss": 0.0681,
+      "theoretical_loss": 3.42688602949854,
+      "tokens_seen": 2240544768
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032418552399293856,
+      "loss": 0.064,
+      "theoretical_loss": 3.426853615135145,
+      "tokens_seen": 2240806912
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032410528005135617,
+      "loss": 0.0663,
+      "theoretical_loss": 3.426821205625199,
+      "tokens_seen": 2241069056
+    },
+    {
+      "epoch": 0.68,
+      "objective/train/advantage_avg": 0.0005579056451097131,
+      "objective/train/docs_used": 816099,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3084162473678589,
+      "objective/train/original_loss": 1.3084161281585693,
+      "objective/train/theoretical_loss": 3.4267888009674072,
+      "objective/train/tokens_used": 2261791200,
+      "objective/train/value_avg": -0.0101470947265625,
+      "objective/train/value_loss": 0.0003198507765773684,
+      "objective/train/value_max": -2.6881694793701172e-05,
+      "objective/train/value_min": -0.7275390625,
+      "objective/train/value_reward_corr": 0.7792313074359762,
+      "objective/train/value_std": 0.0230712890625,
+      "objective/train/weight_avg": 1.0007097721099854,
+      "objective/train/weighted_lm_loss": 1.3082178831100464,
+      "objective/train/weights_max": 1.6977381706237793,
+      "objective/train/weights_min": 0.39976832270622253,
+      "theoretical_loss": 3.4267888009674072,
+      "tokens_seen": 2241331200
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032402503610977373,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4267888009674072,
+      "tokens_seen": 2241331200
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003239447921681913,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4267564011604756,
+      "tokens_seen": 2241593344
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003238645482266089,
+      "loss": 0.0652,
+      "theoretical_loss": 3.426724006203112,
+      "tokens_seen": 2241855488
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032378430428502646,
+      "loss": 0.0648,
+      "theoretical_loss": 3.426691616094022,
+      "tokens_seen": 2242117632
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032370406034344407,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4266592308319144,
+      "tokens_seen": 2242379776
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003236238164018617,
+      "loss": 0.067,
+      "theoretical_loss": 3.426626850415497,
+      "tokens_seen": 2242641920
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032354357246027924,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4265944748434785,
+      "tokens_seen": 2242904064
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032346332851869685,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4265621041145677,
+      "tokens_seen": 2243166208
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032338308457711447,
+      "loss": 0.0641,
+      "theoretical_loss": 3.426529738227475,
+      "tokens_seen": 2243428352
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.000323302840635532,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4264973771809104,
+      "tokens_seen": 2243690496
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003232225966939496,
+      "loss": 0.0666,
+      "theoretical_loss": 3.426465020973584,
+      "tokens_seen": 2243952640
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003231423527523672,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4264326696042073,
+      "tokens_seen": 2244214784
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003230621088107848,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4264003230714923,
+      "tokens_seen": 2244476928
+    },
+    {
+      "epoch": 0.68,
+      "objective/train/advantage_avg": -2.073908763122745e-05,
+      "objective/train/docs_used": 817136,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3620100021362305,
+      "objective/train/original_loss": 1.3620097637176514,
+      "objective/train/theoretical_loss": 3.42638415161848,
+      "objective/train/tokens_used": 2265068000,
+      "objective/train/value_avg": -0.00905609130859375,
+      "objective/train/value_loss": 0.0004213712236378342,
+      "objective/train/value_max": -7.253885269165039e-05,
+      "objective/train/value_min": -0.34912109375,
+      "objective/train/value_reward_corr": 0.6160935511149752,
+      "objective/train/value_std": 0.01435089111328125,
+      "objective/train/weight_avg": 1.0001745223999023,
+      "objective/train/weighted_lm_loss": 1.3625872135162354,
+      "objective/train/weights_max": 1.1662676334381104,
+      "objective/train/weights_min": 0.37542152404785156,
+      "theoretical_loss": 3.42638415161848,
+      "tokens_seen": 2244608000
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032298186486920237,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4263679813741503,
+      "tokens_seen": 2244739072
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032290162092762,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4263356445108943,
+      "tokens_seen": 2245001216
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003228213769860376,
+      "loss": 0.0649,
+      "theoretical_loss": 3.426303312480438,
+      "tokens_seen": 2245263360
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032274113304445515,
+      "loss": 0.0688,
+      "theoretical_loss": 3.426270985281494,
+      "tokens_seen": 2245525504
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003226608891028727,
+      "loss": 0.0668,
+      "theoretical_loss": 3.426238662912777,
+      "tokens_seen": 2245787648
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003225806451612903,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4262063453730014,
+      "tokens_seen": 2246049792
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032250040121970793,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4261740326608825,
+      "tokens_seen": 2246311936
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003224201572781255,
+      "loss": 0.0647,
+      "theoretical_loss": 3.4261417247751353,
+      "tokens_seen": 2246574080
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003223399133365431,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4261094217144765,
+      "tokens_seen": 2246836224
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003222596693949607,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4260771234776226,
+      "tokens_seen": 2247098368
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003221794254533782,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4260448300632906,
+      "tokens_seen": 2247360512
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032209918151179584,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4260125414701976,
+      "tokens_seen": 2247622656
+    },
+    {
+      "epoch": 0.68,
+      "objective/train/advantage_avg": 0.0007315482362173498,
+      "objective/train/docs_used": 818241,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2269222736358643,
+      "objective/train/original_loss": 1.2269222736358643,
+      "objective/train/theoretical_loss": 3.4259802576970624,
+      "objective/train/tokens_used": 2268344800,
+      "objective/train/value_avg": -0.0052032470703125,
+      "objective/train/value_loss": 9.3780858151149e-05,
+      "objective/train/value_max": -4.3332576751708984e-05,
+      "objective/train/value_min": -0.3203125,
+      "objective/train/value_reward_corr": 0.6977430695561531,
+      "objective/train/value_std": 0.01030731201171875,
+      "objective/train/weight_avg": 1.000777244567871,
+      "objective/train/weighted_lm_loss": 1.2275296449661255,
+      "objective/train/weights_max": 1.1834774017333984,
+      "objective/train/weights_min": 0.6207625865936279,
+      "theoretical_loss": 3.4259802576970624,
+      "tokens_seen": 2247884800
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032201893757021345,
+      "loss": 0.0647,
+      "theoretical_loss": 3.4259802576970624,
+      "tokens_seen": 2247884800
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032193869362863106,
+      "loss": 0.0672,
+      "theoretical_loss": 3.425947978742603,
+      "tokens_seen": 2248146944
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003218584496870486,
+      "loss": 0.0666,
+      "theoretical_loss": 3.425915704605538,
+      "tokens_seen": 2248409088
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032177820574546623,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4258834352845877,
+      "tokens_seen": 2248671232
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032169796180388384,
+      "loss": 0.0664,
+      "theoretical_loss": 3.425851170778472,
+      "tokens_seen": 2248933376
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003216177178623014,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4258189110859107,
+      "tokens_seen": 2249195520
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032153747392071896,
+      "loss": 0.0658,
+      "theoretical_loss": 3.425786656205626,
+      "tokens_seen": 2249457664
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003214572299791366,
+      "loss": 0.0672,
+      "theoretical_loss": 3.425754406136338,
+      "tokens_seen": 2249719808
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003213769860375542,
+      "loss": 0.0648,
+      "theoretical_loss": 3.425722160876769,
+      "tokens_seen": 2249981952
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032129674209597174,
+      "loss": 0.0689,
+      "theoretical_loss": 3.425689920425642,
+      "tokens_seen": 2250244096
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032121649815438936,
+      "loss": 0.0667,
+      "theoretical_loss": 3.42565768478168,
+      "tokens_seen": 2250506240
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032113625421280697,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4256254539436055,
+      "tokens_seen": 2250768384
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032105601027122453,
+      "loss": 0.067,
+      "theoretical_loss": 3.4255932279101433,
+      "tokens_seen": 2251030528
+    },
+    {
+      "epoch": 0.68,
+      "objective/train/advantage_avg": 0.0012125978246331215,
+      "objective/train/docs_used": 819526,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4415205717086792,
+      "objective/train/original_loss": 1.4415202140808105,
+      "objective/train/theoretical_loss": 3.4255771166947424,
+      "objective/train/tokens_used": 2271621600,
+      "objective/train/value_avg": -0.01105499267578125,
+      "objective/train/value_loss": 0.00021423044381663203,
+      "objective/train/value_max": -5.3882598876953125e-05,
+      "objective/train/value_min": -0.497314453125,
+      "objective/train/value_reward_corr": 0.826165006112678,
+      "objective/train/value_std": 0.02288818359375,
+      "objective/train/weight_avg": 1.0013140439987183,
+      "objective/train/weighted_lm_loss": 1.4430190324783325,
+      "objective/train/weights_max": 1.2328771352767944,
+      "objective/train/weights_min": 0.37406063079833984,
+      "theoretical_loss": 3.4255771166947424,
+      "tokens_seen": 2251161600
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003209757663296421,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4255610066800166,
+      "tokens_seen": 2251292672
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003208955223880597,
+      "loss": 0.0672,
+      "theoretical_loss": 3.425528790251952,
+      "tokens_seen": 2251554816
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003208152784464773,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4254965786246734,
+      "tokens_seen": 2251816960
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032073503450489487,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4254643717969073,
+      "tokens_seen": 2252079104
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003206547905633125,
+      "loss": 0.0661,
+      "theoretical_loss": 3.42543216976738,
+      "tokens_seen": 2252341248
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003205745466217301,
+      "loss": 0.0645,
+      "theoretical_loss": 3.425399972534818,
+      "tokens_seen": 2252603392
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032049430268014765,
+      "loss": 0.0642,
+      "theoretical_loss": 3.4253677800979494,
+      "tokens_seen": 2252865536
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003204140587385652,
+      "loss": 0.0677,
+      "theoretical_loss": 3.425335592455501,
+      "tokens_seen": 2253127680
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003203338147969828,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4253034096062014,
+      "tokens_seen": 2253389824
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00032025357085540044,
+      "loss": 0.0706,
+      "theoretical_loss": 3.4252712315487797,
+      "tokens_seen": 2253651968
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.000320173326913818,
+      "loss": 0.069,
+      "theoretical_loss": 3.4252390582819645,
+      "tokens_seen": 2253914112
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003200930829722356,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4252068898044863,
+      "tokens_seen": 2254176256
+    },
+    {
+      "epoch": 0.68,
+      "objective/train/advantage_avg": -6.209603725437773e-06,
+      "objective/train/docs_used": 820753,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2472572326660156,
+      "objective/train/original_loss": 1.2472572326660156,
+      "objective/train/theoretical_loss": 3.425174726115075,
+      "objective/train/tokens_used": 2274898400,
+      "objective/train/value_avg": -0.007427215576171875,
+      "objective/train/value_loss": 9.021081496030092e-05,
+      "objective/train/value_max": -4.1961669921875e-05,
+      "objective/train/value_min": -0.2496337890625,
+      "objective/train/value_reward_corr": 0.7706789534692144,
+      "objective/train/value_std": 0.01107025146484375,
+      "objective/train/weight_avg": 1.0000383853912354,
+      "objective/train/weighted_lm_loss": 1.247596263885498,
+      "objective/train/weights_max": 1.1017274856567383,
+      "objective/train/weights_min": 0.736079752445221,
+      "theoretical_loss": 3.425174726115075,
+      "tokens_seen": 2254438400
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003200128390306532,
+      "loss": 0.0646,
+      "theoretical_loss": 3.425174726115075,
+      "tokens_seen": 2254438400
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003199325950890708,
+      "loss": 0.0659,
+      "theoretical_loss": 3.425142567212461,
+      "tokens_seen": 2254700544
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003198523511474884,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4251104130953762,
+      "tokens_seen": 2254962688
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031977210720590595,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4250782637625514,
+      "tokens_seen": 2255224832
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003196918632643235,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4250461192127193,
+      "tokens_seen": 2255486976
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003196116193227411,
+      "loss": 0.067,
+      "theoretical_loss": 3.4250139794446124,
+      "tokens_seen": 2255749120
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031953137538115873,
+      "loss": 0.0634,
+      "theoretical_loss": 3.4249818444569637,
+      "tokens_seen": 2256011264
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031945113143957635,
+      "loss": 0.065,
+      "theoretical_loss": 3.424949714248507,
+      "tokens_seen": 2256273408
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003193708874979939,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4249175888179764,
+      "tokens_seen": 2256535552
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003192906435564115,
+      "loss": 0.0644,
+      "theoretical_loss": 3.424885468164106,
+      "tokens_seen": 2256797696
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003192103996148291,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4248533522856315,
+      "tokens_seen": 2257059840
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031913015567324663,
+      "loss": 0.0704,
+      "theoretical_loss": 3.424821241181288,
+      "tokens_seen": 2257321984
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031904991173166425,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4247891348498114,
+      "tokens_seen": 2257584128
+    },
+    {
+      "epoch": 0.68,
+      "objective/train/advantage_avg": 0.001051482162438333,
+      "objective/train/docs_used": 822060,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2332792282104492,
+      "objective/train/original_loss": 1.2332792282104492,
+      "objective/train/theoretical_loss": 3.4247730834735037,
+      "objective/train/tokens_used": 2278175200,
+      "objective/train/value_avg": -0.00476837158203125,
+      "objective/train/value_loss": 7.368699880316854e-05,
+      "objective/train/value_max": -3.218650817871094e-05,
+      "objective/train/value_min": -0.31494140625,
+      "objective/train/value_reward_corr": 0.7164398341766667,
+      "objective/train/value_std": 0.0090789794921875,
+      "objective/train/weight_avg": 1.0010877847671509,
+      "objective/train/weighted_lm_loss": 1.2348461151123047,
+      "objective/train/weights_max": 1.1168068647384644,
+      "objective/train/weights_min": 0.8260654807090759,
+      "theoretical_loss": 3.4247730834735037,
+      "tokens_seen": 2257715200
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031896966779008186,
+      "loss": 0.0651,
+      "theoretical_loss": 3.424757033289939,
+      "tokens_seen": 2257846272
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031888942384849947,
+      "loss": 0.0678,
+      "theoretical_loss": 3.424724936500407,
+      "tokens_seen": 2258108416
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031880917990691703,
+      "loss": 0.0657,
+      "theoretical_loss": 3.424692844479953,
+      "tokens_seen": 2258370560
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031872893596533464,
+      "loss": 0.0689,
+      "theoretical_loss": 3.424660757227315,
+      "tokens_seen": 2258632704
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031864869202375226,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4246286747412316,
+      "tokens_seen": 2258894848
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031856844808216976,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4245965970204413,
+      "tokens_seen": 2259156992
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003184882041405874,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4245645240636833,
+      "tokens_seen": 2259419136
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.000318407960199005,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4245324558696986,
+      "tokens_seen": 2259681280
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.0003183277162574226,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4245003924372264,
+      "tokens_seen": 2259943424
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 0.00031824747231584016,
+      "loss": 0.0637,
+      "theoretical_loss": 3.424468333765008,
+      "tokens_seen": 2260205568
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031816722837425777,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4244362798517844,
+      "tokens_seen": 2260467712
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003180869844326754,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4244042306962976,
+      "tokens_seen": 2260729856
+    },
+    {
+      "epoch": 0.69,
+      "objective/train/advantage_avg": 4.142043326282874e-05,
+      "objective/train/docs_used": 823330,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.305637240409851,
+      "objective/train/original_loss": 1.3056371212005615,
+      "objective/train/theoretical_loss": 3.42437218629729,
+      "objective/train/tokens_used": 2281452000,
+      "objective/train/value_avg": -0.00960540771484375,
+      "objective/train/value_loss": 0.0003837741387542337,
+      "objective/train/value_max": -5.227327346801758e-05,
+      "objective/train/value_min": -0.70068359375,
+      "objective/train/value_reward_corr": 0.7206865724791649,
+      "objective/train/value_std": 0.0199737548828125,
+      "objective/train/weight_avg": 1.000213861465454,
+      "objective/train/weighted_lm_loss": 1.3052417039871216,
+      "objective/train/weights_max": 1.5599620342254639,
+      "objective/train/weights_min": 0.37613826990127563,
+      "theoretical_loss": 3.42437218629729,
+      "tokens_seen": 2260992000
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003180067404910929,
+      "loss": 0.0677,
+      "theoretical_loss": 3.42437218629729,
+      "tokens_seen": 2260992000
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003179264965495105,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4243401466535044,
+      "tokens_seen": 2261254144
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003178462526079281,
+      "loss": 0.0697,
+      "theoretical_loss": 3.424308111763683,
+      "tokens_seen": 2261516288
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031776600866634567,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4242760816265707,
+      "tokens_seen": 2261778432
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003176857647247633,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4242440562409113,
+      "tokens_seen": 2262040576
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003176055207831809,
+      "loss": 0.0686,
+      "theoretical_loss": 3.424212035605449,
+      "tokens_seen": 2262302720
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003175252768415985,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4241800197189294,
+      "tokens_seen": 2262564864
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.000317445032900016,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4241480085800977,
+      "tokens_seen": 2262827008
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003173647889584336,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4241160021877004,
+      "tokens_seen": 2263089152
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031728454501685124,
+      "loss": 0.0694,
+      "theoretical_loss": 3.424084000540484,
+      "tokens_seen": 2263351296
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003172043010752688,
+      "loss": 0.0662,
+      "theoretical_loss": 3.424052003637195,
+      "tokens_seen": 2263613440
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003171240571336864,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4240200114765815,
+      "tokens_seen": 2263875584
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.000317043813192104,
+      "loss": 0.0653,
+      "theoretical_loss": 3.423988024057391,
+      "tokens_seen": 2264137728
+    },
+    {
+      "epoch": 0.69,
+      "objective/train/advantage_avg": 0.0005332150030881166,
+      "objective/train/docs_used": 824641,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.458601951599121,
+      "objective/train/original_loss": 1.4586018323898315,
+      "objective/train/theoretical_loss": 3.4239720321254388,
+      "objective/train/tokens_used": 2284728800,
+      "objective/train/value_avg": -0.005550384521484375,
+      "objective/train/value_loss": 0.00015455740503966808,
+      "objective/train/value_max": -2.2470951080322266e-05,
+      "objective/train/value_min": -0.2269287109375,
+      "objective/train/value_reward_corr": 0.6511970689407572,
+      "objective/train/value_std": 0.01013946533203125,
+      "objective/train/weight_avg": 1.0006014108657837,
+      "objective/train/weighted_lm_loss": 1.4605647325515747,
+      "objective/train/weights_max": 1.1240557432174683,
+      "objective/train/weights_min": 0.36819323897361755,
+      "theoretical_loss": 3.4239720321254388,
+      "tokens_seen": 2264268800
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031696356925052163,
+      "loss": 0.0692,
+      "theoretical_loss": 3.423956041378373,
+      "tokens_seen": 2264399872
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003168833253089392,
+      "loss": 0.0659,
+      "theoretical_loss": 3.423924063438275,
+      "tokens_seen": 2264662016
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031680308136735675,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4238920902358467,
+      "tokens_seen": 2264924160
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031672283742577436,
+      "loss": 0.0666,
+      "theoretical_loss": 3.423860121769839,
+      "tokens_seen": 2265186304
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003166425934841919,
+      "loss": 0.0662,
+      "theoretical_loss": 3.423828158039001,
+      "tokens_seen": 2265448448
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031656234954260953,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4237961990420844,
+      "tokens_seen": 2265710592
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031648210560102715,
+      "loss": 0.0668,
+      "theoretical_loss": 3.42376424477784,
+      "tokens_seen": 2265972736
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031640186165944476,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4237322952450198,
+      "tokens_seen": 2266234880
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003163216177178623,
+      "loss": 0.0677,
+      "theoretical_loss": 3.423700350442376,
+      "tokens_seen": 2266497024
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003162413737762799,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4236684103686614,
+      "tokens_seen": 2266759168
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003161611298346975,
+      "loss": 0.0681,
+      "theoretical_loss": 3.423636475022629,
+      "tokens_seen": 2267021312
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031608088589311505,
+      "loss": 0.0671,
+      "theoretical_loss": 3.423604544403032,
+      "tokens_seen": 2267283456
+    },
+    {
+      "epoch": 0.69,
+      "objective/train/advantage_avg": 0.00025811055093072355,
+      "objective/train/docs_used": 825734,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2673412561416626,
+      "objective/train/original_loss": 1.2673412561416626,
+      "objective/train/theoretical_loss": 3.4235726185086257,
+      "objective/train/tokens_used": 2288005600,
+      "objective/train/value_avg": -0.006771087646484375,
+      "objective/train/value_loss": 0.00013087765546515584,
+      "objective/train/value_max": -4.83393669128418e-05,
+      "objective/train/value_min": -0.45654296875,
+      "objective/train/value_reward_corr": 0.6977213837669967,
+      "objective/train/value_std": 0.01183319091796875,
+      "objective/train/weight_avg": 1.0003224611282349,
+      "objective/train/weighted_lm_loss": 1.2671499252319336,
+      "objective/train/weights_max": 1.501617670059204,
+      "objective/train/weights_min": 0.6095831990242004,
+      "theoretical_loss": 3.4235726185086257,
+      "tokens_seen": 2267545600
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031600064195153266,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4235726185086257,
+      "tokens_seen": 2267545600
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031592039800995027,
+      "loss": 0.0665,
+      "theoretical_loss": 3.423540697338164,
+      "tokens_seen": 2267807744
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003158401540683679,
+      "loss": 0.0672,
+      "theoretical_loss": 3.423508780890402,
+      "tokens_seen": 2268069888
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031575991012678544,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4234768691640953,
+      "tokens_seen": 2268332032
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.000315679666185203,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4234449621580003,
+      "tokens_seen": 2268594176
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003155994222436206,
+      "loss": 0.067,
+      "theoretical_loss": 3.4234130598708727,
+      "tokens_seen": 2268856320
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031551917830203817,
+      "loss": 0.0667,
+      "theoretical_loss": 3.42338116230147,
+      "tokens_seen": 2269118464
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003154389343604558,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4233492694485497,
+      "tokens_seen": 2269380608
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003153586904188734,
+      "loss": 0.0641,
+      "theoretical_loss": 3.4233173813108695,
+      "tokens_seen": 2269642752
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031527844647729096,
+      "loss": 0.0663,
+      "theoretical_loss": 3.423285497887188,
+      "tokens_seen": 2269904896
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031519820253570857,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4232536191762635,
+      "tokens_seen": 2270167040
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003151179585941262,
+      "loss": 0.0682,
+      "theoretical_loss": 3.423221745176856,
+      "tokens_seen": 2270429184
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031503771465254374,
+      "loss": 0.0648,
+      "theoretical_loss": 3.423189875887725,
+      "tokens_seen": 2270691328
+    },
+    {
+      "epoch": 0.69,
+      "objective/train/advantage_avg": 0.0004905235837213695,
+      "objective/train/docs_used": 826907,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.274688720703125,
+      "objective/train/original_loss": 1.274688720703125,
+      "objective/train/theoretical_loss": 3.4231739430091253,
+      "objective/train/tokens_used": 2291282400,
+      "objective/train/value_avg": -0.005580902099609375,
+      "objective/train/value_loss": 9.647780825616792e-05,
+      "objective/train/value_max": -3.1948089599609375e-05,
+      "objective/train/value_min": -0.318603515625,
+      "objective/train/value_reward_corr": 0.7587791405220843,
+      "objective/train/value_std": 0.0113525390625,
+      "objective/train/weight_avg": 1.0005372762680054,
+      "objective/train/weighted_lm_loss": 1.2754013538360596,
+      "objective/train/weights_max": 1.1139694452285767,
+      "objective/train/weights_min": 0.7064666748046875,
+      "theoretical_loss": 3.4231739430091253,
+      "tokens_seen": 2270822400
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003149574707109613,
+      "loss": 0.0647,
+      "theoretical_loss": 3.4231580113076303,
+      "tokens_seen": 2270953472
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003148772267693789,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4231261514353335,
+      "tokens_seen": 2271215616
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003147969828277965,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4230942962695954,
+      "tokens_seen": 2271477760
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003147167388862141,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4230624458091774,
+      "tokens_seen": 2271739904
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003146364949446317,
+      "loss": 0.0684,
+      "theoretical_loss": 3.423030600052842,
+      "tokens_seen": 2272002048
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003145562510030493,
+      "loss": 0.0646,
+      "theoretical_loss": 3.422998758999352,
+      "tokens_seen": 2272264192
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031447600706146687,
+      "loss": 0.0629,
+      "theoretical_loss": 3.4229669226474697,
+      "tokens_seen": 2272526336
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003143957631198844,
+      "loss": 0.067,
+      "theoretical_loss": 3.422935090995959,
+      "tokens_seen": 2272788480
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031431551917830204,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4229032640435846,
+      "tokens_seen": 2273050624
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031423527523671965,
+      "loss": 0.0655,
+      "theoretical_loss": 3.42287144178911,
+      "tokens_seen": 2273312768
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003141550312951372,
+      "loss": 0.0715,
+      "theoretical_loss": 3.4228396242313006,
+      "tokens_seen": 2273574912
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003140747873535548,
+      "loss": 0.0646,
+      "theoretical_loss": 3.422807811368922,
+      "tokens_seen": 2273837056
+    },
+    {
+      "epoch": 0.69,
+      "objective/train/advantage_avg": 0.0011075290385633707,
+      "objective/train/docs_used": 828122,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2920111417770386,
+      "objective/train/original_loss": 1.292011022567749,
+      "objective/train/theoretical_loss": 3.422776003200739,
+      "objective/train/tokens_used": 2294559200,
+      "objective/train/value_avg": -0.007415771484375,
+      "objective/train/value_loss": 0.00022583450481761247,
+      "objective/train/value_max": -2.193450927734375e-05,
+      "objective/train/value_min": -0.55029296875,
+      "objective/train/value_reward_corr": 0.7401537018891523,
+      "objective/train/value_std": 0.0174713134765625,
+      "objective/train/weight_avg": 1.001213788986206,
+      "objective/train/weighted_lm_loss": 1.2926075458526611,
+      "objective/train/weights_max": 1.284966230392456,
+      "objective/train/weights_min": 0.3930663764476776,
+      "theoretical_loss": 3.422776003200739,
+      "tokens_seen": 2274099200
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031399454341197243,
+      "loss": 0.0658,
+      "theoretical_loss": 3.422776003200739,
+      "tokens_seen": 2274099200
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031391429947039004,
+      "loss": 0.0709,
+      "theoretical_loss": 3.4227441997255195,
+      "tokens_seen": 2274361344
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031383405552880755,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4227124009420296,
+      "tokens_seen": 2274623488
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031375381158722516,
+      "loss": 0.0654,
+      "theoretical_loss": 3.422680606849036,
+      "tokens_seen": 2274885632
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003136735676456428,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4226488174453076,
+      "tokens_seen": 2275147776
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031359332370406033,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4226170327296113,
+      "tokens_seen": 2275409920
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031351307976247795,
+      "loss": 0.0699,
+      "theoretical_loss": 3.422585252700717,
+      "tokens_seen": 2275672064
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031343283582089556,
+      "loss": 0.0677,
+      "theoretical_loss": 3.422553477357393,
+      "tokens_seen": 2275934208
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003133525918793131,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4225217066984093,
+      "tokens_seen": 2276196352
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003132723479377307,
+      "loss": 0.0646,
+      "theoretical_loss": 3.422489940722536,
+      "tokens_seen": 2276458496
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003131921039961483,
+      "loss": 0.0664,
+      "theoretical_loss": 3.422458179428543,
+      "tokens_seen": 2276720640
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003131118600545659,
+      "loss": 0.0651,
+      "theoretical_loss": 3.422426422815202,
+      "tokens_seen": 2276982784
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031303161611298346,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4223946708812845,
+      "tokens_seen": 2277244928
+    },
+    {
+      "epoch": 0.69,
+      "objective/train/advantage_avg": -0.00016944411618169397,
+      "objective/train/docs_used": 829293,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4860248565673828,
+      "objective/train/original_loss": 1.4860246181488037,
+      "objective/train/theoretical_loss": 3.422378796668726,
+      "objective/train/tokens_used": 2297836000,
+      "objective/train/value_avg": -0.0090484619140625,
+      "objective/train/value_loss": 0.00036971012013964355,
+      "objective/train/value_max": -3.349781036376953e-05,
+      "objective/train/value_min": -0.95263671875,
+      "objective/train/value_reward_corr": 0.7871249688210622,
+      "objective/train/value_std": 0.0220489501953125,
+      "objective/train/weight_avg": 0.9999957084655762,
+      "objective/train/weighted_lm_loss": 1.485609769821167,
+      "objective/train/weights_max": 1.4590624570846558,
+      "objective/train/weights_min": 0.37661492824554443,
+      "theoretical_loss": 3.422378796668726,
+      "tokens_seen": 2277376000
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031295137217140107,
+      "loss": 0.068,
+      "theoretical_loss": 3.422362923625562,
+      "tokens_seen": 2277507072
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003128711282298187,
+      "loss": 0.0669,
+      "theoretical_loss": 3.422331181046807,
+      "tokens_seen": 2277769216
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031279088428823624,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4222994431437925,
+      "tokens_seen": 2278031360
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003127106403466538,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4222677099152916,
+      "tokens_seen": 2278293504
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003126303964050714,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4222359813600782,
+      "tokens_seen": 2278555648
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.000312550152463489,
+      "loss": 0.0646,
+      "theoretical_loss": 3.422204257476927,
+      "tokens_seen": 2278817792
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003124699085219066,
+      "loss": 0.0645,
+      "theoretical_loss": 3.4221725382646118,
+      "tokens_seen": 2279079936
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003123896645803242,
+      "loss": 0.0663,
+      "theoretical_loss": 3.422140823721908,
+      "tokens_seen": 2279342080
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003123094206387418,
+      "loss": 0.0684,
+      "theoretical_loss": 3.422109113847592,
+      "tokens_seen": 2279604224
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031222917669715937,
+      "loss": 0.0643,
+      "theoretical_loss": 3.422077408640439,
+      "tokens_seen": 2279866368
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.000312148932755577,
+      "loss": 0.0671,
+      "theoretical_loss": 3.422045708099226,
+      "tokens_seen": 2280128512
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031206868881399454,
+      "loss": 0.0687,
+      "theoretical_loss": 3.42201401222273,
+      "tokens_seen": 2280390656
+    },
+    {
+      "epoch": 0.69,
+      "objective/train/advantage_avg": 0.000552852638065815,
+      "objective/train/docs_used": 830439,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.256763219833374,
+      "objective/train/original_loss": 1.256762981414795,
+      "objective/train/theoretical_loss": 3.421982321009728,
+      "objective/train/tokens_used": 2301112800,
+      "objective/train/value_avg": -0.007610321044921875,
+      "objective/train/value_loss": 0.0004640833067242056,
+      "objective/train/value_max": -4.297494888305664e-05,
+      "objective/train/value_min": -0.60205078125,
+      "objective/train/value_reward_corr": 0.6248717318173257,
+      "objective/train/value_std": 0.0151824951171875,
+      "objective/train/weight_avg": 1.000746250152588,
+      "objective/train/weighted_lm_loss": 1.257375717163086,
+      "objective/train/weights_max": 1.4147088527679443,
+      "objective/train/weights_min": 0.3931143283843994,
+      "theoretical_loss": 3.421982321009728,
+      "tokens_seen": 2280652800
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031198844487241215,
+      "loss": 0.0666,
+      "theoretical_loss": 3.421982321009728,
+      "tokens_seen": 2280652800
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003119082009308297,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4219506344589985,
+      "tokens_seen": 2280914944
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003118279569892473,
+      "loss": 0.0658,
+      "theoretical_loss": 3.421918952569319,
+      "tokens_seen": 2281177088
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031174771304766494,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4218872753394702,
+      "tokens_seen": 2281439232
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003116674691060825,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4218556027682294,
+      "tokens_seen": 2281701376
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003115872251645001,
+      "loss": 0.0645,
+      "theoretical_loss": 3.4218239348543777,
+      "tokens_seen": 2281963520
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031150698122291766,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4217922715966944,
+      "tokens_seen": 2282225664
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003114267372813352,
+      "loss": 0.0662,
+      "theoretical_loss": 3.421760612993961,
+      "tokens_seen": 2282487808
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031134649333975284,
+      "loss": 0.0663,
+      "theoretical_loss": 3.421728959044958,
+      "tokens_seen": 2282749952
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031126624939817045,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4216973097484678,
+      "tokens_seen": 2283012096
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031118600545658806,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4216656651032715,
+      "tokens_seen": 2283274240
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003111057615150056,
+      "loss": 0.0654,
+      "theoretical_loss": 3.421634025108152,
+      "tokens_seen": 2283536384
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031102551757342323,
+      "loss": 0.0659,
+      "theoretical_loss": 3.421602389761893,
+      "tokens_seen": 2283798528
+    },
+    {
+      "epoch": 0.69,
+      "objective/train/advantage_avg": 0.0004953542375005782,
+      "objective/train/docs_used": 831582,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3190346956253052,
+      "objective/train/original_loss": 1.3190345764160156,
+      "objective/train/theoretical_loss": 3.4215865738317057,
+      "objective/train/tokens_used": 2304389600,
+      "objective/train/value_avg": -0.005458831787109375,
+      "objective/train/value_loss": 0.00012545128993224353,
+      "objective/train/value_max": -2.467632293701172e-05,
+      "objective/train/value_min": -0.343017578125,
+      "objective/train/value_reward_corr": 0.6331208949384891,
+      "objective/train/value_std": 0.0095672607421875,
+      "objective/train/weight_avg": 1.0005526542663574,
+      "objective/train/weighted_lm_loss": 1.3202760219573975,
+      "objective/train/weights_max": 1.2151522636413574,
+      "objective/train/weights_min": 0.3685888350009918,
+      "theoretical_loss": 3.4215865738317057,
+      "tokens_seen": 2283929600
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003109452736318408,
+      "loss": 0.0664,
+      "theoretical_loss": 3.421570759063277,
+      "tokens_seen": 2284060672
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031086502969025835,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4215391330110885,
+      "tokens_seen": 2284322816
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031078478574867596,
+      "loss": 0.0677,
+      "theoretical_loss": 3.421507511604111,
+      "tokens_seen": 2284584960
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003107045418070936,
+      "loss": 0.0648,
+      "theoretical_loss": 3.42147589484113,
+      "tokens_seen": 2284847104
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003106242978655112,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4214442827209313,
+      "tokens_seen": 2285109248
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031054405392392874,
+      "loss": 0.0632,
+      "theoretical_loss": 3.4214126752423,
+      "tokens_seen": 2285371392
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031046380998234636,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4213810724040217,
+      "tokens_seen": 2285633536
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00031038356604076397,
+      "loss": 0.0649,
+      "theoretical_loss": 3.421349474204884,
+      "tokens_seen": 2285895680
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003103033220991815,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4213178806436737,
+      "tokens_seen": 2286157824
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003102230781575991,
+      "loss": 0.0637,
+      "theoretical_loss": 3.4212862917191784,
+      "tokens_seen": 2286419968
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003101428342160167,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4212547074301862,
+      "tokens_seen": 2286682112
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003100625902744343,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4212231277754848,
+      "tokens_seen": 2286944256
+    },
+    {
+      "epoch": 0.69,
+      "objective/train/advantage_avg": -0.0007914595771580935,
+      "objective/train/docs_used": 832763,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3256409168243408,
+      "objective/train/original_loss": 1.3256409168243408,
+      "objective/train/theoretical_loss": 3.4211915527538643,
+      "objective/train/tokens_used": 2307666400,
+      "objective/train/value_avg": -0.007171630859375,
+      "objective/train/value_loss": 0.00014895742060616612,
+      "objective/train/value_max": -2.9981136322021484e-05,
+      "objective/train/value_min": -0.1917724609375,
+      "objective/train/value_reward_corr": 0.8010973247790141,
+      "objective/train/value_std": 0.0142364501953125,
+      "objective/train/weight_avg": 0.9992815256118774,
+      "objective/train/weighted_lm_loss": 1.3248015642166138,
+      "objective/train/weights_max": 1.183910846710205,
+      "objective/train/weights_min": 0.6507918834686279,
+      "theoretical_loss": 3.4211915527538643,
+      "tokens_seen": 2287206400
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00030998234633285187,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4211915527538643,
+      "tokens_seen": 2287206400
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003099021023912695,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4211599823641134,
+      "tokens_seen": 2287468544
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003098218584496871,
+      "loss": 0.0668,
+      "theoretical_loss": 3.421128416605022,
+      "tokens_seen": 2287730688
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003097416145081046,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4210968554753807,
+      "tokens_seen": 2287992832
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003096613705665222,
+      "loss": 0.0651,
+      "theoretical_loss": 3.42106529897398,
+      "tokens_seen": 2288254976
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003095811266249398,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4210337470996106,
+      "tokens_seen": 2288517120
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003095008826833574,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4210021998510647,
+      "tokens_seen": 2288779264
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.000309420638741775,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4209706572271346,
+      "tokens_seen": 2289041408
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003093403948001926,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4209391192266128,
+      "tokens_seen": 2289303552
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003092601508586102,
+      "loss": 0.0672,
+      "theoretical_loss": 3.420907585848292,
+      "tokens_seen": 2289565696
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003091799069170278,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4208760570909655,
+      "tokens_seen": 2289827840
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00030909966297544534,
+      "loss": 0.0648,
+      "theoretical_loss": 3.420844532953428,
+      "tokens_seen": 2290089984
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00030901941903386295,
+      "loss": 0.0677,
+      "theoretical_loss": 3.420813013434473,
+      "tokens_seen": 2290352128
+    },
+    {
+      "epoch": 0.69,
+      "objective/train/advantage_avg": 0.0008832691819407046,
+      "objective/train/docs_used": 833966,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3830344676971436,
+      "objective/train/original_loss": 1.3830342292785645,
+      "objective/train/theoretical_loss": 3.4207972554065877,
+      "objective/train/tokens_used": 2310943200,
+      "objective/train/value_avg": -0.00794219970703125,
+      "objective/train/value_loss": 0.0003235115436837077,
+      "objective/train/value_max": -1.2636184692382812e-05,
+      "objective/train/value_min": -0.5380859375,
+      "objective/train/value_reward_corr": 0.663530368934879,
+      "objective/train/value_std": 0.0168304443359375,
+      "objective/train/weight_avg": 1.001022219657898,
+      "objective/train/weighted_lm_loss": 1.3845484256744385,
+      "objective/train/weights_max": 1.5663223266601562,
+      "objective/train/weights_min": 0.3698508143424988,
+      "theoretical_loss": 3.4207972554065877,
+      "tokens_seen": 2290483200
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003089391750922805,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4207814985328957,
+      "tokens_seen": 2290614272
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003088589311506981,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4207499882474917,
+      "tokens_seen": 2290876416
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00030877868720911573,
+      "loss": 0.0633,
+      "theoretical_loss": 3.4207184825770565,
+      "tokens_seen": 2291138560
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00030869844326753335,
+      "loss": 0.0644,
+      "theoretical_loss": 3.420686981520386,
+      "tokens_seen": 2291400704
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003086181993259509,
+      "loss": 0.0664,
+      "theoretical_loss": 3.420655485076277,
+      "tokens_seen": 2291662848
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00030853795538436846,
+      "loss": 0.0652,
+      "theoretical_loss": 3.420623993243527,
+      "tokens_seen": 2291924992
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.0003084577114427861,
+      "loss": 0.0661,
+      "theoretical_loss": 3.420592506020933,
+      "tokens_seen": 2292187136
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00030837746750120363,
+      "loss": 0.0646,
+      "theoretical_loss": 3.420561023407293,
+      "tokens_seen": 2292449280
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00030829722355962125,
+      "loss": 0.0674,
+      "theoretical_loss": 3.420529545401406,
+      "tokens_seen": 2292711424
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00030821697961803886,
+      "loss": 0.067,
+      "theoretical_loss": 3.42049807200207,
+      "tokens_seen": 2292973568
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 0.00030813673567645647,
+      "loss": 0.066,
+      "theoretical_loss": 3.420466603208085,
+      "tokens_seen": 2293235712
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030805649173487403,
+      "loss": 0.0634,
+      "theoretical_loss": 3.4204351390182506,
+      "tokens_seen": 2293497856
+    },
+    {
+      "debugging/Compilability": 1.0,
+      "debugging/distinct-1-grams": 0.7361949066418735,
+      "debugging/entropy-1-grams": 5.412433223050153,
+      "debugging/length": 455.6666666666667,
+      "debugging/num_segments": 15,
+      "debugging/raw_token_scores_avg": 0.004264768213033676,
+      "debugging/raw_token_scores_std": 0.017437797039747238,
+      "debugging/score": 0.0066926720884253575,
+      "debugging/score_std": 0.006256546529562537,
+      "epoch": 0.7,
+      "objective/train/advantage_avg": 0.0012232189765200019,
+      "objective/train/docs_used": 835219,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2494778633117676,
+      "objective/train/original_loss": 1.249477744102478,
+      "objective/train/theoretical_loss": 3.4204036794313675,
+      "objective/train/tokens_used": 2314220000,
+      "objective/train/value_avg": -0.005489349365234375,
+      "objective/train/value_loss": 0.00015417773101944476,
+      "objective/train/value_max": -4.5418739318847656e-05,
+      "objective/train/value_min": -0.390380859375,
+      "objective/train/value_reward_corr": 0.7059713839052515,
+      "objective/train/value_std": 0.01250457763671875,
+      "objective/train/weight_avg": 1.0012928247451782,
+      "objective/train/weighted_lm_loss": 1.251007080078125,
+      "objective/train/weights_max": 1.470705509185791,
+      "objective/train/weights_min": 0.3033490478992462,
+      "theoretical_loss": 3.4204036794313675,
+      "tokens_seen": 2293760000
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003079762477932916,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4204036794313675,
+      "tokens_seen": 2293760000
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003078960038517092,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4203722244462353,
+      "tokens_seen": 2294022144
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030781575991012676,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4203407740616565,
+      "tokens_seen": 2294284288
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003077355159685444,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4203093282764314,
+      "tokens_seen": 2294546432
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.000307655272026962,
+      "loss": 0.0653,
+      "theoretical_loss": 3.420277887089363,
+      "tokens_seen": 2294808576
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003075750280853796,
+      "loss": 0.0639,
+      "theoretical_loss": 3.420246450499253,
+      "tokens_seen": 2295070720
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030749478414379716,
+      "loss": 0.0641,
+      "theoretical_loss": 3.420215018504905,
+      "tokens_seen": 2295332864
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030741454020221477,
+      "loss": 0.063,
+      "theoretical_loss": 3.4201835911051224,
+      "tokens_seen": 2295595008
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030733429626063233,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4201521682987086,
+      "tokens_seen": 2295857152
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003072540523190499,
+      "loss": 0.068,
+      "theoretical_loss": 3.420120750084468,
+      "tokens_seen": 2296119296
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003071738083774675,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4200893364612055,
+      "tokens_seen": 2296381440
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003070935644358851,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4200579274277256,
+      "tokens_seen": 2296643584
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030701332049430267,
+      "loss": 0.067,
+      "theoretical_loss": 3.420026522982835,
+      "tokens_seen": 2296905728
+    },
+    {
+      "epoch": 0.7,
+      "objective/train/advantage_avg": 0.001160003594122827,
+      "objective/train/docs_used": 836410,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2669320106506348,
+      "objective/train/original_loss": 1.2669320106506348,
+      "objective/train/theoretical_loss": 3.4200108224807373,
+      "objective/train/tokens_used": 2317496800,
+      "objective/train/value_avg": -0.00878143310546875,
+      "objective/train/value_loss": 0.00025325262686237693,
+      "objective/train/value_max": -1.9252300262451172e-05,
+      "objective/train/value_min": -0.50830078125,
+      "objective/train/value_reward_corr": 0.6975027171589683,
+      "objective/train/value_std": 0.01409149169921875,
+      "objective/train/weight_avg": 1.0012704133987427,
+      "objective/train/weighted_lm_loss": 1.268456220626831,
+      "objective/train/weights_max": 1.1904091835021973,
+      "objective/train/weights_min": 0.3858168125152588,
+      "theoretical_loss": 3.4200108224807373,
+      "tokens_seen": 2297036800
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003069330765527203,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4199951231253394,
+      "tokens_seen": 2297167872
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003068528326111379,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4199637278540447,
+      "tokens_seen": 2297430016
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030677258866955545,
+      "loss": 0.0646,
+      "theoretical_loss": 3.4199323371677584,
+      "tokens_seen": 2297692160
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.000306692344727973,
+      "loss": 0.0659,
+      "theoretical_loss": 3.419900951065288,
+      "tokens_seen": 2297954304
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003066121007863906,
+      "loss": 0.0662,
+      "theoretical_loss": 3.419869569545441,
+      "tokens_seen": 2298216448
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030653185684480824,
+      "loss": 0.0654,
+      "theoretical_loss": 3.419838192607026,
+      "tokens_seen": 2298478592
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003064516129032258,
+      "loss": 0.0629,
+      "theoretical_loss": 3.4198068202488514,
+      "tokens_seen": 2298740736
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003063713689616434,
+      "loss": 0.067,
+      "theoretical_loss": 3.419775452469727,
+      "tokens_seen": 2299002880
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.000306291125020061,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4197440892684616,
+      "tokens_seen": 2299265024
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003062108810784786,
+      "loss": 0.066,
+      "theoretical_loss": 3.4197127306438664,
+      "tokens_seen": 2299527168
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030613063713689614,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4196813765947507,
+      "tokens_seen": 2299789312
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030605039319531375,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4196500271199266,
+      "tokens_seen": 2300051456
+    },
+    {
+      "epoch": 0.7,
+      "objective/train/advantage_avg": -8.435761628788896e-06,
+      "objective/train/docs_used": 837515,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3882620334625244,
+      "objective/train/original_loss": 1.3882619142532349,
+      "objective/train/theoretical_loss": 3.4196186822182044,
+      "objective/train/tokens_used": 2320773600,
+      "objective/train/value_avg": -0.0076446533203125,
+      "objective/train/value_loss": 0.0006114847492426634,
+      "objective/train/value_max": -7.486343383789062e-05,
+      "objective/train/value_min": -0.92431640625,
+      "objective/train/value_reward_corr": 0.6498722767181959,
+      "objective/train/value_std": 0.02294921875,
+      "objective/train/weight_avg": 1.000264286994934,
+      "objective/train/weighted_lm_loss": 1.3892766237258911,
+      "objective/train/weights_max": 1.806443452835083,
+      "objective/train/weights_min": 0.36988890171051025,
+      "theoretical_loss": 3.4196186822182044,
+      "tokens_seen": 2300313600
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030597014925373136,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4196186822182044,
+      "tokens_seen": 2300313600
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003058899053121489,
+      "loss": 0.0653,
+      "theoretical_loss": 3.419587341888397,
+      "tokens_seen": 2300575744
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030580966137056653,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4195560061293158,
+      "tokens_seen": 2300837888
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030572941742898415,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4195246749397743,
+      "tokens_seen": 2301100032
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030564917348740176,
+      "loss": 0.064,
+      "theoretical_loss": 3.4194933483185856,
+      "tokens_seen": 2301362176
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030556892954581926,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4194620262645627,
+      "tokens_seen": 2301624320
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003054886856042369,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4194307087765203,
+      "tokens_seen": 2301886464
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003054084416626545,
+      "loss": 0.0698,
+      "theoretical_loss": 3.4193993958532722,
+      "tokens_seen": 2302148608
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030532819772107205,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4193680874936345,
+      "tokens_seen": 2302410752
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030524795377948966,
+      "loss": 0.0669,
+      "theoretical_loss": 3.419336783696422,
+      "tokens_seen": 2302672896
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030516770983790727,
+      "loss": 0.0666,
+      "theoretical_loss": 3.41930548446045,
+      "tokens_seen": 2302935040
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030508746589632483,
+      "loss": 0.0689,
+      "theoretical_loss": 3.419274189784536,
+      "tokens_seen": 2303197184
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003050072219547424,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4192428996674957,
+      "tokens_seen": 2303459328
+    },
+    {
+      "epoch": 0.7,
+      "objective/train/advantage_avg": -9.100284660235047e-05,
+      "objective/train/docs_used": 838679,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.309449315071106,
+      "objective/train/original_loss": 1.309449315071106,
+      "objective/train/theoretical_loss": 3.419227256318184,
+      "objective/train/tokens_used": 2324050400,
+      "objective/train/value_avg": -0.00968170166015625,
+      "objective/train/value_loss": 0.0003752364427782595,
+      "objective/train/value_max": -3.272294998168945e-05,
+      "objective/train/value_min": -0.994140625,
+      "objective/train/value_reward_corr": 0.7353203166864406,
+      "objective/train/value_std": 0.0184783935546875,
+      "objective/train/weight_avg": 1.0000666379928589,
+      "objective/train/weighted_lm_loss": 1.3097712993621826,
+      "objective/train/weights_max": 1.30930495262146,
+      "objective/train/weights_min": 0.22673387825489044,
+      "theoretical_loss": 3.419227256318184,
+      "tokens_seen": 2303590400
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030492697801316,
+      "loss": 0.0696,
+      "theoretical_loss": 3.4192116141081463,
+      "tokens_seen": 2303721472
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003048467340715776,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4191803331053063,
+      "tokens_seen": 2303983616
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030476649012999517,
+      "loss": 0.0675,
+      "theoretical_loss": 3.419149056657793,
+      "tokens_seen": 2304245760
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003046862461884128,
+      "loss": 0.0648,
+      "theoretical_loss": 3.4191177847644254,
+      "tokens_seen": 2304507904
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003046060022468304,
+      "loss": 0.0676,
+      "theoretical_loss": 3.419086517424022,
+      "tokens_seen": 2304770048
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030452575830524796,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4190552546354023,
+      "tokens_seen": 2305032192
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030444551436366557,
+      "loss": 0.0673,
+      "theoretical_loss": 3.419023996397386,
+      "tokens_seen": 2305294336
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003043652704220831,
+      "loss": 0.0651,
+      "theoretical_loss": 3.418992742708794,
+      "tokens_seen": 2305556480
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030428502648050074,
+      "loss": 0.0654,
+      "theoretical_loss": 3.418961493568446,
+      "tokens_seen": 2305818624
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003042047825389183,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4189302489751636,
+      "tokens_seen": 2306080768
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003041245385973359,
+      "loss": 0.0665,
+      "theoretical_loss": 3.418899008927769,
+      "tokens_seen": 2306342912
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003040442946557535,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4188677734250836,
+      "tokens_seen": 2306605056
+    },
+    {
+      "epoch": 0.7,
+      "objective/train/advantage_avg": 0.000588314956985414,
+      "objective/train/docs_used": 839893,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4173892736434937,
+      "objective/train/original_loss": 1.417389154434204,
+      "objective/train/theoretical_loss": 3.4188365424659293,
+      "objective/train/tokens_used": 2327327200,
+      "objective/train/value_avg": -0.0101318359375,
+      "objective/train/value_loss": 0.00021919727441854775,
+      "objective/train/value_max": -3.707408905029297e-05,
+      "objective/train/value_min": -0.5791015625,
+      "objective/train/value_reward_corr": 0.7850794856275933,
+      "objective/train/value_std": 0.019073486328125,
+      "objective/train/weight_avg": 1.0006909370422363,
+      "objective/train/weighted_lm_loss": 1.4180923700332642,
+      "objective/train/weights_max": 1.4146628379821777,
+      "objective/train/weights_min": 0.38787099719047546,
+      "theoretical_loss": 3.4188365424659293,
+      "tokens_seen": 2306867200
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003039640507141711,
+      "loss": 0.067,
+      "theoretical_loss": 3.4188365424659293,
+      "tokens_seen": 2306867200
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003038838067725887,
+      "loss": 0.066,
+      "theoretical_loss": 3.4188053160491303,
+      "tokens_seen": 2307129344
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030380356283100625,
+      "loss": 0.0664,
+      "theoretical_loss": 3.418774094173509,
+      "tokens_seen": 2307391488
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030372331888942387,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4187428768378894,
+      "tokens_seen": 2307653632
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003036430749478414,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4187116640410955,
+      "tokens_seen": 2307915776
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030356283100625904,
+      "loss": 0.0657,
+      "theoretical_loss": 3.418680455781953,
+      "tokens_seen": 2308177920
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030348258706467665,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4186492520592853,
+      "tokens_seen": 2308440064
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003034023431230942,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4186180528719188,
+      "tokens_seen": 2308702208
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003033220991815118,
+      "loss": 0.0684,
+      "theoretical_loss": 3.41858685821868,
+      "tokens_seen": 2308964352
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003032418552399294,
+      "loss": 0.0681,
+      "theoretical_loss": 3.418555668098395,
+      "tokens_seen": 2309226496
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030316161129834694,
+      "loss": 0.0683,
+      "theoretical_loss": 3.41852448250989,
+      "tokens_seen": 2309488640
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030308136735676455,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4184933014519925,
+      "tokens_seen": 2309750784
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030300112341518216,
+      "loss": 0.065,
+      "theoretical_loss": 3.4184621249235305,
+      "tokens_seen": 2310012928
+    },
+    {
+      "epoch": 0.7,
+      "objective/train/advantage_avg": 0.00043035694397985935,
+      "objective/train/docs_used": 841155,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.224597454071045,
+      "objective/train/original_loss": 1.224597454071045,
+      "objective/train/theoretical_loss": 3.418446538357472,
+      "objective/train/tokens_used": 2330604000,
+      "objective/train/value_avg": -0.0084686279296875,
+      "objective/train/value_loss": 0.00034951031557284296,
+      "objective/train/value_max": -5.066394805908203e-05,
+      "objective/train/value_min": -0.443115234375,
+      "objective/train/value_reward_corr": 0.6553620318807173,
+      "objective/train/value_std": 0.014984130859375,
+      "objective/train/weight_avg": 1.0005806684494019,
+      "objective/train/weighted_lm_loss": 1.2254345417022705,
+      "objective/train/weights_max": 1.3982270956039429,
+      "objective/train/weights_min": 0.3952797055244446,
+      "theoretical_loss": 3.418446538357472,
+      "tokens_seen": 2310144000
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003029208794735998,
+      "loss": 0.0658,
+      "theoretical_loss": 3.418430952923332,
+      "tokens_seen": 2310275072
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030284063553201733,
+      "loss": 0.0684,
+      "theoretical_loss": 3.418399785450226,
+      "tokens_seen": 2310537216
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030276039159043495,
+      "loss": 0.065,
+      "theoretical_loss": 3.418368622503041,
+      "tokens_seen": 2310799360
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030268014764885256,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4183374640806066,
+      "tokens_seen": 2311061504
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030259990370727006,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4183063101817526,
+      "tokens_seen": 2311323648
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003025196597656877,
+      "loss": 0.0686,
+      "theoretical_loss": 3.41827516080531,
+      "tokens_seen": 2311585792
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003024394158241053,
+      "loss": 0.0699,
+      "theoretical_loss": 3.418244015950108,
+      "tokens_seen": 2311847936
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003023591718825229,
+      "loss": 0.0655,
+      "theoretical_loss": 3.418212875614979,
+      "tokens_seen": 2312110080
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030227892794094046,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4181817397987553,
+      "tokens_seen": 2312372224
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030219868399935807,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4181506085002673,
+      "tokens_seen": 2312634368
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003021184400577757,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4181194817183487,
+      "tokens_seen": 2312896512
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003020381961161932,
+      "loss": 0.0668,
+      "theoretical_loss": 3.418088359451832,
+      "tokens_seen": 2313158656
+    },
+    {
+      "epoch": 0.7,
+      "objective/train/advantage_avg": 0.0005539424018934369,
+      "objective/train/docs_used": 842373,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2508080005645752,
+      "objective/train/original_loss": 1.250807762145996,
+      "objective/train/theoretical_loss": 3.41805724169955,
+      "objective/train/tokens_used": 2333880800,
+      "objective/train/value_avg": -0.005340576171875,
+      "objective/train/value_loss": 0.00012613584112841636,
+      "objective/train/value_max": -3.0934810638427734e-05,
+      "objective/train/value_min": -0.2215576171875,
+      "objective/train/value_reward_corr": 0.6626319954391378,
+      "objective/train/value_std": 0.00977325439453125,
+      "objective/train/weight_avg": 1.0006122589111328,
+      "objective/train/weighted_lm_loss": 1.251070261001587,
+      "objective/train/weights_max": 1.145806074142456,
+      "objective/train/weights_min": 0.3807637691497803,
+      "theoretical_loss": 3.41805724169955,
+      "tokens_seen": 2313420800
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003019579521746108,
+      "loss": 0.0679,
+      "theoretical_loss": 3.41805724169955,
+      "tokens_seen": 2313420800
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003018777082330284,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4180261284603373,
+      "tokens_seen": 2313682944
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.000301797464291446,
+      "loss": 0.0693,
+      "theoretical_loss": 3.417995019733028,
+      "tokens_seen": 2313945088
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003017172203498636,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4179639155164567,
+      "tokens_seen": 2314207232
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003016369764082812,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4179328158094586,
+      "tokens_seen": 2314469376
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003015567324666988,
+      "loss": 0.0661,
+      "theoretical_loss": 3.417901720610869,
+      "tokens_seen": 2314731520
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003014764885251163,
+      "loss": 0.0642,
+      "theoretical_loss": 3.417870629919524,
+      "tokens_seen": 2314993664
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003013962445835339,
+      "loss": 0.0678,
+      "theoretical_loss": 3.41783954373426,
+      "tokens_seen": 2315255808
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030131600064195154,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4178084620539138,
+      "tokens_seen": 2315517952
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003012357567003691,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4177773848773225,
+      "tokens_seen": 2315780096
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003011555127587867,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4177463122033243,
+      "tokens_seen": 2316042240
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003010752688172043,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4177152440307568,
+      "tokens_seen": 2316304384
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030099502487562194,
+      "loss": 0.067,
+      "theoretical_loss": 3.417684180358459,
+      "tokens_seen": 2316566528
+    },
+    {
+      "epoch": 0.7,
+      "objective/train/advantage_avg": 0.0001357531436951831,
+      "objective/train/docs_used": 843608,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4189345836639404,
+      "objective/train/original_loss": 1.4189344644546509,
+      "objective/train/theoretical_loss": 3.417668650209548,
+      "objective/train/tokens_used": 2337157600,
+      "objective/train/value_avg": -0.00809478759765625,
+      "objective/train/value_loss": 0.00034320837585255504,
+      "objective/train/value_max": -5.3882598876953125e-05,
+      "objective/train/value_min": -0.5009765625,
+      "objective/train/value_reward_corr": 0.620115318252616,
+      "objective/train/value_std": 0.0153350830078125,
+      "objective/train/weight_avg": 1.0002853870391846,
+      "objective/train/weighted_lm_loss": 1.4190088510513306,
+      "objective/train/weights_max": 1.3826122283935547,
+      "objective/train/weights_min": 0.36843207478523254,
+      "theoretical_loss": 3.417668650209548,
+      "tokens_seen": 2316697600
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003009147809340395,
+      "loss": 0.0687,
+      "theoretical_loss": 3.417653121185269,
+      "tokens_seen": 2316828672
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030083453699245705,
+      "loss": 0.0704,
+      "theoretical_loss": 3.4176220665100274,
+      "tokens_seen": 2317090816
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030075429305087466,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4175910163315733,
+      "tokens_seen": 2317352960
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003006740491092922,
+      "loss": 0.0684,
+      "theoretical_loss": 3.417559970648747,
+      "tokens_seen": 2317615104
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030059380516770984,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4175289294603894,
+      "tokens_seen": 2317877248
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030051356122612745,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4174978927653417,
+      "tokens_seen": 2318139392
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030043331728454506,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4174668605624454,
+      "tokens_seen": 2318401536
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003003530733429626,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4174358328505425,
+      "tokens_seen": 2318663680
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003002728294013802,
+      "loss": 0.067,
+      "theoretical_loss": 3.4174048096284757,
+      "tokens_seen": 2318925824
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0003001925854597978,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4173737908950876,
+      "tokens_seen": 2319187968
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030011234151821535,
+      "loss": 0.0671,
+      "theoretical_loss": 3.417342776649221,
+      "tokens_seen": 2319450112
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00030003209757663296,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4173117668897204,
+      "tokens_seen": 2319712256
+    },
+    {
+      "epoch": 0.7,
+      "objective/train/advantage_avg": -0.00041747145587578416,
+      "objective/train/docs_used": 844817,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1701602935791016,
+      "objective/train/original_loss": 1.1701600551605225,
+      "objective/train/theoretical_loss": 3.4172807616154293,
+      "objective/train/tokens_used": 2340434400,
+      "objective/train/value_avg": -0.00962066650390625,
+      "objective/train/value_loss": 0.00041081372182816267,
+      "objective/train/value_max": -4.363059997558594e-05,
+      "objective/train/value_min": -0.92236328125,
+      "objective/train/value_reward_corr": 0.7432690809678663,
+      "objective/train/value_std": 0.0228118896484375,
+      "objective/train/weight_avg": 0.9997594356536865,
+      "objective/train/weighted_lm_loss": 1.1693514585494995,
+      "objective/train/weights_max": 2.4330854415893555,
+      "objective/train/weights_min": 0.15194809436798096,
+      "theoretical_loss": 3.4172807616154293,
+      "tokens_seen": 2319974400
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002999518536350506,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4172807616154293,
+      "tokens_seen": 2319974400
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002998716096934682,
+      "loss": 0.0679,
+      "theoretical_loss": 3.417249760825193,
+      "tokens_seen": 2320236544
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00029979136575188574,
+      "loss": 0.0648,
+      "theoretical_loss": 3.417218764517856,
+      "tokens_seen": 2320498688
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00029971112181030336,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4171877726922633,
+      "tokens_seen": 2320760832
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002996308778687209,
+      "loss": 0.067,
+      "theoretical_loss": 3.417156785347262,
+      "tokens_seen": 2321022976
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002995506339271385,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4171258024816975,
+      "tokens_seen": 2321285120
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002994703899855561,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4170948240944163,
+      "tokens_seen": 2321547264
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002993901460439737,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4170638501842663,
+      "tokens_seen": 2321809408
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00029930990210239126,
+      "loss": 0.0676,
+      "theoretical_loss": 3.417032880750094,
+      "tokens_seen": 2322071552
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00029922965816080887,
+      "loss": 0.0658,
+      "theoretical_loss": 3.417001915790749,
+      "tokens_seen": 2322333696
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002991494142192265,
+      "loss": 0.0647,
+      "theoretical_loss": 3.416970955305078,
+      "tokens_seen": 2322595840
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00029906917027764404,
+      "loss": 0.0677,
+      "theoretical_loss": 3.416939999291931,
+      "tokens_seen": 2322857984
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002989889263360616,
+      "loss": 0.0683,
+      "theoretical_loss": 3.416909047750157,
+      "tokens_seen": 2323120128
+    },
+    {
+      "epoch": 0.7,
+      "objective/train/advantage_avg": 0.0003804284206125885,
+      "objective/train/docs_used": 845757,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2605044841766357,
+      "objective/train/original_loss": 1.2605046033859253,
+      "objective/train/theoretical_loss": 3.4168935736556754,
+      "objective/train/tokens_used": 2343711200,
+      "objective/train/value_avg": -0.00946044921875,
+      "objective/train/value_loss": 0.0002300405321875587,
+      "objective/train/value_max": -2.6881694793701172e-05,
+      "objective/train/value_min": -0.8173828125,
+      "objective/train/value_reward_corr": 0.7705075250052723,
+      "objective/train/value_std": 0.018768310546875,
+      "objective/train/weight_avg": 1.0004947185516357,
+      "objective/train/weighted_lm_loss": 1.2603367567062378,
+      "objective/train/weights_max": 1.7017693519592285,
+      "objective/train/weights_min": 0.7406313419342041,
+      "theoretical_loss": 3.4168935736556754,
+      "tokens_seen": 2323251200
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002989086823944792,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4168781006786055,
+      "tokens_seen": 2323382272
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002988284384528968,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4168471580761266,
+      "tokens_seen": 2323644416
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002987481945113144,
+      "loss": 0.0671,
+      "theoretical_loss": 3.416816219941571,
+      "tokens_seen": 2323906560
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.000298667950569732,
+      "loss": 0.066,
+      "theoretical_loss": 3.4167852862737895,
+      "tokens_seen": 2324168704
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002985877066281496,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4167543570716337,
+      "tokens_seen": 2324430848
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00029850746268656717,
+      "loss": 0.0661,
+      "theoretical_loss": 3.416723432333956,
+      "tokens_seen": 2324692992
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002984272187449847,
+      "loss": 0.0661,
+      "theoretical_loss": 3.416692512059607,
+      "tokens_seen": 2324955136
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00029834697480340234,
+      "loss": 0.065,
+      "theoretical_loss": 3.416661596247441,
+      "tokens_seen": 2325217280
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00029826673086181995,
+      "loss": 0.0692,
+      "theoretical_loss": 3.41663068489631,
+      "tokens_seen": 2325479424
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002981864869202375,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4165997780050685,
+      "tokens_seen": 2325741568
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.0002981062429786551,
+      "loss": 0.0675,
+      "theoretical_loss": 3.41656887557257,
+      "tokens_seen": 2326003712
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 0.00029802599903707273,
+      "loss": 0.0687,
+      "theoretical_loss": 3.416537977597668,
+      "tokens_seen": 2326265856
+    },
+    {
+      "epoch": 0.7,
+      "objective/train/advantage_avg": 0.001136346603743732,
+      "objective/train/docs_used": 847012,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3613319396972656,
+      "objective/train/original_loss": 1.3613317012786865,
+      "objective/train/theoretical_loss": 3.4165070840792184,
+      "objective/train/tokens_used": 2346988000,
+      "objective/train/value_avg": -0.006256103515625,
+      "objective/train/value_loss": 0.00011094258661614731,
+      "objective/train/value_max": -2.9802322387695312e-05,
+      "objective/train/value_min": -0.837890625,
+      "objective/train/value_reward_corr": 0.7206140669800114,
+      "objective/train/value_std": 0.01192474365234375,
+      "objective/train/weight_avg": 1.0011875629425049,
+      "objective/train/weighted_lm_loss": 1.3630400896072388,
+      "objective/train/weights_max": 1.1782879829406738,
+      "objective/train/weights_min": 0.373849481344223,
+      "theoretical_loss": 3.4165070840792184,
+      "tokens_seen": 2326528000
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029794575509549035,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4165070840792184,
+      "tokens_seen": 2326528000
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029786551115390785,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4164761950160765,
+      "tokens_seen": 2326790144
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029778526721232546,
+      "loss": 0.066,
+      "theoretical_loss": 3.4164453104070973,
+      "tokens_seen": 2327052288
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002977050232707431,
+      "loss": 0.0644,
+      "theoretical_loss": 3.416414430251137,
+      "tokens_seen": 2327314432
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029762477932916063,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4163835545470524,
+      "tokens_seen": 2327576576
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029754453538757825,
+      "loss": 0.0615,
+      "theoretical_loss": 3.4163526832937,
+      "tokens_seen": 2327838720
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029746429144599586,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4163218164899374,
+      "tokens_seen": 2328100864
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029738404750441347,
+      "loss": 0.0663,
+      "theoretical_loss": 3.416290954134622,
+      "tokens_seen": 2328363008
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.000297303803562831,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4162600962266128,
+      "tokens_seen": 2328625152
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002972235596212486,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4162292427647674,
+      "tokens_seen": 2328887296
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002971433156796662,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4161983937479454,
+      "tokens_seen": 2329149440
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029706307173808376,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4161675491750056,
+      "tokens_seen": 2329411584
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002969828277965014,
+      "loss": 0.0684,
+      "theoretical_loss": 3.416136709044809,
+      "tokens_seen": 2329673728
+    },
+    {
+      "epoch": 0.71,
+      "objective/train/advantage_avg": 0.0002161753218388185,
+      "objective/train/docs_used": 848291,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2818440198898315,
+      "objective/train/original_loss": 1.281843900680542,
+      "objective/train/theoretical_loss": 3.416121290645383,
+      "objective/train/tokens_used": 2350264800,
+      "objective/train/value_avg": -0.007343292236328125,
+      "objective/train/value_loss": 0.00019503793737385422,
+      "objective/train/value_max": -6.812810897827148e-05,
+      "objective/train/value_min": -0.22900390625,
+      "objective/train/value_reward_corr": 0.672381420425753,
+      "objective/train/value_std": 0.0130462646484375,
+      "objective/train/weight_avg": 1.0003025531768799,
+      "objective/train/weighted_lm_loss": 1.2819976806640625,
+      "objective/train/weights_max": 1.2044645547866821,
+      "objective/train/weights_min": 0.3711864948272705,
+      "theoretical_loss": 3.416121290645383,
+      "tokens_seen": 2329804800
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.000296902583854919,
+      "loss": 0.0651,
+      "theoretical_loss": 3.416105873356215,
+      "tokens_seen": 2329935872
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029682233991333654,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4160750421080843,
+      "tokens_seen": 2330198016
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002967420959717541,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4160442152992783,
+      "tokens_seen": 2330460160
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002966618520301717,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4160133929286585,
+      "tokens_seen": 2330722304
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029658160808858933,
+      "loss": 0.067,
+      "theoretical_loss": 3.4159825749950867,
+      "tokens_seen": 2330984448
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002965013641470069,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4159517614974257,
+      "tokens_seen": 2331246592
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002964211202054245,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4159209524345373,
+      "tokens_seen": 2331508736
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002963408762638421,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4158901478052863,
+      "tokens_seen": 2331770880
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029626063232225967,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4158593476085346,
+      "tokens_seen": 2332033024
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002961803883806773,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4158285518431475,
+      "tokens_seen": 2332295168
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029610014443909484,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4157977605079894,
+      "tokens_seen": 2332557312
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029601990049751245,
+      "loss": 0.0632,
+      "theoretical_loss": 3.415766973601924,
+      "tokens_seen": 2332819456
+    },
+    {
+      "epoch": 0.71,
+      "objective/train/advantage_avg": 0.000803662755060941,
+      "objective/train/docs_used": 849464,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3308948278427124,
+      "objective/train/original_loss": 1.3308947086334229,
+      "objective/train/theoretical_loss": 3.415736191123818,
+      "objective/train/tokens_used": 2353541600,
+      "objective/train/value_avg": -0.00688934326171875,
+      "objective/train/value_loss": 0.0001628411264391616,
+      "objective/train/value_max": -6.0617923736572266e-05,
+      "objective/train/value_min": -0.380859375,
+      "objective/train/value_reward_corr": 0.7531091940899427,
+      "objective/train/value_std": 0.0137786865234375,
+      "objective/train/weight_avg": 1.000876784324646,
+      "objective/train/weighted_lm_loss": 1.3320674896240234,
+      "objective/train/weights_max": 1.4635417461395264,
+      "objective/train/weights_min": 0.37232381105422974,
+      "theoretical_loss": 3.415736191123818,
+      "tokens_seen": 2333081600
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029593965655593,
+      "loss": 0.0665,
+      "theoretical_loss": 3.415736191123818,
+      "tokens_seen": 2333081600
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002958594126143476,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4157054130725366,
+      "tokens_seen": 2333343744
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029577916867276524,
+      "loss": 0.0686,
+      "theoretical_loss": 3.415674639446946,
+      "tokens_seen": 2333605888
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002956989247311828,
+      "loss": 0.0675,
+      "theoretical_loss": 3.415643870245913,
+      "tokens_seen": 2333868032
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002956186807896004,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4156131054683034,
+      "tokens_seen": 2334130176
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029553843684801797,
+      "loss": 0.0681,
+      "theoretical_loss": 3.415582345112986,
+      "tokens_seen": 2334392320
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002954581929064356,
+      "loss": 0.0716,
+      "theoretical_loss": 3.415551589178828,
+      "tokens_seen": 2334654464
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029537794896485314,
+      "loss": 0.0679,
+      "theoretical_loss": 3.415520837664698,
+      "tokens_seen": 2334916608
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029529770502327075,
+      "loss": 0.0658,
+      "theoretical_loss": 3.415490090569464,
+      "tokens_seen": 2335178752
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029521746108168836,
+      "loss": 0.0685,
+      "theoretical_loss": 3.415459347891996,
+      "tokens_seen": 2335440896
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002951372171401059,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4154286096311623,
+      "tokens_seen": 2335703040
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029505697319852353,
+      "loss": 0.0645,
+      "theoretical_loss": 3.4153978757858336,
+      "tokens_seen": 2335965184
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029497672925694115,
+      "loss": 0.066,
+      "theoretical_loss": 3.41536714635488,
+      "tokens_seen": 2336227328
+    },
+    {
+      "epoch": 0.71,
+      "objective/train/advantage_avg": 0.001232532667927444,
+      "objective/train/docs_used": 850768,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.270790457725525,
+      "objective/train/original_loss": 1.2707905769348145,
+      "objective/train/theoretical_loss": 3.415351783294441,
+      "objective/train/tokens_used": 2356818400,
+      "objective/train/value_avg": -0.005054473876953125,
+      "objective/train/value_loss": 0.0001378721499349922,
+      "objective/train/value_max": -2.586841583251953e-05,
+      "objective/train/value_min": -0.1971435546875,
+      "objective/train/value_reward_corr": 0.6732705521244904,
+      "objective/train/value_std": 0.0104217529296875,
+      "objective/train/weight_avg": 1.0012966394424438,
+      "objective/train/weighted_lm_loss": 1.272763967514038,
+      "objective/train/weights_max": 1.1968432664871216,
+      "objective/train/weights_min": 0.36855649948120117,
+      "theoretical_loss": 3.415351783294441,
+      "tokens_seen": 2336358400
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029489648531535865,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4153364213371726,
+      "tokens_seen": 2336489472
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029481624137377626,
+      "loss": 0.0655,
+      "theoretical_loss": 3.415305700731582,
+      "tokens_seen": 2336751616
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002947359974321939,
+      "loss": 0.0683,
+      "theoretical_loss": 3.41527498453698,
+      "tokens_seen": 2337013760
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002946557534906115,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4152442727522385,
+      "tokens_seen": 2337275904
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029457550954902905,
+      "loss": 0.0696,
+      "theoretical_loss": 3.41521356537623,
+      "tokens_seen": 2337538048
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029449526560744666,
+      "loss": 0.0668,
+      "theoretical_loss": 3.415182862407827,
+      "tokens_seen": 2337800192
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029441502166586427,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4151521638459035,
+      "tokens_seen": 2338062336
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002943347777242818,
+      "loss": 0.0697,
+      "theoretical_loss": 3.4151214696893324,
+      "tokens_seen": 2338324480
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002942545337826994,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4150907799369876,
+      "tokens_seen": 2338586624
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.000294174289841117,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4150600945877443,
+      "tokens_seen": 2338848768
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002940940458995346,
+      "loss": 0.069,
+      "theoretical_loss": 3.415029413640477,
+      "tokens_seen": 2339110912
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029401380195795217,
+      "loss": 0.0671,
+      "theoretical_loss": 3.414998737094061,
+      "tokens_seen": 2339373056
+    },
+    {
+      "epoch": 0.71,
+      "objective/train/advantage_avg": -0.00019418902229517698,
+      "objective/train/docs_used": 851815,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.414489984512329,
+      "objective/train/original_loss": 1.414489984512329,
+      "objective/train/theoretical_loss": 3.414968064947372,
+      "objective/train/tokens_used": 2360095200,
+      "objective/train/value_avg": -0.00991058349609375,
+      "objective/train/value_loss": 0.0003953047562390566,
+      "objective/train/value_max": -0.00011324882507324219,
+      "objective/train/value_min": -0.529296875,
+      "objective/train/value_reward_corr": 0.6825710898331492,
+      "objective/train/value_std": 0.0172119140625,
+      "objective/train/weight_avg": 0.9999842047691345,
+      "objective/train/weighted_lm_loss": 1.4128410816192627,
+      "objective/train/weights_max": 1.31961989402771,
+      "objective/train/weights_min": 0.37487199902534485,
+      "theoretical_loss": 3.414968064947372,
+      "tokens_seen": 2339635200
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002939335580163698,
+      "loss": 0.0673,
+      "theoretical_loss": 3.414968064947372,
+      "tokens_seen": 2339635200
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002938533140747874,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4149373971992856,
+      "tokens_seen": 2339897344
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002937730701332049,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4149067338486794,
+      "tokens_seen": 2340159488
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002936928261916225,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4148760748944293,
+      "tokens_seen": 2340421632
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002936125822500401,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4148454203354133,
+      "tokens_seen": 2340683776
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029353233830845774,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4148147701705094,
+      "tokens_seen": 2340945920
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002934520943668753,
+      "loss": 0.0669,
+      "theoretical_loss": 3.414784124398595,
+      "tokens_seen": 2341208064
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002933718504252929,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4147534830185493,
+      "tokens_seen": 2341470208
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002932916064837105,
+      "loss": 0.0674,
+      "theoretical_loss": 3.414722846029251,
+      "tokens_seen": 2341732352
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002932113625421281,
+      "loss": 0.0695,
+      "theoretical_loss": 3.4146922134295794,
+      "tokens_seen": 2341994496
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029313111860054564,
+      "loss": 0.0685,
+      "theoretical_loss": 3.414661585218415,
+      "tokens_seen": 2342256640
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029305087465896325,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4146309613946366,
+      "tokens_seen": 2342518784
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002929706307173808,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4146003419571267,
+      "tokens_seen": 2342780928
+    },
+    {
+      "epoch": 0.71,
+      "objective/train/advantage_avg": 0.00029292894760146737,
+      "objective/train/docs_used": 852909,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2010201215744019,
+      "objective/train/original_loss": 1.2010200023651123,
+      "objective/train/theoretical_loss": 3.4145850338828723,
+      "objective/train/tokens_used": 2363372000,
+      "objective/train/value_avg": -0.007549285888671875,
+      "objective/train/value_loss": 0.0002901647530961782,
+      "objective/train/value_max": -1.5437602996826172e-05,
+      "objective/train/value_min": -0.414306640625,
+      "objective/train/value_reward_corr": 0.6223776772873889,
+      "objective/train/value_std": 0.01342010498046875,
+      "objective/train/weight_avg": 1.000418782234192,
+      "objective/train/weighted_lm_loss": 1.2023299932479858,
+      "objective/train/weights_max": 1.4874529838562012,
+      "objective/train/weights_min": 0.40334662795066833,
+      "theoretical_loss": 3.4145850338828723,
+      "tokens_seen": 2342912000
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002928903867757984,
+      "loss": 0.0682,
+      "theoretical_loss": 3.414569726904765,
+      "tokens_seen": 2343043072
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029281014283421604,
+      "loss": 0.0673,
+      "theoretical_loss": 3.414539116236434,
+      "tokens_seen": 2343305216
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029272989889263365,
+      "loss": 0.066,
+      "theoretical_loss": 3.4145085099510144,
+      "tokens_seen": 2343567360
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002926496549510512,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4144779080473895,
+      "tokens_seen": 2343829504
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029256941100946877,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4144473105244413,
+      "tokens_seen": 2344091648
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002924891670678864,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4144167173810533,
+      "tokens_seen": 2344353792
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029240892312630394,
+      "loss": 0.068,
+      "theoretical_loss": 3.414386128616109,
+      "tokens_seen": 2344615936
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029232867918472155,
+      "loss": 0.0685,
+      "theoretical_loss": 3.414355544228492,
+      "tokens_seen": 2344878080
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029224843524313916,
+      "loss": 0.0697,
+      "theoretical_loss": 3.414324964217087,
+      "tokens_seen": 2345140224
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002921681913015568,
+      "loss": 0.0658,
+      "theoretical_loss": 3.414294388580779,
+      "tokens_seen": 2345402368
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029208794735997433,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4142638173184525,
+      "tokens_seen": 2345664512
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002920077034183919,
+      "loss": 0.066,
+      "theoretical_loss": 3.4142332504289934,
+      "tokens_seen": 2345926656
+    },
+    {
+      "epoch": 0.71,
+      "objective/train/advantage_avg": 0.0005077946116216481,
+      "objective/train/docs_used": 854185,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4195713996887207,
+      "objective/train/original_loss": 1.4195713996887207,
+      "objective/train/theoretical_loss": 3.414202687911288,
+      "objective/train/tokens_used": 2366648800,
+      "objective/train/value_avg": -0.0115814208984375,
+      "objective/train/value_loss": 0.0005665815551765263,
+      "objective/train/value_max": -5.02467155456543e-05,
+      "objective/train/value_min": -0.97509765625,
+      "objective/train/value_reward_corr": 0.7525287514571477,
+      "objective/train/value_std": 0.0290374755859375,
+      "objective/train/weight_avg": 1.0007684230804443,
+      "objective/train/weighted_lm_loss": 1.4191945791244507,
+      "objective/train/weights_max": 1.751630187034607,
+      "objective/train/weights_min": 0.3687090575695038,
+      "theoretical_loss": 3.414202687911288,
+      "tokens_seen": 2346188800
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002919274594768095,
+      "loss": 0.0681,
+      "theoretical_loss": 3.414202687911288,
+      "tokens_seen": 2346188800
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029184721553522706,
+      "loss": 0.0672,
+      "theoretical_loss": 3.414172129764222,
+      "tokens_seen": 2346450944
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002917669715936447,
+      "loss": 0.0681,
+      "theoretical_loss": 3.414141575986682,
+      "tokens_seen": 2346713088
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002916867276520623,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4141110265775563,
+      "tokens_seen": 2346975232
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002916064837104799,
+      "loss": 0.0664,
+      "theoretical_loss": 3.414080481535732,
+      "tokens_seen": 2347237376
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029152623976889746,
+      "loss": 0.0644,
+      "theoretical_loss": 3.414049940860097,
+      "tokens_seen": 2347499520
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029144599582731507,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4140194045495393,
+      "tokens_seen": 2347761664
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029136575188573263,
+      "loss": 0.0661,
+      "theoretical_loss": 3.413988872602949,
+      "tokens_seen": 2348023808
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002912855079441502,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4139583450192137,
+      "tokens_seen": 2348285952
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002912052640025678,
+      "loss": 0.0655,
+      "theoretical_loss": 3.413927821797224,
+      "tokens_seen": 2348548096
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002911250200609854,
+      "loss": 0.0681,
+      "theoretical_loss": 3.41389730293587,
+      "tokens_seen": 2348810240
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029104477611940297,
+      "loss": 0.065,
+      "theoretical_loss": 3.413866788434042,
+      "tokens_seen": 2349072384
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002909645321778206,
+      "loss": 0.0678,
+      "theoretical_loss": 3.413836278290631,
+      "tokens_seen": 2349334528
+    },
+    {
+      "epoch": 0.71,
+      "objective/train/advantage_avg": 0.0006469209911301732,
+      "objective/train/docs_used": 855374,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4136708974838257,
+      "objective/train/original_loss": 1.4136707782745361,
+      "objective/train/theoretical_loss": 3.413821024852985,
+      "objective/train/tokens_used": 2369925600,
+      "objective/train/value_avg": -0.0093231201171875,
+      "objective/train/value_loss": 0.0005805552937090397,
+      "objective/train/value_max": -1.722574234008789e-05,
+      "objective/train/value_min": -0.9482421875,
+      "objective/train/value_reward_corr": 0.6200741516231335,
+      "objective/train/value_std": 0.0212554931640625,
+      "objective/train/weight_avg": 1.0009031295776367,
+      "objective/train/weighted_lm_loss": 1.4141875505447388,
+      "objective/train/weights_max": 1.8178532123565674,
+      "objective/train/weights_min": 0.3728525638580322,
+      "theoretical_loss": 3.413821024852985,
+      "tokens_seen": 2349465600
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002908842882362382,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4138057725045274,
+      "tokens_seen": 2349596672
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029080404429465576,
+      "loss": 0.0698,
+      "theoretical_loss": 3.413775271074624,
+      "tokens_seen": 2349858816
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002907238003530733,
+      "loss": 0.0671,
+      "theoretical_loss": 3.413744773999812,
+      "tokens_seen": 2350120960
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002906435564114909,
+      "loss": 0.0654,
+      "theoretical_loss": 3.413714281278985,
+      "tokens_seen": 2350383104
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029056331246990854,
+      "loss": 0.0674,
+      "theoretical_loss": 3.413683792911035,
+      "tokens_seen": 2350645248
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002904830685283261,
+      "loss": 0.067,
+      "theoretical_loss": 3.413653308894855,
+      "tokens_seen": 2350907392
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002904028245867437,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4136228292293396,
+      "tokens_seen": 2351169536
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002903225806451613,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4135923539133826,
+      "tokens_seen": 2351431680
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029024233670357894,
+      "loss": 0.067,
+      "theoretical_loss": 3.413561882945878,
+      "tokens_seen": 2351693824
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029016209276199644,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4135314163257213,
+      "tokens_seen": 2351955968
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029008184882041405,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4135009540518073,
+      "tokens_seen": 2352218112
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00029000160487883166,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4134704961230327,
+      "tokens_seen": 2352480256
+    },
+    {
+      "epoch": 0.71,
+      "objective/train/advantage_avg": 0.0007358468137681484,
+      "objective/train/docs_used": 856640,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4628479480743408,
+      "objective/train/original_loss": 1.4628478288650513,
+      "objective/train/theoretical_loss": 3.4134400425382925,
+      "objective/train/tokens_used": 2373202400,
+      "objective/train/value_avg": -0.010711669921875,
+      "objective/train/value_loss": 0.00020279559248592705,
+      "objective/train/value_max": -1.7762184143066406e-05,
+      "objective/train/value_min": -0.291259765625,
+      "objective/train/value_reward_corr": 0.7869439892749879,
+      "objective/train/value_std": 0.0175018310546875,
+      "objective/train/weight_avg": 1.0008323192596436,
+      "objective/train/weighted_lm_loss": 1.4642149209976196,
+      "objective/train/weights_max": 1.1599798202514648,
+      "objective/train/weights_min": 0.3715236186981201,
+      "theoretical_loss": 3.4134400425382925,
+      "tokens_seen": 2352742400
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002899213609372492,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4134400425382925,
+      "tokens_seen": 2352742400
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00028984111699566684,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4134095932964836,
+      "tokens_seen": 2353004544
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00028976087305408445,
+      "loss": 0.0681,
+      "theoretical_loss": 3.413379148396503,
+      "tokens_seen": 2353266688
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00028968062911250206,
+      "loss": 0.0655,
+      "theoretical_loss": 3.413348707837248,
+      "tokens_seen": 2353528832
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00028960038517091957,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4133182716176167,
+      "tokens_seen": 2353790976
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002895201412293372,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4132878397365065,
+      "tokens_seen": 2354053120
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002894398972877548,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4132574121928165,
+      "tokens_seen": 2354315264
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00028935965334617235,
+      "loss": 0.0661,
+      "theoretical_loss": 3.413226988985446,
+      "tokens_seen": 2354577408
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00028927940940458996,
+      "loss": 0.065,
+      "theoretical_loss": 3.4131965701132927,
+      "tokens_seen": 2354839552
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002891991654630076,
+      "loss": 0.0641,
+      "theoretical_loss": 3.4131661555752584,
+      "tokens_seen": 2355101696
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002891189215214252,
+      "loss": 0.0663,
+      "theoretical_loss": 3.413135745370242,
+      "tokens_seen": 2355363840
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002890386775798427,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4131053394971445,
+      "tokens_seen": 2355625984
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002889584336382603,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4130749379548666,
+      "tokens_seen": 2355888128
+    },
+    {
+      "epoch": 0.71,
+      "objective/train/advantage_avg": -0.0010917705949395895,
+      "objective/train/docs_used": 857700,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3285942077636719,
+      "objective/train/original_loss": 1.3285940885543823,
+      "objective/train/theoretical_loss": 3.4130597388074415,
+      "objective/train/tokens_used": 2376479200,
+      "objective/train/value_avg": -0.014007568359375,
+      "objective/train/value_loss": 0.00037045456701889634,
+      "objective/train/value_max": -5.739927291870117e-05,
+      "objective/train/value_min": -0.51806640625,
+      "objective/train/value_reward_corr": 0.8645235383798338,
+      "objective/train/value_std": 0.0283355712890625,
+      "objective/train/weight_avg": 0.999081552028656,
+      "objective/train/weighted_lm_loss": 1.326906442642212,
+      "objective/train/weights_max": 1.4649479389190674,
+      "objective/train/weights_min": 0.37004411220550537,
+      "theoretical_loss": 3.4130597388074415,
+      "tokens_seen": 2356019200
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002888781896966779,
+      "loss": 0.067,
+      "theoretical_loss": 3.41304454074231,
+      "tokens_seen": 2356150272
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002887979457550955,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4130141478583758,
+      "tokens_seen": 2356412416
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002887177018135131,
+      "loss": 0.066,
+      "theoretical_loss": 3.4129837593019667,
+      "tokens_seen": 2356674560
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002886374578719307,
+      "loss": 0.0672,
+      "theoretical_loss": 3.412953375071985,
+      "tokens_seen": 2356936704
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00028855721393034826,
+      "loss": 0.0662,
+      "theoretical_loss": 3.412922995167334,
+      "tokens_seen": 2357198848
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00028847696998876587,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4128926195869163,
+      "tokens_seen": 2357460992
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00028839672604718343,
+      "loss": 0.0634,
+      "theoretical_loss": 3.4128622483296365,
+      "tokens_seen": 2357723136
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.00028831648210560104,
+      "loss": 0.0674,
+      "theoretical_loss": 3.412831881394399,
+      "tokens_seen": 2357985280
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002882362381640186,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4128015187801073,
+      "tokens_seen": 2358247424
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002881559942224362,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4127711604856668,
+      "tokens_seen": 2358509568
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002880757502808538,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4127408065099827,
+      "tokens_seen": 2358771712
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.0002879955063392714,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4127104568519613,
+      "tokens_seen": 2359033856
+    },
+    {
+      "epoch": 0.71,
+      "objective/train/advantage_avg": -9.175229934044182e-06,
+      "objective/train/docs_used": 858941,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2591488361358643,
+      "objective/train/original_loss": 1.2591485977172852,
+      "objective/train/theoretical_loss": 3.412680111510508,
+      "objective/train/tokens_used": 2379756000,
+      "objective/train/value_avg": -0.0064544677734375,
+      "objective/train/value_loss": 0.00019788608187809587,
+      "objective/train/value_max": -3.510713577270508e-05,
+      "objective/train/value_min": -0.3271484375,
+      "objective/train/value_reward_corr": 0.6843129270556736,
+      "objective/train/value_std": 0.0122528076171875,
+      "objective/train/weight_avg": 1.0000793933868408,
+      "objective/train/weighted_lm_loss": 1.2583329677581787,
+      "objective/train/weights_max": 1.2182163000106812,
+      "objective/train/weights_min": 0.3679918348789215,
+      "theoretical_loss": 3.412680111510508,
+      "tokens_seen": 2359296000
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 0.000287915262397689,
+      "loss": 0.0645,
+      "theoretical_loss": 3.412680111510508,
+      "tokens_seen": 2359296000
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028783501845610655,
+      "loss": 0.0677,
+      "theoretical_loss": 3.41264977048453,
+      "tokens_seen": 2359558144
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028775477451452417,
+      "loss": 0.0692,
+      "theoretical_loss": 3.412619433772934,
+      "tokens_seen": 2359820288
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002876745305729417,
+      "loss": 0.0677,
+      "theoretical_loss": 3.412589101374627,
+      "tokens_seen": 2360082432
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028759428663135934,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4125587732885174,
+      "tokens_seen": 2360344576
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028751404268977695,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4125284495135126,
+      "tokens_seen": 2360606720
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002874337987481945,
+      "loss": 0.0668,
+      "theoretical_loss": 3.412498130048522,
+      "tokens_seen": 2360868864
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002873535548066121,
+      "loss": 0.0681,
+      "theoretical_loss": 3.412467814892454,
+      "tokens_seen": 2361131008
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002872733108650297,
+      "loss": 0.066,
+      "theoretical_loss": 3.4124375040442176,
+      "tokens_seen": 2361393152
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002871930669234473,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4124071975027235,
+      "tokens_seen": 2361655296
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028711282298186485,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4123768952668807,
+      "tokens_seen": 2361917440
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028703257904028246,
+      "loss": 0.0678,
+      "theoretical_loss": 3.412346597335601,
+      "tokens_seen": 2362179584
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002869523350987001,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4123163037077937,
+      "tokens_seen": 2362441728
+    },
+    {
+      "epoch": 0.72,
+      "objective/train/advantage_avg": 0.0010815730784088373,
+      "objective/train/docs_used": 860261,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.416516900062561,
+      "objective/train/original_loss": 1.4165170192718506,
+      "objective/train/theoretical_loss": 3.4123011585073524,
+      "objective/train/tokens_used": 2383032800,
+      "objective/train/value_avg": -0.005901336669921875,
+      "objective/train/value_loss": 8.771286957198754e-05,
+      "objective/train/value_max": -2.9802322387695312e-05,
+      "objective/train/value_min": -0.2247314453125,
+      "objective/train/value_reward_corr": 0.7080153186998583,
+      "objective/train/value_std": 0.01003265380859375,
+      "objective/train/weight_avg": 1.0011248588562012,
+      "objective/train/weighted_lm_loss": 1.4178658723831177,
+      "objective/train/weights_max": 1.1204938888549805,
+      "objective/train/weights_min": 0.7841757535934448,
+      "theoretical_loss": 3.4123011585073524,
+      "tokens_seen": 2362572800
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028687209115711763,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4122860143823717,
+      "tokens_seen": 2362703872
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028679184721553525,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4122557293582454,
+      "tokens_seen": 2362966016
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028671160327395286,
+      "loss": 0.0662,
+      "theoretical_loss": 3.412225448634328,
+      "tokens_seen": 2363228160
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028663135933237036,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4121951722095316,
+      "tokens_seen": 2363490304
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.000286551115390788,
+      "loss": 0.0685,
+      "theoretical_loss": 3.412164900082769,
+      "tokens_seen": 2363752448
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002864708714492056,
+      "loss": 0.065,
+      "theoretical_loss": 3.412134632252953,
+      "tokens_seen": 2364014592
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002863906275076232,
+      "loss": 0.0641,
+      "theoretical_loss": 3.4121043687189982,
+      "tokens_seen": 2364276736
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028631038356604076,
+      "loss": 0.0693,
+      "theoretical_loss": 3.4120741094798186,
+      "tokens_seen": 2364538880
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002862301396244584,
+      "loss": 0.0657,
+      "theoretical_loss": 3.412043854534328,
+      "tokens_seen": 2364801024
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.000286149895682876,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4120136038814417,
+      "tokens_seen": 2365063168
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002860696517412935,
+      "loss": 0.0674,
+      "theoretical_loss": 3.411983357520075,
+      "tokens_seen": 2365325312
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002859894077997111,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4119531154491436,
+      "tokens_seen": 2365587456
+    },
+    {
+      "epoch": 0.72,
+      "objective/train/advantage_avg": 0.0008249745587818325,
+      "objective/train/docs_used": 861421,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1259411573410034,
+      "objective/train/original_loss": 1.1259410381317139,
+      "objective/train/theoretical_loss": 3.411922877667563,
+      "objective/train/tokens_used": 2386309600,
+      "objective/train/value_avg": -0.00743865966796875,
+      "objective/train/value_loss": 0.00017807947006076574,
+      "objective/train/value_max": -4.869699478149414e-05,
+      "objective/train/value_min": -0.32421875,
+      "objective/train/value_reward_corr": 0.6612214544166373,
+      "objective/train/value_std": 0.01175689697265625,
+      "objective/train/weight_avg": 1.0009032487869263,
+      "objective/train/weighted_lm_loss": 1.1269689798355103,
+      "objective/train/weights_max": 1.2590364217758179,
+      "objective/train/weights_min": 0.3725539743900299,
+      "theoretical_loss": 3.411922877667563,
+      "tokens_seen": 2365849600
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002859091638581287,
+      "loss": 0.0651,
+      "theoretical_loss": 3.411922877667563,
+      "tokens_seen": 2365849600
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028582891991654633,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4118926441742508,
+      "tokens_seen": 2366111744
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002857486759749639,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4118624149681227,
+      "tokens_seen": 2366373888
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002856684320333815,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4118321900480963,
+      "tokens_seen": 2366636032
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002855881880917991,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4118019694130894,
+      "tokens_seen": 2366898176
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028550794415021667,
+      "loss": 0.0692,
+      "theoretical_loss": 3.41177175306202,
+      "tokens_seen": 2367160320
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028542770020863423,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4117415409938068,
+      "tokens_seen": 2367422464
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028534745626705184,
+      "loss": 0.0667,
+      "theoretical_loss": 3.411711333207368,
+      "tokens_seen": 2367684608
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028526721232546945,
+      "loss": 0.0657,
+      "theoretical_loss": 3.411681129701623,
+      "tokens_seen": 2367946752
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.000285186968383887,
+      "loss": 0.0648,
+      "theoretical_loss": 3.411650930475491,
+      "tokens_seen": 2368208896
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002851067244423046,
+      "loss": 0.066,
+      "theoretical_loss": 3.4116207355278934,
+      "tokens_seen": 2368471040
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028502648050072224,
+      "loss": 0.0668,
+      "theoretical_loss": 3.411590544857749,
+      "tokens_seen": 2368733184
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002849462365591398,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4115603584639795,
+      "tokens_seen": 2368995328
+    },
+    {
+      "epoch": 0.72,
+      "objective/train/advantage_avg": 0.0003487260837573558,
+      "objective/train/docs_used": 862685,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2241593599319458,
+      "objective/train/original_loss": 1.2241593599319458,
+      "objective/train/theoretical_loss": 3.411545266870398,
+      "objective/train/tokens_used": 2389586400,
+      "objective/train/value_avg": -0.00579833984375,
+      "objective/train/value_loss": 0.00012634368613362312,
+      "objective/train/value_max": -2.9325485229492188e-05,
+      "objective/train/value_min": -0.55712890625,
+      "objective/train/value_reward_corr": 0.6420565850004247,
+      "objective/train/value_std": 0.00954437255859375,
+      "objective/train/weight_avg": 1.0004075765609741,
+      "objective/train/weighted_lm_loss": 1.2246220111846924,
+      "objective/train/weights_max": 1.3289666175842285,
+      "objective/train/weights_min": 0.371211975812912,
+      "theoretical_loss": 3.411545266870398,
+      "tokens_seen": 2369126400
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028486599261755735,
+      "loss": 0.065,
+      "theoretical_loss": 3.4115301763455057,
+      "tokens_seen": 2369257472
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028478574867597497,
+      "loss": 0.0666,
+      "theoretical_loss": 3.411499998501249,
+      "tokens_seen": 2369519616
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002847055047343925,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4114698249301316,
+      "tokens_seen": 2369781760
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028462526079281014,
+      "loss": 0.0657,
+      "theoretical_loss": 3.411439655631076,
+      "tokens_seen": 2370043904
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028454501685122775,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4114094906030052,
+      "tokens_seen": 2370306048
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028446477290964536,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4113793298448414,
+      "tokens_seen": 2370568192
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002843845289680629,
+      "loss": 0.0681,
+      "theoretical_loss": 3.411349173355508,
+      "tokens_seen": 2370830336
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002843042850264805,
+      "loss": 0.0691,
+      "theoretical_loss": 3.4113190211339304,
+      "tokens_seen": 2371092480
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002842240410848981,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4112888731790316,
+      "tokens_seen": 2371354624
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028414379714331565,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4112587294897367,
+      "tokens_seen": 2371616768
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028406355320173326,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4112285900649706,
+      "tokens_seen": 2371878912
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002839833092601509,
+      "loss": 0.0649,
+      "theoretical_loss": 3.411198454903659,
+      "tokens_seen": 2372141056
+    },
+    {
+      "epoch": 0.72,
+      "objective/train/advantage_avg": 0.0007301963050849736,
+      "objective/train/docs_used": 863878,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3005666732788086,
+      "objective/train/original_loss": 1.3005667924880981,
+      "objective/train/theoretical_loss": 3.4111683240047275,
+      "objective/train/tokens_used": 2392863200,
+      "objective/train/value_avg": -0.00737762451171875,
+      "objective/train/value_loss": 0.0002043836284428835,
+      "objective/train/value_max": -3.218650817871094e-05,
+      "objective/train/value_min": -0.67529296875,
+      "objective/train/value_reward_corr": 0.7171774505277242,
+      "objective/train/value_std": 0.0146636962890625,
+      "objective/train/weight_avg": 1.0008224248886108,
+      "objective/train/weighted_lm_loss": 1.3015154600143433,
+      "objective/train/weights_max": 1.142139196395874,
+      "objective/train/weights_min": 0.37391793727874756,
+      "theoretical_loss": 3.4111683240047275,
+      "tokens_seen": 2372403200
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002839030653185685,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4111683240047275,
+      "tokens_seen": 2372403200
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028382282137698605,
+      "loss": 0.0626,
+      "theoretical_loss": 3.4111381973671024,
+      "tokens_seen": 2372665344
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028374257743540366,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4111080749897105,
+      "tokens_seen": 2372927488
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002836623334938212,
+      "loss": 0.067,
+      "theoretical_loss": 3.4110779568714786,
+      "tokens_seen": 2373189632
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002835820895522388,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4110478430113345,
+      "tokens_seen": 2373451776
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002835018456106564,
+      "loss": 0.0641,
+      "theoretical_loss": 3.4110177334082055,
+      "tokens_seen": 2373713920
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.000283421601669074,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4109876280610196,
+      "tokens_seen": 2373976064
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002833413577274916,
+      "loss": 0.0626,
+      "theoretical_loss": 3.4109575269687062,
+      "tokens_seen": 2374238208
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028326111378590917,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4109274301301937,
+      "tokens_seen": 2374500352
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002831808698443268,
+      "loss": 0.0647,
+      "theoretical_loss": 3.410897337544412,
+      "tokens_seen": 2374762496
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028310062590274434,
+      "loss": 0.0645,
+      "theoretical_loss": 3.41086724921029,
+      "tokens_seen": 2375024640
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002830203819611619,
+      "loss": 0.0669,
+      "theoretical_loss": 3.410837165126758,
+      "tokens_seen": 2375286784
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002829401380195795,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4108070852927472,
+      "tokens_seen": 2375548928
+    },
+    {
+      "epoch": 0.72,
+      "objective/train/advantage_avg": 6.751207547495142e-05,
+      "objective/train/docs_used": 865150,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3704978227615356,
+      "objective/train/original_loss": 1.370497703552246,
+      "objective/train/theoretical_loss": 3.410792046968978,
+      "objective/train/tokens_used": 2396140000,
+      "objective/train/value_avg": -0.00868988037109375,
+      "objective/train/value_loss": 0.00020805127860512584,
+      "objective/train/value_max": -3.7610530853271484e-05,
+      "objective/train/value_min": -0.970703125,
+      "objective/train/value_reward_corr": 0.7735973443730089,
+      "objective/train/value_std": 0.016876220703125,
+      "objective/train/weight_avg": 1.0001626014709473,
+      "objective/train/weighted_lm_loss": 1.3697094917297363,
+      "objective/train/weights_max": 1.2211940288543701,
+      "objective/train/weights_min": 0.3830130398273468,
+      "theoretical_loss": 3.410792046968978,
+      "tokens_seen": 2375680000
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028285989407799713,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4107770097071874,
+      "tokens_seen": 2375811072
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002827796501364147,
+      "loss": 0.0651,
+      "theoretical_loss": 3.410746938369011,
+      "tokens_seen": 2376073216
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002826994061948323,
+      "loss": 0.0681,
+      "theoretical_loss": 3.410716871277149,
+      "tokens_seen": 2376335360
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002826191622532499,
+      "loss": 0.065,
+      "theoretical_loss": 3.4106868084305337,
+      "tokens_seen": 2376597504
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028253891831166747,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4106567498280973,
+      "tokens_seen": 2376859648
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028245867437008503,
+      "loss": 0.0672,
+      "theoretical_loss": 3.410626695468773,
+      "tokens_seen": 2377121792
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028237843042850264,
+      "loss": 0.0653,
+      "theoretical_loss": 3.410596645351494,
+      "tokens_seen": 2377383936
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028229818648692025,
+      "loss": 0.065,
+      "theoretical_loss": 3.4105665994751933,
+      "tokens_seen": 2377646080
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002822179425453378,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4105365578388054,
+      "tokens_seen": 2377908224
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002821376986037554,
+      "loss": 0.0673,
+      "theoretical_loss": 3.410506520441265,
+      "tokens_seen": 2378170368
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028205745466217304,
+      "loss": 0.0662,
+      "theoretical_loss": 3.410476487281506,
+      "tokens_seen": 2378432512
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028197721072059065,
+      "loss": 0.065,
+      "theoretical_loss": 3.410446458358464,
+      "tokens_seen": 2378694656
+    },
+    {
+      "epoch": 0.72,
+      "objective/train/advantage_avg": 0.0003571182896848768,
+      "objective/train/docs_used": 866449,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.122450828552246,
+      "objective/train/original_loss": 1.122450828552246,
+      "objective/train/theoretical_loss": 3.4104164336710747,
+      "objective/train/tokens_used": 2399416800,
+      "objective/train/value_avg": -0.00919342041015625,
+      "objective/train/value_loss": 0.00014748242392670363,
+      "objective/train/value_max": -5.346536636352539e-05,
+      "objective/train/value_min": -0.291015625,
+      "objective/train/value_reward_corr": 0.7777066932130247,
+      "objective/train/value_std": 0.015655517578125,
+      "objective/train/weight_avg": 1.0004256963729858,
+      "objective/train/weighted_lm_loss": 1.1227384805679321,
+      "objective/train/weights_max": 1.2215328216552734,
+      "objective/train/weights_min": 0.40386396646499634,
+      "theoretical_loss": 3.4104164336710747,
+      "tokens_seen": 2378956800
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028189696677900815,
+      "loss": 0.065,
+      "theoretical_loss": 3.4104164336710747,
+      "tokens_seen": 2378956800
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028181672283742577,
+      "loss": 0.0648,
+      "theoretical_loss": 3.4103864132182733,
+      "tokens_seen": 2379218944
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002817364788958434,
+      "loss": 0.0667,
+      "theoretical_loss": 3.410356396998997,
+      "tokens_seen": 2379481088
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028165623495426094,
+      "loss": 0.0703,
+      "theoretical_loss": 3.410326385012182,
+      "tokens_seen": 2379743232
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028157599101267855,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4102963772567656,
+      "tokens_seen": 2380005376
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028149574707109616,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4102663737316847,
+      "tokens_seen": 2380267520
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002814155031295138,
+      "loss": 0.065,
+      "theoretical_loss": 3.410236374435878,
+      "tokens_seen": 2380529664
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002813352591879313,
+      "loss": 0.0653,
+      "theoretical_loss": 3.410206379368283,
+      "tokens_seen": 2380791808
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002812550152463489,
+      "loss": 0.066,
+      "theoretical_loss": 3.4101763885278387,
+      "tokens_seen": 2381053952
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002811747713047665,
+      "loss": 0.0675,
+      "theoretical_loss": 3.410146401913484,
+      "tokens_seen": 2381316096
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028109452736318406,
+      "loss": 0.0654,
+      "theoretical_loss": 3.410116419524158,
+      "tokens_seen": 2381578240
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002810142834216017,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4100864413588003,
+      "tokens_seen": 2381840384
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002809340394800193,
+      "loss": 0.0652,
+      "theoretical_loss": 3.410056467416352,
+      "tokens_seen": 2382102528
+    },
+    {
+      "epoch": 0.72,
+      "objective/train/advantage_avg": -0.000377790245693177,
+      "objective/train/docs_used": 867590,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3330551385879517,
+      "objective/train/original_loss": 1.3330551385879517,
+      "objective/train/theoretical_loss": 3.410041482028387,
+      "objective/train/tokens_used": 2402693600,
+      "objective/train/value_avg": -0.01041412353515625,
+      "objective/train/value_loss": 0.00034967754618264735,
+      "objective/train/value_max": -1.3828277587890625e-05,
+      "objective/train/value_min": -0.447998046875,
+      "objective/train/value_reward_corr": 0.7093849786916704,
+      "objective/train/value_std": 0.019927978515625,
+      "objective/train/weight_avg": 0.9997868537902832,
+      "objective/train/weighted_lm_loss": 1.3321644067764282,
+      "objective/train/weights_max": 1.4849282503128052,
+      "objective/train/weights_min": 0.376959890127182,
+      "theoretical_loss": 3.410041482028387,
+      "tokens_seen": 2382233600
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002808537955384369,
+      "loss": 0.067,
+      "theoretical_loss": 3.4100264976957524,
+      "tokens_seen": 2382364672
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028077355159685446,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4099965321959433,
+      "tokens_seen": 2382626816
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.000280693307655272,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4099665709158655,
+      "tokens_seen": 2382888960
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028061306371368963,
+      "loss": 0.0659,
+      "theoretical_loss": 3.409936613854461,
+      "tokens_seen": 2383151104
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002805328197721072,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4099066610106714,
+      "tokens_seen": 2383413248
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002804525758305248,
+      "loss": 0.0659,
+      "theoretical_loss": 3.40987671238344,
+      "tokens_seen": 2383675392
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002803723318889424,
+      "loss": 0.066,
+      "theoretical_loss": 3.4098467679717084,
+      "tokens_seen": 2383937536
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028029208794735997,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4098168277744203,
+      "tokens_seen": 2384199680
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002802118440057776,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4097868917905196,
+      "tokens_seen": 2384461824
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028013160006419514,
+      "loss": 0.0654,
+      "theoretical_loss": 3.40975696001895,
+      "tokens_seen": 2384723968
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00028005135612261276,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4097270324586555,
+      "tokens_seen": 2384986112
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002799711121810303,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4096971091085817,
+      "tokens_seen": 2385248256
+    },
+    {
+      "epoch": 0.72,
+      "objective/train/advantage_avg": 0.0015336308861151338,
+      "objective/train/docs_used": 868828,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.295317530632019,
+      "objective/train/original_loss": 1.29531729221344,
+      "objective/train/theoretical_loss": 3.4096671899676725,
+      "objective/train/tokens_used": 2405970400,
+      "objective/train/value_avg": -0.01094818115234375,
+      "objective/train/value_loss": 0.00022329458442982286,
+      "objective/train/value_max": -6.657838821411133e-05,
+      "objective/train/value_min": -0.353759765625,
+      "objective/train/value_reward_corr": 0.7366697275718553,
+      "objective/train/value_std": 0.017852783203125,
+      "objective/train/weight_avg": 1.0016379356384277,
+      "objective/train/weighted_lm_loss": 1.2963742017745972,
+      "objective/train/weights_max": 1.3134781122207642,
+      "objective/train/weights_min": 0.3717447817325592,
+      "theoretical_loss": 3.4096671899676725,
+      "tokens_seen": 2385510400
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002798908682394479,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4096671899676725,
+      "tokens_seen": 2385510400
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027981062429786554,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4096372750348745,
+      "tokens_seen": 2385772544
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002797303803562831,
+      "loss": 0.0667,
+      "theoretical_loss": 3.409607364309133,
+      "tokens_seen": 2386034688
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002796501364147007,
+      "loss": 0.0699,
+      "theoretical_loss": 3.4095774577893936,
+      "tokens_seen": 2386296832
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027956989247311827,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4095475554746044,
+      "tokens_seen": 2386558976
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002794896485315359,
+      "loss": 0.0678,
+      "theoretical_loss": 3.409517657363711,
+      "tokens_seen": 2386821120
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027940940458995344,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4094877634556617,
+      "tokens_seen": 2387083264
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027932916064837105,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4094578737494037,
+      "tokens_seen": 2387345408
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027924891670678866,
+      "loss": 0.068,
+      "theoretical_loss": 3.4094279882438854,
+      "tokens_seen": 2387607552
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002791686727652062,
+      "loss": 0.0636,
+      "theoretical_loss": 3.409398106938055,
+      "tokens_seen": 2387869696
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027908842882362384,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4093682298308616,
+      "tokens_seen": 2388131840
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027900818488204145,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4093383569212548,
+      "tokens_seen": 2388393984
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027892794094045895,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4093084882081834,
+      "tokens_seen": 2388656128
+    },
+    {
+      "epoch": 0.72,
+      "objective/train/advantage_avg": 9.959724411601201e-05,
+      "objective/train/docs_used": 870071,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3700393438339233,
+      "objective/train/original_loss": 1.370039463043213,
+      "objective/train/theoretical_loss": 3.409293555425021,
+      "objective/train/tokens_used": 2409247200,
+      "objective/train/value_avg": -0.00676727294921875,
+      "objective/train/value_loss": 0.00039645720971748233,
+      "objective/train/value_max": -6.866455078125e-05,
+      "objective/train/value_min": -0.9521484375,
+      "objective/train/value_reward_corr": 0.675174519670669,
+      "objective/train/value_std": 0.01494598388671875,
+      "objective/train/weight_avg": 1.000266432762146,
+      "objective/train/weighted_lm_loss": 1.3696532249450684,
+      "objective/train/weights_max": 1.2567331790924072,
+      "objective/train/weights_min": 0.3683270514011383,
+      "theoretical_loss": 3.409293555425021,
+      "tokens_seen": 2388787200
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027884769699887657,
+      "loss": 0.0718,
+      "theoretical_loss": 3.4092786236905983,
+      "tokens_seen": 2388918272
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002787674530572942,
+      "loss": 0.065,
+      "theoretical_loss": 3.4092487633674495,
+      "tokens_seen": 2389180416
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002786872091157118,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4092189072376877,
+      "tokens_seen": 2389442560
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027860696517412935,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4091890553002644,
+      "tokens_seen": 2389704704
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.00027852672123254696,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4091592075541306,
+      "tokens_seen": 2389966848
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002784464772909646,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4091293639982387,
+      "tokens_seen": 2390228992
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002783662333493821,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4090995246315403,
+      "tokens_seen": 2390491136
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002782859894077997,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4090696894529886,
+      "tokens_seen": 2390753280
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002782057454662173,
+      "loss": 0.0667,
+      "theoretical_loss": 3.409039858461537,
+      "tokens_seen": 2391015424
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002781255015246349,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4090100316561385,
+      "tokens_seen": 2391277568
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002780452575830525,
+      "loss": 0.0672,
+      "theoretical_loss": 3.408980209035746,
+      "tokens_seen": 2391539712
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002779650136414701,
+      "loss": 0.067,
+      "theoretical_loss": 3.4089503905993155,
+      "tokens_seen": 2391801856
+    },
+    {
+      "epoch": 0.72,
+      "objective/train/advantage_avg": -2.867977491405327e-05,
+      "objective/train/docs_used": 871244,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4574215412139893,
+      "objective/train/original_loss": 1.4574213027954102,
+      "objective/train/theoretical_loss": 3.4089205763458,
+      "objective/train/tokens_used": 2412524000,
+      "objective/train/value_avg": -0.00861358642578125,
+      "objective/train/value_loss": 0.00014166087203193456,
+      "objective/train/value_max": -4.5418739318847656e-05,
+      "objective/train/value_min": -0.31787109375,
+      "objective/train/value_reward_corr": 0.7858477079902386,
+      "objective/train/value_std": 0.0163726806640625,
+      "objective/train/weight_avg": 1.000041127204895,
+      "objective/train/weighted_lm_loss": 1.457660436630249,
+      "objective/train/weights_max": 1.1719762086868286,
+      "objective/train/weights_min": 0.7179513573646545,
+      "theoretical_loss": 3.4089205763458,
+      "tokens_seen": 2392064000
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002778847696998877,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4089205763458,
+      "tokens_seen": 2392064000
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 0.0002778045257583052,
+      "loss": 0.0676,
+      "theoretical_loss": 3.408890766274155,
+      "tokens_seen": 2392326144
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002777242818167228,
+      "loss": 0.067,
+      "theoretical_loss": 3.408860960383336,
+      "tokens_seen": 2392588288
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027764403787514043,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4088311586722986,
+      "tokens_seen": 2392850432
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027756379393355804,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4088013611399983,
+      "tokens_seen": 2393112576
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002774835499919756,
+      "loss": 0.0669,
+      "theoretical_loss": 3.408771567785392,
+      "tokens_seen": 2393374720
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002774033060503932,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4087417786074363,
+      "tokens_seen": 2393636864
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002773230621088108,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4087119936050887,
+      "tokens_seen": 2393899008
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002772428181672284,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4086822127773067,
+      "tokens_seen": 2394161152
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027716257422564594,
+      "loss": 0.0645,
+      "theoretical_loss": 3.4086524361230475,
+      "tokens_seen": 2394423296
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027708233028406355,
+      "loss": 0.0659,
+      "theoretical_loss": 3.40862266364127,
+      "tokens_seen": 2394685440
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027700208634248117,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4085928953309335,
+      "tokens_seen": 2394947584
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002769218424008987,
+      "loss": 0.0677,
+      "theoretical_loss": 3.408563131190996,
+      "tokens_seen": 2395209728
+    },
+    {
+      "epoch": 0.73,
+      "objective/train/advantage_avg": 0.00042437895899638534,
+      "objective/train/docs_used": 872384,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2476065158843994,
+      "objective/train/original_loss": 1.2476065158843994,
+      "objective/train/theoretical_loss": 3.408548250684601,
+      "objective/train/tokens_used": 2415800800,
+      "objective/train/value_avg": -0.006153106689453125,
+      "objective/train/value_loss": 0.00014658182044513524,
+      "objective/train/value_max": -3.4809112548828125e-05,
+      "objective/train/value_min": -0.2880859375,
+      "objective/train/value_reward_corr": 0.5964869899054704,
+      "objective/train/value_std": 0.0099029541015625,
+      "objective/train/weight_avg": 1.0004934072494507,
+      "objective/train/weighted_lm_loss": 1.2491647005081177,
+      "objective/train/weights_max": 1.2718428373336792,
+      "objective/train/weights_min": 0.37988168001174927,
+      "theoretical_loss": 3.408548250684601,
+      "tokens_seen": 2395340800
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027684159845931634,
+      "loss": 0.0662,
+      "theoretical_loss": 3.408533371220417,
+      "tokens_seen": 2395471872
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027676135451773395,
+      "loss": 0.0664,
+      "theoretical_loss": 3.408503615418156,
+      "tokens_seen": 2395734016
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002766811105761515,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4084738637831746,
+      "tokens_seen": 2395996160
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027660086663456907,
+      "loss": 0.0647,
+      "theoretical_loss": 3.408444116314432,
+      "tokens_seen": 2396258304
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002765206226929867,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4084143730108893,
+      "tokens_seen": 2396520448
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027644037875140424,
+      "loss": 0.0671,
+      "theoretical_loss": 3.408384633871509,
+      "tokens_seen": 2396782592
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027636013480982185,
+      "loss": 0.0685,
+      "theoretical_loss": 3.408354898895251,
+      "tokens_seen": 2397044736
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027627989086823946,
+      "loss": 0.067,
+      "theoretical_loss": 3.408325168081078,
+      "tokens_seen": 2397306880
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002761996469266571,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4082954414279523,
+      "tokens_seen": 2397569024
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027611940298507464,
+      "loss": 0.066,
+      "theoretical_loss": 3.408265718934837,
+      "tokens_seen": 2397831168
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027603915904349225,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4082360006006955,
+      "tokens_seen": 2398093312
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002759589151019098,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4082062864244906,
+      "tokens_seen": 2398355456
+    },
+    {
+      "epoch": 0.73,
+      "objective/train/advantage_avg": 0.0004930827999487519,
+      "objective/train/docs_used": 873551,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2975027561187744,
+      "objective/train/original_loss": 1.2975027561187744,
+      "objective/train/theoretical_loss": 3.4081765764051863,
+      "objective/train/tokens_used": 2419077600,
+      "objective/train/value_avg": -0.008544921875,
+      "objective/train/value_loss": 0.00024381437106058002,
+      "objective/train/value_max": -2.9981136322021484e-05,
+      "objective/train/value_min": -0.282958984375,
+      "objective/train/value_reward_corr": 0.7140946149248976,
+      "objective/train/value_std": 0.0169219970703125,
+      "objective/train/weight_avg": 1.0006061792373657,
+      "objective/train/weighted_lm_loss": 1.2970629930496216,
+      "objective/train/weights_max": 1.1696314811706543,
+      "objective/train/weights_min": 0.3694954514503479,
+      "theoretical_loss": 3.4081765764051863,
+      "tokens_seen": 2398617600
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027587867116032736,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4081765764051863,
+      "tokens_seen": 2398617600
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.000275798427218745,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4081468705417475,
+      "tokens_seen": 2398879744
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002757181832771626,
+      "loss": 0.0641,
+      "theoretical_loss": 3.408117168833138,
+      "tokens_seen": 2399141888
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002756379393355802,
+      "loss": 0.0649,
+      "theoretical_loss": 3.408087471278323,
+      "tokens_seen": 2399404032
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027555769539399776,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4080577778762686,
+      "tokens_seen": 2399666176
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002754774514524154,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4080280886259393,
+      "tokens_seen": 2399928320
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027539720751083293,
+      "loss": 0.0664,
+      "theoretical_loss": 3.4079984035263022,
+      "tokens_seen": 2400190464
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002753169635692505,
+      "loss": 0.0647,
+      "theoretical_loss": 3.4079687225763236,
+      "tokens_seen": 2400452608
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002752367196276681,
+      "loss": 0.0676,
+      "theoretical_loss": 3.40793904577497,
+      "tokens_seen": 2400714752
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002751564756860857,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4079093731212087,
+      "tokens_seen": 2400976896
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027507623174450333,
+      "loss": 0.0648,
+      "theoretical_loss": 3.4078797046140075,
+      "tokens_seen": 2401239040
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002749959878029209,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4078500402523346,
+      "tokens_seen": 2401501184
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002749157438613385,
+      "loss": 0.0654,
+      "theoretical_loss": 3.407820380035157,
+      "tokens_seen": 2401763328
+    },
+    {
+      "epoch": 0.73,
+      "objective/train/advantage_avg": 0.00013451933045871556,
+      "objective/train/docs_used": 874334,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2702131271362305,
+      "objective/train/original_loss": 1.2702131271362305,
+      "objective/train/theoretical_loss": 3.4078055514804326,
+      "objective/train/tokens_used": 2422354400,
+      "objective/train/value_avg": -0.007625579833984375,
+      "objective/train/value_loss": 0.00018812972120940685,
+      "objective/train/value_max": -1.2993812561035156e-05,
+      "objective/train/value_min": -0.2490234375,
+      "objective/train/value_reward_corr": 0.7233506520390734,
+      "objective/train/value_std": 0.01470947265625,
+      "objective/train/weight_avg": 1.0002230405807495,
+      "objective/train/weighted_lm_loss": 1.2700724601745605,
+      "objective/train/weights_max": 1.20888352394104,
+      "objective/train/weights_min": 0.36869922280311584,
+      "theoretical_loss": 3.4078055514804326,
+      "tokens_seen": 2401894400
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027483549991975606,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4077907239614453,
+      "tokens_seen": 2402025472
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002747552559781736,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4077610720301674,
+      "tokens_seen": 2402287616
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027467501203659123,
+      "loss": 0.0691,
+      "theoretical_loss": 3.407731424240293,
+      "tokens_seen": 2402549760
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027459476809500884,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4077017805907923,
+      "tokens_seen": 2402811904
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002745145241534264,
+      "loss": 0.0707,
+      "theoretical_loss": 3.4076721410806345,
+      "tokens_seen": 2403074048
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.000274434280211844,
+      "loss": 0.0658,
+      "theoretical_loss": 3.407642505708791,
+      "tokens_seen": 2403336192
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002743540362702616,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4076128744742324,
+      "tokens_seen": 2403598336
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027427379232867924,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4075832473759298,
+      "tokens_seen": 2403860480
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027419354838709674,
+      "loss": 0.0714,
+      "theoretical_loss": 3.4075536244128553,
+      "tokens_seen": 2404122624
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027411330444551435,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4075240055839804,
+      "tokens_seen": 2404384768
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027403306050393197,
+      "loss": 0.0666,
+      "theoretical_loss": 3.407494390888278,
+      "tokens_seen": 2404646912
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002739528165623495,
+      "loss": 0.068,
+      "theoretical_loss": 3.4074647803247204,
+      "tokens_seen": 2404909056
+    },
+    {
+      "epoch": 0.73,
+      "objective/train/advantage_avg": 0.0008375584147870541,
+      "objective/train/docs_used": 875512,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2580220699310303,
+      "objective/train/original_loss": 1.2580218315124512,
+      "objective/train/theoretical_loss": 3.407435173892281,
+      "objective/train/tokens_used": 2425631200,
+      "objective/train/value_avg": -0.004520416259765625,
+      "objective/train/value_loss": 0.0001223483559442684,
+      "objective/train/value_max": -3.3736228942871094e-05,
+      "objective/train/value_min": -0.1993408203125,
+      "objective/train/value_reward_corr": 0.5143376907001702,
+      "objective/train/value_std": 0.006969451904296875,
+      "objective/train/weight_avg": 1.0008904933929443,
+      "objective/train/weighted_lm_loss": 1.2593088150024414,
+      "objective/train/weights_max": 1.1975740194320679,
+      "objective/train/weights_min": 0.37003564834594727,
+      "theoretical_loss": 3.407435173892281,
+      "tokens_seen": 2405171200
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027387257262076714,
+      "loss": 0.0641,
+      "theoretical_loss": 3.407435173892281,
+      "tokens_seen": 2405171200
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027379232867918475,
+      "loss": 0.0696,
+      "theoretical_loss": 3.407405571589933,
+      "tokens_seen": 2405433344
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027371208473760236,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4073759734166504,
+      "tokens_seen": 2405695488
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027363184079601987,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4073463793714076,
+      "tokens_seen": 2405957632
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002735515968544375,
+      "loss": 0.0696,
+      "theoretical_loss": 3.407316789453179,
+      "tokens_seen": 2406219776
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002734713529128551,
+      "loss": 0.0664,
+      "theoretical_loss": 3.40728720366094,
+      "tokens_seen": 2406481920
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027339110897127265,
+      "loss": 0.0688,
+      "theoretical_loss": 3.407257621993665,
+      "tokens_seen": 2406744064
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027331086502969026,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4072280444503305,
+      "tokens_seen": 2407006208
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002732306210881079,
+      "loss": 0.0679,
+      "theoretical_loss": 3.407198471029912,
+      "tokens_seen": 2407268352
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002731503771465255,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4071689017313864,
+      "tokens_seen": 2407530496
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.000273070133204943,
+      "loss": 0.0687,
+      "theoretical_loss": 3.40713933655373,
+      "tokens_seen": 2407792640
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002729898892633606,
+      "loss": 0.0712,
+      "theoretical_loss": 3.40710977549592,
+      "tokens_seen": 2408054784
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002729096453217782,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4070802185569344,
+      "tokens_seen": 2408316928
+    },
+    {
+      "epoch": 0.73,
+      "objective/train/advantage_avg": 0.000534991966560483,
+      "objective/train/docs_used": 876655,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3676766157150269,
+      "objective/train/original_loss": 1.3676767349243164,
+      "objective/train/theoretical_loss": 3.4070654416316812,
+      "objective/train/tokens_used": 2428908000,
+      "objective/train/value_avg": -0.0054931640625,
+      "objective/train/value_loss": 0.0001075658219633624,
+      "objective/train/value_max": -6.0617923736572266e-05,
+      "objective/train/value_min": -0.206298828125,
+      "objective/train/value_reward_corr": 0.6255496585240548,
+      "objective/train/value_std": 0.00858306884765625,
+      "objective/train/weight_avg": 1.0005849599838257,
+      "objective/train/weighted_lm_loss": 1.3686153888702393,
+      "objective/train/weights_max": 1.2291204929351807,
+      "objective/train/weights_min": 0.396826833486557,
+      "theoretical_loss": 3.4070654416316812,
+      "tokens_seen": 2408448000
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002728294013801958,
+      "loss": 0.0686,
+      "theoretical_loss": 3.407050665735751,
+      "tokens_seen": 2408579072
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002727491574386134,
+      "loss": 0.0671,
+      "theoretical_loss": 3.407021117031347,
+      "tokens_seen": 2408841216
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.000272668913497031,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4069915724427027,
+      "tokens_seen": 2409103360
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002725886695554486,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4069620319687957,
+      "tokens_seen": 2409365504
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027250842561386617,
+      "loss": 0.0693,
+      "theoretical_loss": 3.406932495608606,
+      "tokens_seen": 2409627648
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027242818167228373,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4069029633611128,
+      "tokens_seen": 2409889792
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027234793773070134,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4068734352252967,
+      "tokens_seen": 2410151936
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002722676937891189,
+      "loss": 0.0667,
+      "theoretical_loss": 3.406843911200138,
+      "tokens_seen": 2410414080
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002721874498475365,
+      "loss": 0.0687,
+      "theoretical_loss": 3.406814391284617,
+      "tokens_seen": 2410676224
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027210720590595413,
+      "loss": 0.0654,
+      "theoretical_loss": 3.4067848754777152,
+      "tokens_seen": 2410938368
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002720269619643717,
+      "loss": 0.068,
+      "theoretical_loss": 3.4067553637784145,
+      "tokens_seen": 2411200512
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002719467180227893,
+      "loss": 0.0683,
+      "theoretical_loss": 3.406725856185696,
+      "tokens_seen": 2411462656
+    },
+    {
+      "epoch": 0.73,
+      "objective/train/advantage_avg": 0.00015087482461240143,
+      "objective/train/docs_used": 877919,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4046005010604858,
+      "objective/train/original_loss": 1.4046005010604858,
+      "objective/train/theoretical_loss": 3.4066963526985425,
+      "objective/train/tokens_used": 2432184800,
+      "objective/train/value_avg": -0.008453369140625,
+      "objective/train/value_loss": 0.00016879015311133116,
+      "objective/train/value_max": -5.8770179748535156e-05,
+      "objective/train/value_min": -0.48583984375,
+      "objective/train/value_reward_corr": 0.6949893811513895,
+      "objective/train/value_std": 0.01419830322265625,
+      "objective/train/weight_avg": 1.0002318620681763,
+      "objective/train/weighted_lm_loss": 1.4044239521026611,
+      "objective/train/weights_max": 1.409144401550293,
+      "objective/train/weights_min": 0.3955934941768646,
+      "theoretical_loss": 3.4066963526985425,
+      "tokens_seen": 2411724800
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027186647408120686,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4066963526985425,
+      "tokens_seen": 2411724800
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027178623013962447,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4066668533159365,
+      "tokens_seen": 2411986944
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027170598619804203,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4066373580368614,
+      "tokens_seen": 2412249088
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027162574225645964,
+      "loss": 0.0692,
+      "theoretical_loss": 3.4066078668602997,
+      "tokens_seen": 2412511232
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027154549831487725,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4065783797852354,
+      "tokens_seen": 2412773376
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002714652543732948,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4065488968106528,
+      "tokens_seen": 2413035520
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002713850104317124,
+      "loss": 0.069,
+      "theoretical_loss": 3.406519417935536,
+      "tokens_seen": 2413297664
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027130476649013004,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4064899431588698,
+      "tokens_seen": 2413559808
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002712245225485476,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4064604724796395,
+      "tokens_seen": 2413821952
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027114427860696515,
+      "loss": 0.0682,
+      "theoretical_loss": 3.4064310058968306,
+      "tokens_seen": 2414084096
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027106403466538277,
+      "loss": 0.0637,
+      "theoretical_loss": 3.406401543409429,
+      "tokens_seen": 2414346240
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002709837907238004,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4063720850164203,
+      "tokens_seen": 2414608384
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027090354678221794,
+      "loss": 0.0674,
+      "theoretical_loss": 3.406342630716792,
+      "tokens_seen": 2414870528
+    },
+    {
+      "epoch": 0.73,
+      "objective/train/advantage_avg": -0.00010549866419751197,
+      "objective/train/docs_used": 879191,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2191466093063354,
+      "objective/train/original_loss": 1.219146728515625,
+      "objective/train/theoretical_loss": 3.406327905101679,
+      "objective/train/tokens_used": 2435461600,
+      "objective/train/value_avg": -0.006084442138671875,
+      "objective/train/value_loss": 0.00020990363555029035,
+      "objective/train/value_max": -2.9981136322021484e-05,
+      "objective/train/value_min": -0.39013671875,
+      "objective/train/value_reward_corr": 0.6207213945178252,
+      "objective/train/value_std": 0.0113372802734375,
+      "objective/train/weight_avg": 0.999985933303833,
+      "objective/train/weighted_lm_loss": 1.2187398672103882,
+      "objective/train/weights_max": 1.2805371284484863,
+      "objective/train/weights_min": 0.37227270007133484,
+      "theoretical_loss": 3.406327905101679,
+      "tokens_seen": 2415001600
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027082330284063555,
+      "loss": 0.068,
+      "theoretical_loss": 3.406313180509531,
+      "tokens_seen": 2415132672
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027074305889905316,
+      "loss": 0.0677,
+      "theoretical_loss": 3.406283734393624,
+      "tokens_seen": 2415394816
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027066281495747067,
+      "loss": 0.066,
+      "theoretical_loss": 3.4062542923680588,
+      "tokens_seen": 2415656960
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002705825710158883,
+      "loss": 0.0682,
+      "theoretical_loss": 3.406224854431824,
+      "tokens_seen": 2415919104
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002705023270743059,
+      "loss": 0.0685,
+      "theoretical_loss": 3.406195420583907,
+      "tokens_seen": 2416181248
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002704220831327235,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4061659908232973,
+      "tokens_seen": 2416443392
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00027034183919114106,
+      "loss": 0.067,
+      "theoretical_loss": 3.4061365651489837,
+      "tokens_seen": 2416705536
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002702615952495587,
+      "loss": 0.0687,
+      "theoretical_loss": 3.406107143559956,
+      "tokens_seen": 2416967680
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002701813513079763,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4060777260552038,
+      "tokens_seen": 2417229824
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002701011073663938,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4060483126337173,
+      "tokens_seen": 2417491968
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002700208634248114,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4060189032944868,
+      "tokens_seen": 2417754112
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.000269940619483229,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4059894980365035,
+      "tokens_seen": 2418016256
+    },
+    {
+      "epoch": 0.73,
+      "objective/train/advantage_avg": -0.00011731717677321285,
+      "objective/train/docs_used": 880852,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3383729457855225,
+      "objective/train/original_loss": 1.3383729457855225,
+      "objective/train/theoretical_loss": 3.4059600968587587,
+      "objective/train/tokens_used": 2438738400,
+      "objective/train/value_avg": -0.00749969482421875,
+      "objective/train/value_loss": 0.0005007804138585925,
+      "objective/train/value_max": -2.485513687133789e-05,
+      "objective/train/value_min": -0.61767578125,
+      "objective/train/value_reward_corr": 0.7091478847452439,
+      "objective/train/value_std": 0.0192413330078125,
+      "objective/train/weight_avg": 1.0000945329666138,
+      "objective/train/weighted_lm_loss": 1.3375656604766846,
+      "objective/train/weights_max": 1.3170450925827026,
+      "objective/train/weights_min": 0.3693002760410309,
+      "theoretical_loss": 3.4059600968587587,
+      "tokens_seen": 2418278400
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026986037554164663,
+      "loss": 0.0676,
+      "theoretical_loss": 3.4059600968587587,
+      "tokens_seen": 2418278400
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002697801316000642,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4059306997602437,
+      "tokens_seen": 2418540544
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002696998876584818,
+      "loss": 0.066,
+      "theoretical_loss": 3.405901306739951,
+      "tokens_seen": 2418802688
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002696196437168994,
+      "loss": 0.0658,
+      "theoretical_loss": 3.405871917796872,
+      "tokens_seen": 2419064832
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026953939977531697,
+      "loss": 0.0696,
+      "theoretical_loss": 3.405842532930001,
+      "tokens_seen": 2419326976
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026945915583373453,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4058131521383292,
+      "tokens_seen": 2419589120
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026937891189215214,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4057837754208515,
+      "tokens_seen": 2419851264
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026929866795056976,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4057544027765605,
+      "tokens_seen": 2420113408
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002692184240089873,
+      "loss": 0.0701,
+      "theoretical_loss": 3.4057250342044516,
+      "tokens_seen": 2420375552
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002691381800674049,
+      "loss": 0.0651,
+      "theoretical_loss": 3.405695669703518,
+      "tokens_seen": 2420637696
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026905793612582254,
+      "loss": 0.0663,
+      "theoretical_loss": 3.405666309272755,
+      "tokens_seen": 2420899840
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002689776921842401,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4056369529111583,
+      "tokens_seen": 2421161984
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026889744824265766,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4056076006177234,
+      "tokens_seen": 2421424128
+    },
+    {
+      "epoch": 0.73,
+      "objective/train/advantage_avg": 0.0005559303681366146,
+      "objective/train/docs_used": 881885,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2067722082138062,
+      "objective/train/original_loss": 1.2067723274230957,
+      "objective/train/theoretical_loss": 3.4055929259962525,
+      "objective/train/tokens_used": 2442015200,
+      "objective/train/value_avg": -0.004528045654296875,
+      "objective/train/value_loss": 7.930587162263691e-05,
+      "objective/train/value_max": -4.4345855712890625e-05,
+      "objective/train/value_min": -0.2314453125,
+      "objective/train/value_reward_corr": 0.7267003287627125,
+      "objective/train/value_std": 0.0087890625,
+      "objective/train/weight_avg": 1.0005950927734375,
+      "objective/train/weighted_lm_loss": 1.207557201385498,
+      "objective/train/weights_max": 1.1026510000228882,
+      "objective/train/weights_min": 0.8195127248764038,
+      "theoretical_loss": 3.4055929259962525,
+      "tokens_seen": 2421555200
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026881720430107527,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4055782523914453,
+      "tokens_seen": 2421686272
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002687369603594929,
+      "loss": 0.0679,
+      "theoretical_loss": 3.405548908231321,
+      "tokens_seen": 2421948416
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026865671641791044,
+      "loss": 0.0666,
+      "theoretical_loss": 3.405519568136347,
+      "tokens_seen": 2422210560
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026857647247632805,
+      "loss": 0.0714,
+      "theoretical_loss": 3.40549023210552,
+      "tokens_seen": 2422472704
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026849622853474566,
+      "loss": 0.0705,
+      "theoretical_loss": 3.4054609001378378,
+      "tokens_seen": 2422734848
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002684159845931632,
+      "loss": 0.0684,
+      "theoretical_loss": 3.4054315722322976,
+      "tokens_seen": 2422996992
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002683357406515808,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4054022483878983,
+      "tokens_seen": 2423259136
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002682554967099984,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4053729286036374,
+      "tokens_seen": 2423521280
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026817525276841595,
+      "loss": 0.0662,
+      "theoretical_loss": 3.405343612878514,
+      "tokens_seen": 2423783424
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026809500882683357,
+      "loss": 0.0652,
+      "theoretical_loss": 3.405314301211527,
+      "tokens_seen": 2424045568
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002680147648852512,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4052849936016765,
+      "tokens_seen": 2424307712
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002679345209436688,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4052556900479614,
+      "tokens_seen": 2424569856
+    },
+    {
+      "epoch": 0.73,
+      "objective/train/advantage_avg": 0.0001301930460613221,
+      "objective/train/docs_used": 882990,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4599616527557373,
+      "objective/train/original_loss": 1.4599618911743164,
+      "objective/train/theoretical_loss": 3.405226390549383,
+      "objective/train/tokens_used": 2445292000,
+      "objective/train/value_avg": -0.006389617919921875,
+      "objective/train/value_loss": 0.0001866083184722811,
+      "objective/train/value_max": -4.100799560546875e-05,
+      "objective/train/value_min": -0.35546875,
+      "objective/train/value_reward_corr": 0.6365628212502612,
+      "objective/train/value_std": 0.01094818115234375,
+      "objective/train/weight_avg": 1.0002139806747437,
+      "objective/train/weighted_lm_loss": 1.4597046375274658,
+      "objective/train/weights_max": 1.174597978591919,
+      "objective/train/weights_min": 0.3728525638580322,
+      "theoretical_loss": 3.405226390549383,
+      "tokens_seen": 2424832000
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026785427700208635,
+      "loss": 0.0676,
+      "theoretical_loss": 3.405226390549383,
+      "tokens_seen": 2424832000
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.00026777403306050396,
+      "loss": 0.0685,
+      "theoretical_loss": 3.4051970951049406,
+      "tokens_seen": 2425094144
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 0.0002676937891189215,
+      "loss": 0.0694,
+      "theoretical_loss": 3.4051678037136357,
+      "tokens_seen": 2425356288
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002676135451773391,
+      "loss": 0.0674,
+      "theoretical_loss": 3.40513851637447,
+      "tokens_seen": 2425618432
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002675333012357567,
+      "loss": 0.0679,
+      "theoretical_loss": 3.4051092330864443,
+      "tokens_seen": 2425880576
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002674530572941743,
+      "loss": 0.0674,
+      "theoretical_loss": 3.405079953848561,
+      "tokens_seen": 2426142720
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002673728133525919,
+      "loss": 0.068,
+      "theoretical_loss": 3.4050506786598222,
+      "tokens_seen": 2426404864
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002672925694110095,
+      "loss": 0.0687,
+      "theoretical_loss": 3.405021407519231,
+      "tokens_seen": 2426667008
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002672123254694271,
+      "loss": 0.068,
+      "theoretical_loss": 3.4049921404257897,
+      "tokens_seen": 2426929152
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026713208152784465,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4049628773785017,
+      "tokens_seen": 2427191296
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002670518375862622,
+      "loss": 0.0648,
+      "theoretical_loss": 3.4049336183763717,
+      "tokens_seen": 2427453440
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002669715936446798,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4049043634184026,
+      "tokens_seen": 2427715584
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026689134970309743,
+      "loss": 0.068,
+      "theoretical_loss": 3.4048751125035994,
+      "tokens_seen": 2427977728
+    },
+    {
+      "epoch": 0.74,
+      "objective/train/advantage_avg": 0.0006799577386118472,
+      "objective/train/docs_used": 884091,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.223225474357605,
+      "objective/train/original_loss": 1.2232253551483154,
+      "objective/train/theoretical_loss": 3.404860488562074,
+      "objective/train/tokens_used": 2448568800,
+      "objective/train/value_avg": -0.006168365478515625,
+      "objective/train/value_loss": 9.300230158260092e-05,
+      "objective/train/value_max": -3.0219554901123047e-05,
+      "objective/train/value_min": -0.19287109375,
+      "objective/train/value_reward_corr": 0.6994599916806957,
+      "objective/train/value_std": 0.0108184814453125,
+      "objective/train/weight_avg": 1.0007256269454956,
+      "objective/train/weighted_lm_loss": 1.2240463495254517,
+      "objective/train/weights_max": 1.1006338596343994,
+      "objective/train/weights_min": 0.6186822056770325,
+      "theoretical_loss": 3.404860488562074,
+      "tokens_seen": 2428108800
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026681110576151504,
+      "loss": 0.0671,
+      "theoretical_loss": 3.404845865630967,
+      "tokens_seen": 2428239872
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002667308618199326,
+      "loss": 0.0674,
+      "theoretical_loss": 3.40481662279951,
+      "tokens_seen": 2428502016
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002666506178783502,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4047873840082348,
+      "tokens_seen": 2428764160
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002665703739367678,
+      "loss": 0.0673,
+      "theoretical_loss": 3.404758149256146,
+      "tokens_seen": 2429026304
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026649012999518533,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4047289185422507,
+      "tokens_seen": 2429288448
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026640988605360294,
+      "loss": 0.0653,
+      "theoretical_loss": 3.404699691865555,
+      "tokens_seen": 2429550592
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026632964211202055,
+      "loss": 0.068,
+      "theoretical_loss": 3.4046704692250658,
+      "tokens_seen": 2429812736
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002662493981704381,
+      "loss": 0.0663,
+      "theoretical_loss": 3.404641250619791,
+      "tokens_seen": 2430074880
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002661691542288557,
+      "loss": 0.0694,
+      "theoretical_loss": 3.404612036048737,
+      "tokens_seen": 2430337024
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026608891028727334,
+      "loss": 0.0683,
+      "theoretical_loss": 3.404582825510913,
+      "tokens_seen": 2430599168
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026600866634569095,
+      "loss": 0.0667,
+      "theoretical_loss": 3.404553619005326,
+      "tokens_seen": 2430861312
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026592842240410846,
+      "loss": 0.0672,
+      "theoretical_loss": 3.404524416530986,
+      "tokens_seen": 2431123456
+    },
+    {
+      "epoch": 0.74,
+      "objective/train/advantage_avg": 0.0008247354999184608,
+      "objective/train/docs_used": 885315,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3540970087051392,
+      "objective/train/original_loss": 1.3540968894958496,
+      "objective/train/theoretical_loss": 3.4044952180869013,
+      "objective/train/tokens_used": 2451845600,
+      "objective/train/value_avg": -0.00623321533203125,
+      "objective/train/value_loss": 0.00027920593856833875,
+      "objective/train/value_max": -4.13060188293457e-05,
+      "objective/train/value_min": -0.845703125,
+      "objective/train/value_reward_corr": 0.6324752214183904,
+      "objective/train/value_std": 0.0144500732421875,
+      "objective/train/weight_avg": 1.000954508781433,
+      "objective/train/weighted_lm_loss": 1.3552262783050537,
+      "objective/train/weights_max": 2.2237422466278076,
+      "objective/train/weights_min": 0.38558143377304077,
+      "theoretical_loss": 3.4044952180869013,
+      "tokens_seen": 2431385600
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026584817846252607,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4044952180869013,
+      "tokens_seen": 2431385600
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002657679345209437,
+      "loss": 0.0688,
+      "theoretical_loss": 3.4044660236720805,
+      "tokens_seen": 2431647744
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026568769057936124,
+      "loss": 0.0655,
+      "theoretical_loss": 3.404436833285535,
+      "tokens_seen": 2431909888
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026560744663777885,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4044076469262734,
+      "tokens_seen": 2432172032
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026552720269619646,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4043784645933064,
+      "tokens_seen": 2432434176
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002654469587546141,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4043492862856457,
+      "tokens_seen": 2432696320
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002653667148130316,
+      "loss": 0.0685,
+      "theoretical_loss": 3.404320112002301,
+      "tokens_seen": 2432958464
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002652864708714492,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4042909417422846,
+      "tokens_seen": 2433220608
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002652062269298668,
+      "loss": 0.0678,
+      "theoretical_loss": 3.404261775504608,
+      "tokens_seen": 2433482752
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026512598298828436,
+      "loss": 0.0657,
+      "theoretical_loss": 3.404232613288283,
+      "tokens_seen": 2433744896
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.000265045739046702,
+      "loss": 0.0675,
+      "theoretical_loss": 3.404203455092323,
+      "tokens_seen": 2434007040
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002649654951051196,
+      "loss": 0.0651,
+      "theoretical_loss": 3.40417430091574,
+      "tokens_seen": 2434269184
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002648852511635372,
+      "loss": 0.0675,
+      "theoretical_loss": 3.404145150757548,
+      "tokens_seen": 2434531328
+    },
+    {
+      "epoch": 0.74,
+      "objective/train/advantage_avg": 0.0004898194456472993,
+      "objective/train/docs_used": 886508,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3692785501480103,
+      "objective/train/original_loss": 1.369278907775879,
+      "objective/train/theoretical_loss": 3.40413057718504,
+      "objective/train/tokens_used": 2455122400,
+      "objective/train/value_avg": -0.005901336669921875,
+      "objective/train/value_loss": 0.0002257556770928204,
+      "objective/train/value_max": -3.170967102050781e-05,
+      "objective/train/value_min": -0.65185546875,
+      "objective/train/value_reward_corr": 0.6215079431063899,
+      "objective/train/value_std": 0.01128387451171875,
+      "objective/train/weight_avg": 1.0005888938903809,
+      "objective/train/weighted_lm_loss": 1.3697773218154907,
+      "objective/train/weights_max": 1.3338719606399536,
+      "objective/train/weights_min": 0.3692946434020996,
+      "theoretical_loss": 3.40413057718504,
+      "tokens_seen": 2434662400
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026480500722195476,
+      "loss": 0.0683,
+      "theoretical_loss": 3.4041160046167596,
+      "tokens_seen": 2434793472
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002647247632803723,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4040868624923895,
+      "tokens_seen": 2435055616
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026464451933878993,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4040577243834513,
+      "tokens_seen": 2435317760
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002645642753972075,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4040285902889598,
+      "tokens_seen": 2435579904
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002644840314556251,
+      "loss": 0.0682,
+      "theoretical_loss": 3.40399946020793,
+      "tokens_seen": 2435842048
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002644037875140427,
+      "loss": 0.0665,
+      "theoretical_loss": 3.403970334139377,
+      "tokens_seen": 2436104192
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026432354357246033,
+      "loss": 0.0655,
+      "theoretical_loss": 3.403941212082317,
+      "tokens_seen": 2436366336
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002642432996308779,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4039120940357654,
+      "tokens_seen": 2436628480
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026416305568929544,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4038829799987385,
+      "tokens_seen": 2436890624
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026408281174771306,
+      "loss": 0.0674,
+      "theoretical_loss": 3.4038538699702534,
+      "tokens_seen": 2437152768
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002640025678061306,
+      "loss": 0.0664,
+      "theoretical_loss": 3.403824763949326,
+      "tokens_seen": 2437414912
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026392232386454823,
+      "loss": 0.0673,
+      "theoretical_loss": 3.403795661934976,
+      "tokens_seen": 2437677056
+    },
+    {
+      "epoch": 0.74,
+      "objective/train/advantage_avg": -0.0008872678154148161,
+      "objective/train/docs_used": 887686,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3116480112075806,
+      "objective/train/original_loss": 1.3116480112075806,
+      "objective/train/theoretical_loss": 3.4037665639262182,
+      "objective/train/tokens_used": 2458399200,
+      "objective/train/value_avg": -0.007190704345703125,
+      "objective/train/value_loss": 0.000523583497852087,
+      "objective/train/value_max": -2.753734588623047e-05,
+      "objective/train/value_min": -0.70947265625,
+      "objective/train/value_reward_corr": 0.6604272960853255,
+      "objective/train/value_std": 0.0171356201171875,
+      "objective/train/weight_avg": 0.9993419051170349,
+      "objective/train/weighted_lm_loss": 1.309940218925476,
+      "objective/train/weights_max": 1.3254317045211792,
+      "objective/train/weights_min": 0.37620142102241516,
+      "theoretical_loss": 3.4037665639262182,
+      "tokens_seen": 2437939200
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026384207992296584,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4037665639262182,
+      "tokens_seen": 2437939200
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002637618359813834,
+      "loss": 0.0654,
+      "theoretical_loss": 3.403737469922073,
+      "tokens_seen": 2438201344
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.000263681592039801,
+      "loss": 0.0675,
+      "theoretical_loss": 3.4037083799215573,
+      "tokens_seen": 2438463488
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026360134809821857,
+      "loss": 0.0647,
+      "theoretical_loss": 3.40367929392369,
+      "tokens_seen": 2438725632
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002635211041566362,
+      "loss": 0.066,
+      "theoretical_loss": 3.4036502119274914,
+      "tokens_seen": 2438987776
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026344086021505374,
+      "loss": 0.0633,
+      "theoretical_loss": 3.4036211339319795,
+      "tokens_seen": 2439249920
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026336061627347135,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4035920599361753,
+      "tokens_seen": 2439512064
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026328037233188897,
+      "loss": 0.0686,
+      "theoretical_loss": 3.403562989939098,
+      "tokens_seen": 2439774208
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002632001283903065,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4035339239397686,
+      "tokens_seen": 2440036352
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026311988444872414,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4035048619372072,
+      "tokens_seen": 2440298496
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026303964050714175,
+      "loss": 0.0664,
+      "theoretical_loss": 3.403475803930436,
+      "tokens_seen": 2440560640
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002629593965655593,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4034467499184755,
+      "tokens_seen": 2440822784
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026287915262397687,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4034176999003485,
+      "tokens_seen": 2441084928
+    },
+    {
+      "epoch": 0.74,
+      "objective/train/advantage_avg": 8.833639731165022e-05,
+      "objective/train/docs_used": 888977,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1548739671707153,
+      "objective/train/original_loss": 1.1548740863800049,
+      "objective/train/theoretical_loss": 3.4034031763886667,
+      "objective/train/tokens_used": 2461676000,
+      "objective/train/value_avg": -0.00930023193359375,
+      "objective/train/value_loss": 0.00023279798915609717,
+      "objective/train/value_max": -6.973743438720703e-05,
+      "objective/train/value_min": -0.62548828125,
+      "objective/train/value_reward_corr": 0.76527345823869,
+      "objective/train/value_std": 0.016143798828125,
+      "objective/train/weight_avg": 1.0001966953277588,
+      "objective/train/weighted_lm_loss": 1.1562342643737793,
+      "objective/train/weights_max": 1.419206142425537,
+      "objective/train/weights_min": 0.4449949264526367,
+      "theoretical_loss": 3.4034031763886667,
+      "tokens_seen": 2441216000
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002627989086823945,
+      "loss": 0.0649,
+      "theoretical_loss": 3.4033886538750764,
+      "tokens_seen": 2441347072
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002627186647408121,
+      "loss": 0.0667,
+      "theoretical_loss": 3.4033596118416822,
+      "tokens_seen": 2441609216
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026263842079922965,
+      "loss": 0.0632,
+      "theoretical_loss": 3.4033305737991886,
+      "tokens_seen": 2441871360
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026255817685764726,
+      "loss": 0.068,
+      "theoretical_loss": 3.403301539746619,
+      "tokens_seen": 2442133504
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002624779329160649,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4032725096829965,
+      "tokens_seen": 2442395648
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002623976889744824,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4032434836073455,
+      "tokens_seen": 2442657792
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002623174450329,
+      "loss": 0.0688,
+      "theoretical_loss": 3.40321446151869,
+      "tokens_seen": 2442919936
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002622372010913176,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4031854434160556,
+      "tokens_seen": 2443182080
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002621569571497352,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4031564292984653,
+      "tokens_seen": 2443444224
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002620767132081528,
+      "loss": 0.0667,
+      "theoretical_loss": 3.403127419164946,
+      "tokens_seen": 2443706368
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002619964692665704,
+      "loss": 0.0681,
+      "theoretical_loss": 3.4030984130145225,
+      "tokens_seen": 2443968512
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.000261916225324988,
+      "loss": 0.0689,
+      "theoretical_loss": 3.4030694108462214,
+      "tokens_seen": 2444230656
+    },
+    {
+      "epoch": 0.74,
+      "objective/train/advantage_avg": 0.0005917379749007523,
+      "objective/train/docs_used": 890203,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4914954900741577,
+      "objective/train/original_loss": 1.4914956092834473,
+      "objective/train/theoretical_loss": 3.4030404126590685,
+      "objective/train/tokens_used": 2464952800,
+      "objective/train/value_avg": -0.00826263427734375,
+      "objective/train/value_loss": 0.00019185063138138503,
+      "objective/train/value_max": -5.066394805908203e-05,
+      "objective/train/value_min": -0.35595703125,
+      "objective/train/value_reward_corr": 0.7125985312722368,
+      "objective/train/value_std": 0.01407623291015625,
+      "objective/train/weight_avg": 1.0006788969039917,
+      "objective/train/weighted_lm_loss": 1.4919967651367188,
+      "objective/train/weights_max": 1.2016677856445312,
+      "objective/train/weights_min": 0.3696744740009308,
+      "theoretical_loss": 3.4030404126590685,
+      "tokens_seen": 2444492800
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026183598138340556,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4030404126590685,
+      "tokens_seen": 2444492800
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002617557374418231,
+      "loss": 0.0686,
+      "theoretical_loss": 3.403011418452091,
+      "tokens_seen": 2444754944
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026167549350024073,
+      "loss": 0.0682,
+      "theoretical_loss": 3.402982428224315,
+      "tokens_seen": 2445017088
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026159524955865834,
+      "loss": 0.0687,
+      "theoretical_loss": 3.4029534419747685,
+      "tokens_seen": 2445279232
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002615150056170759,
+      "loss": 0.0677,
+      "theoretical_loss": 3.4029244597024793,
+      "tokens_seen": 2445541376
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002614347616754935,
+      "loss": 0.0657,
+      "theoretical_loss": 3.402895481406475,
+      "tokens_seen": 2445803520
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026135451773391113,
+      "loss": 0.065,
+      "theoretical_loss": 3.4028665070857844,
+      "tokens_seen": 2446065664
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002612742737923287,
+      "loss": 0.0672,
+      "theoretical_loss": 3.402837536739436,
+      "tokens_seen": 2446327808
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026119402985074624,
+      "loss": 0.0671,
+      "theoretical_loss": 3.4028085703664583,
+      "tokens_seen": 2446589952
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026111378590916386,
+      "loss": 0.0672,
+      "theoretical_loss": 3.402779607965882,
+      "tokens_seen": 2446852096
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026103354196758147,
+      "loss": 0.0686,
+      "theoretical_loss": 3.4027506495367357,
+      "tokens_seen": 2447114240
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026095329802599903,
+      "loss": 0.0666,
+      "theoretical_loss": 3.4027216950780494,
+      "tokens_seen": 2447376384
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026087305408441664,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4026927445888546,
+      "tokens_seen": 2447638528
+    },
+    {
+      "epoch": 0.74,
+      "objective/train/advantage_avg": 0.0009709555306471884,
+      "objective/train/docs_used": 890833,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2022894620895386,
+      "objective/train/original_loss": 1.2022895812988281,
+      "objective/train/theoretical_loss": 3.402678270832513,
+      "objective/train/tokens_used": 2468229600,
+      "objective/train/value_avg": -0.006378173828125,
+      "objective/train/value_loss": 0.00013867423695046455,
+      "objective/train/value_max": -7.140636444091797e-05,
+      "objective/train/value_min": -0.228515625,
+      "objective/train/value_reward_corr": 0.6964469034006823,
+      "objective/train/value_std": 0.0112457275390625,
+      "objective/train/weight_avg": 1.001035451889038,
+      "objective/train/weighted_lm_loss": 1.2034066915512085,
+      "objective/train/weights_max": 1.19721257686615,
+      "objective/train/weights_min": 0.37144145369529724,
+      "theoretical_loss": 3.402678270832513,
+      "tokens_seen": 2447769600
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026079281014283425,
+      "loss": 0.061,
+      "theoretical_loss": 3.4026637980681813,
+      "tokens_seen": 2447900672
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002607125662012518,
+      "loss": 0.0663,
+      "theoretical_loss": 3.4026348555150605,
+      "tokens_seen": 2448162816
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026063232225966937,
+      "loss": 0.0668,
+      "theoretical_loss": 3.402605916928524,
+      "tokens_seen": 2448424960
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.000260552078318087,
+      "loss": 0.0659,
+      "theoretical_loss": 3.4025769823076035,
+      "tokens_seen": 2448687104
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002604718343765046,
+      "loss": 0.0671,
+      "theoretical_loss": 3.402548051651331,
+      "tokens_seen": 2448949248
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026039159043492215,
+      "loss": 0.0689,
+      "theoretical_loss": 3.402519124958739,
+      "tokens_seen": 2449211392
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026031134649333977,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4024902022288606,
+      "tokens_seen": 2449473536
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002602311025517574,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4024612834607284,
+      "tokens_seen": 2449735680
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026015085861017494,
+      "loss": 0.0618,
+      "theoretical_loss": 3.402432368653376,
+      "tokens_seen": 2449997824
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00026007061466859255,
+      "loss": 0.0668,
+      "theoretical_loss": 3.4024034578058373,
+      "tokens_seen": 2450259968
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002599903707270101,
+      "loss": 0.0646,
+      "theoretical_loss": 3.4023745509171466,
+      "tokens_seen": 2450522112
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025991012678542767,
+      "loss": 0.0638,
+      "theoretical_loss": 3.4023456479863383,
+      "tokens_seen": 2450784256
+    },
+    {
+      "epoch": 0.74,
+      "objective/train/advantage_avg": 0.000830281525850296,
+      "objective/train/docs_used": 891914,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.410335659980774,
+      "objective/train/original_loss": 1.4103357791900635,
+      "objective/train/theoretical_loss": 3.402316749012447,
+      "objective/train/tokens_used": 2471506400,
+      "objective/train/value_avg": -0.006683349609375,
+      "objective/train/value_loss": 0.00012804150173906237,
+      "objective/train/value_max": -2.4318695068359375e-05,
+      "objective/train/value_min": -0.2186279296875,
+      "objective/train/value_reward_corr": 0.6649692993754099,
+      "objective/train/value_std": 0.01036834716796875,
+      "objective/train/weight_avg": 1.0008869171142578,
+      "objective/train/weighted_lm_loss": 1.4112725257873535,
+      "objective/train/weights_max": 1.172834873199463,
+      "objective/train/weights_min": 0.3791404366493225,
+      "theoretical_loss": 3.402316749012447,
+      "tokens_seen": 2451046400
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002598298828438453,
+      "loss": 0.0649,
+      "theoretical_loss": 3.402316749012447,
+      "tokens_seen": 2451046400
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002597496389022629,
+      "loss": 0.0623,
+      "theoretical_loss": 3.4022878539945087,
+      "tokens_seen": 2451308544
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002596693949606805,
+      "loss": 0.0664,
+      "theoretical_loss": 3.402258962931558,
+      "tokens_seen": 2451570688
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025958915101909806,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4022300758226307,
+      "tokens_seen": 2451832832
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002595089070775157,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4022011926667637,
+      "tokens_seen": 2452094976
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025942866313593323,
+      "loss": 0.0637,
+      "theoretical_loss": 3.402172313462993,
+      "tokens_seen": 2452357120
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002593484191943508,
+      "loss": 0.0652,
+      "theoretical_loss": 3.4021434382103557,
+      "tokens_seen": 2452619264
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002592681752527684,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4021145669078887,
+      "tokens_seen": 2452881408
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.000259187931311186,
+      "loss": 0.0651,
+      "theoretical_loss": 3.40208569955463,
+      "tokens_seen": 2453143552
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025910768736960363,
+      "loss": 0.0651,
+      "theoretical_loss": 3.402056836149617,
+      "tokens_seen": 2453405696
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002590274434280212,
+      "loss": 0.0634,
+      "theoretical_loss": 3.4020279766918886,
+      "tokens_seen": 2453667840
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002589471994864388,
+      "loss": 0.0661,
+      "theoretical_loss": 3.401999121180482,
+      "tokens_seen": 2453929984
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025886695554485636,
+      "loss": 0.0604,
+      "theoretical_loss": 3.401970269614438,
+      "tokens_seen": 2454192128
+    },
+    {
+      "epoch": 0.74,
+      "objective/train/advantage_avg": 0.0005356064066290855,
+      "objective/train/docs_used": 893100,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3348721265792847,
+      "objective/train/original_loss": 1.3348722457885742,
+      "objective/train/theoretical_loss": 3.401955845310626,
+      "objective/train/tokens_used": 2474783200,
+      "objective/train/value_avg": -0.006534576416015625,
+      "objective/train/value_loss": 0.00014540940173901618,
+      "objective/train/value_max": -1.806020736694336e-05,
+      "objective/train/value_min": -0.311767578125,
+      "objective/train/value_reward_corr": 0.7150836623406631,
+      "objective/train/value_std": 0.0132293701171875,
+      "objective/train/weight_avg": 1.0006065368652344,
+      "objective/train/weighted_lm_loss": 1.335172414779663,
+      "objective/train/weights_max": 1.1883995532989502,
+      "objective/train/weights_min": 0.7188173532485962,
+      "theoretical_loss": 3.401955845310626,
+      "tokens_seen": 2454323200
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002587867116032739,
+      "loss": 0.065,
+      "theoretical_loss": 3.4019414219927944,
+      "tokens_seen": 2454454272
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025870646766169153,
+      "loss": 0.0622,
+      "theoretical_loss": 3.4019125783145907,
+      "tokens_seen": 2454716416
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025862622372010914,
+      "loss": 0.0629,
+      "theoretical_loss": 3.4018837385788676,
+      "tokens_seen": 2454978560
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025854597977852676,
+      "loss": 0.0666,
+      "theoretical_loss": 3.401854902784665,
+      "tokens_seen": 2455240704
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002584657358369443,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4018260709310235,
+      "tokens_seen": 2455502848
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002583854918953619,
+      "loss": 0.0634,
+      "theoretical_loss": 3.4017972430169836,
+      "tokens_seen": 2455764992
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025830524795377954,
+      "loss": 0.0658,
+      "theoretical_loss": 3.4017684190415873,
+      "tokens_seen": 2456027136
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025822500401219704,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4017395990038755,
+      "tokens_seen": 2456289280
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025814476007061466,
+      "loss": 0.0655,
+      "theoretical_loss": 3.4017107829028905,
+      "tokens_seen": 2456551424
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025806451612903227,
+      "loss": 0.0632,
+      "theoretical_loss": 3.401681970737674,
+      "tokens_seen": 2456813568
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002579842721874498,
+      "loss": 0.0653,
+      "theoretical_loss": 3.4016531625072695,
+      "tokens_seen": 2457075712
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025790402824586744,
+      "loss": 0.0672,
+      "theoretical_loss": 3.4016243582107193,
+      "tokens_seen": 2457337856
+    },
+    {
+      "epoch": 0.74,
+      "objective/train/advantage_avg": 0.0009694792679511011,
+      "objective/train/docs_used": 894282,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2399897575378418,
+      "objective/train/original_loss": 1.2399897575378418,
+      "objective/train/theoretical_loss": 3.4015955578470667,
+      "objective/train/tokens_used": 2478060000,
+      "objective/train/value_avg": -0.005939483642578125,
+      "objective/train/value_loss": 0.00014201925660017878,
+      "objective/train/value_max": -2.6702880859375e-05,
+      "objective/train/value_min": -0.2325439453125,
+      "objective/train/value_reward_corr": 0.7003671634648543,
+      "objective/train/value_std": 0.01181793212890625,
+      "objective/train/weight_avg": 1.0010325908660889,
+      "objective/train/weighted_lm_loss": 1.2412679195404053,
+      "objective/train/weights_max": 1.2071141004562378,
+      "objective/train/weights_min": 0.3796498477458954,
+      "theoretical_loss": 3.4015955578470667,
+      "tokens_seen": 2457600000
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025782378430428505,
+      "loss": 0.063,
+      "theoretical_loss": 3.4015955578470667,
+      "tokens_seen": 2457600000
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025774354036270266,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4015667614153555,
+      "tokens_seen": 2457862144
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.00025766329642112017,
+      "loss": 0.0632,
+      "theoretical_loss": 3.4015379689146292,
+      "tokens_seen": 2458124288
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 0.0002575830524795378,
+      "loss": 0.0661,
+      "theoretical_loss": 3.4015091803439326,
+      "tokens_seen": 2458386432
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002575028085379554,
+      "loss": 0.0669,
+      "theoretical_loss": 3.4014803957023094,
+      "tokens_seen": 2458648576
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025742256459637295,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4014516149888054,
+      "tokens_seen": 2458910720
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025734232065479057,
+      "loss": 0.0626,
+      "theoretical_loss": 3.401422838202466,
+      "tokens_seen": 2459172864
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002572620767132082,
+      "loss": 0.0618,
+      "theoretical_loss": 3.401394065342336,
+      "tokens_seen": 2459435008
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002571818327716258,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4013652964074614,
+      "tokens_seen": 2459697152
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025710158883004335,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4013365313968893,
+      "tokens_seen": 2459959296
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002570213448884609,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4013077703096655,
+      "tokens_seen": 2460221440
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002569411009468785,
+      "loss": 0.065,
+      "theoretical_loss": 3.4012790131448374,
+      "tokens_seen": 2460483584
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002568608570052961,
+      "loss": 0.061,
+      "theoretical_loss": 3.4012502599014516,
+      "tokens_seen": 2460745728
+    },
+    {
+      "epoch": 0.75,
+      "objective/train/advantage_avg": -0.00022634593187831342,
+      "objective/train/docs_used": 895477,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3499811887741089,
+      "objective/train/original_loss": 1.3499810695648193,
+      "objective/train/theoretical_loss": 3.401235884750002,
+      "objective/train/tokens_used": 2481336800,
+      "objective/train/value_avg": -0.006145477294921875,
+      "objective/train/value_loss": 0.00018551845278125256,
+      "objective/train/value_max": -8.094310760498047e-05,
+      "objective/train/value_min": -0.245849609375,
+      "objective/train/value_reward_corr": 0.6423102891155956,
+      "objective/train/value_std": 0.01097869873046875,
+      "objective/train/weight_avg": 0.99985671043396,
+      "objective/train/weighted_lm_loss": 1.350335955619812,
+      "objective/train/weights_max": 1.1451069116592407,
+      "objective/train/weights_min": 0.3721874952316284,
+      "theoretical_loss": 3.401235884750002,
+      "tokens_seen": 2460876800
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002567806130637137,
+      "loss": 0.0643,
+      "theoretical_loss": 3.4012215105785564,
+      "tokens_seen": 2461007872
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002567003691221313,
+      "loss": 0.0642,
+      "theoretical_loss": 3.401192765175199,
+      "tokens_seen": 2461270016
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002566201251805489,
+      "loss": 0.0656,
+      "theoretical_loss": 3.4011640236904284,
+      "tokens_seen": 2461532160
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002565398812389665,
+      "loss": 0.062,
+      "theoretical_loss": 3.401135286123293,
+      "tokens_seen": 2461794304
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025645963729738403,
+      "loss": 0.0647,
+      "theoretical_loss": 3.401106552472841,
+      "tokens_seen": 2462056448
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025637939335580165,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4010778227381233,
+      "tokens_seen": 2462318592
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002562991494142192,
+      "loss": 0.0651,
+      "theoretical_loss": 3.4010490969181877,
+      "tokens_seen": 2462580736
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002562189054726368,
+      "loss": 0.0665,
+      "theoretical_loss": 3.401020375012085,
+      "tokens_seen": 2462842880
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025613866153105443,
+      "loss": 0.0659,
+      "theoretical_loss": 3.400991657018865,
+      "tokens_seen": 2463105024
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.000256058417589472,
+      "loss": 0.0661,
+      "theoretical_loss": 3.400962942937579,
+      "tokens_seen": 2463367168
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002559781736478896,
+      "loss": 0.0644,
+      "theoretical_loss": 3.400934232767277,
+      "tokens_seen": 2463629312
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025589792970630716,
+      "loss": 0.0629,
+      "theoretical_loss": 3.400905526507011,
+      "tokens_seen": 2463891456
+    },
+    {
+      "epoch": 0.75,
+      "objective/train/advantage_avg": -0.00027899775886908174,
+      "objective/train/docs_used": 896700,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1354491710662842,
+      "objective/train/original_loss": 1.1354490518569946,
+      "objective/train/theoretical_loss": 3.4008768241558323,
+      "objective/train/tokens_used": 2484613600,
+      "objective/train/value_avg": -0.01045989990234375,
+      "objective/train/value_loss": 0.00030195803265087306,
+      "objective/train/value_max": -1.2993812561035156e-05,
+      "objective/train/value_min": -0.337890625,
+      "objective/train/value_reward_corr": 0.8496636557732059,
+      "objective/train/value_std": 0.027252197265625,
+      "objective/train/weight_avg": 0.9998582005500793,
+      "objective/train/weighted_lm_loss": 1.1343961954116821,
+      "objective/train/weights_max": 1.1410603523254395,
+      "objective/train/weights_min": 0.369240403175354,
+      "theoretical_loss": 3.4008768241558323,
+      "tokens_seen": 2464153600
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025581768576472477,
+      "loss": 0.0647,
+      "theoretical_loss": 3.4008768241558323,
+      "tokens_seen": 2464153600
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025573744182314233,
+      "loss": 0.066,
+      "theoretical_loss": 3.4008481257127925,
+      "tokens_seen": 2464415744
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025565719788155994,
+      "loss": 0.0632,
+      "theoretical_loss": 3.400819431176944,
+      "tokens_seen": 2464677888
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025557695393997755,
+      "loss": 0.0619,
+      "theoretical_loss": 3.4007907405473405,
+      "tokens_seen": 2464940032
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002554967099983951,
+      "loss": 0.0665,
+      "theoretical_loss": 3.400762053823033,
+      "tokens_seen": 2465202176
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002554164660568127,
+      "loss": 0.064,
+      "theoretical_loss": 3.400733371003076,
+      "tokens_seen": 2465464320
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025533622211523034,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4007046920865225,
+      "tokens_seen": 2465726464
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002552559781736479,
+      "loss": 0.063,
+      "theoretical_loss": 3.4006760170724264,
+      "tokens_seen": 2465988608
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025517573423206546,
+      "loss": 0.0665,
+      "theoretical_loss": 3.4006473459598423,
+      "tokens_seen": 2466250752
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025509549029048307,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4006186787478243,
+      "tokens_seen": 2466512896
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002550152463489007,
+      "loss": 0.0657,
+      "theoretical_loss": 3.4005900154354274,
+      "tokens_seen": 2466775040
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025493500240731824,
+      "loss": 0.0653,
+      "theoretical_loss": 3.400561356021707,
+      "tokens_seen": 2467037184
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025485475846573585,
+      "loss": 0.0663,
+      "theoretical_loss": 3.400532700505719,
+      "tokens_seen": 2467299328
+    },
+    {
+      "epoch": 0.75,
+      "objective/train/advantage_avg": -0.0005027269362471998,
+      "objective/train/docs_used": 897881,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.282994031906128,
+      "objective/train/original_loss": 1.282994031906128,
+      "objective/train/theoretical_loss": 3.4005183742090788,
+      "objective/train/tokens_used": 2487890400,
+      "objective/train/value_avg": -0.0072021484375,
+      "objective/train/value_loss": 0.0002903847489506006,
+      "objective/train/value_max": -3.707408905029297e-05,
+      "objective/train/value_min": -0.595703125,
+      "objective/train/value_reward_corr": 0.7093589137672446,
+      "objective/train/value_std": 0.0144195556640625,
+      "objective/train/weight_avg": 0.9996247887611389,
+      "objective/train/weighted_lm_loss": 1.2816740274429321,
+      "objective/train/weights_max": 1.3052014112472534,
+      "objective/train/weights_min": 0.3687020242214203,
+      "theoretical_loss": 3.4005183742090788,
+      "tokens_seen": 2467430400
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025477451452415346,
+      "loss": 0.0669,
+      "theoretical_loss": 3.400504048886518,
+      "tokens_seen": 2467561472
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.000254694270582571,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4004754011631615,
+      "tokens_seen": 2467823616
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002546140266409886,
+      "loss": 0.0645,
+      "theoretical_loss": 3.4004467573347053,
+      "tokens_seen": 2468085760
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002545337826994062,
+      "loss": 0.0659,
+      "theoretical_loss": 3.400418117400206,
+      "tokens_seen": 2468347904
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002544535387578238,
+      "loss": 0.0639,
+      "theoretical_loss": 3.4003894813587223,
+      "tokens_seen": 2468610048
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025437329481624136,
+      "loss": 0.0644,
+      "theoretical_loss": 3.4003608492093096,
+      "tokens_seen": 2468872192
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.000254293050874659,
+      "loss": 0.0652,
+      "theoretical_loss": 3.400332220951027,
+      "tokens_seen": 2469134336
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002542128069330766,
+      "loss": 0.065,
+      "theoretical_loss": 3.400303596582933,
+      "tokens_seen": 2469396480
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002541325629914941,
+      "loss": 0.063,
+      "theoretical_loss": 3.4002749761040847,
+      "tokens_seen": 2469658624
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002540523190499117,
+      "loss": 0.0644,
+      "theoretical_loss": 3.400246359513542,
+      "tokens_seen": 2469920768
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002539720751083293,
+      "loss": 0.0678,
+      "theoretical_loss": 3.4002177468103643,
+      "tokens_seen": 2470182912
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025389183116674693,
+      "loss": 0.0646,
+      "theoretical_loss": 3.40018913799361,
+      "tokens_seen": 2470445056
+    },
+    {
+      "epoch": 0.75,
+      "objective/train/advantage_avg": -0.0009792648488655686,
+      "objective/train/docs_used": 899059,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.327161431312561,
+      "objective/train/original_loss": 1.3271613121032715,
+      "objective/train/theoretical_loss": 3.4001605330623397,
+      "objective/train/tokens_used": 2491167200,
+      "objective/train/value_avg": -0.01041412353515625,
+      "objective/train/value_loss": 0.00031207365100272,
+      "objective/train/value_max": -2.5272369384765625e-05,
+      "objective/train/value_min": -0.96484375,
+      "objective/train/value_reward_corr": 0.8798099044091047,
+      "objective/train/value_std": 0.0277099609375,
+      "objective/train/weight_avg": 0.9991644620895386,
+      "objective/train/weighted_lm_loss": 1.3267914056777954,
+      "objective/train/weights_max": 1.9819515943527222,
+      "objective/train/weights_min": 0.3967057168483734,
+      "theoretical_loss": 3.4001605330623397,
+      "tokens_seen": 2470707200
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002538115872251645,
+      "loss": 0.0634,
+      "theoretical_loss": 3.4001605330623397,
+      "tokens_seen": 2470707200
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002537313432835821,
+      "loss": 0.0644,
+      "theoretical_loss": 3.400131932015613,
+      "tokens_seen": 2470969344
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002536510993419997,
+      "loss": 0.0634,
+      "theoretical_loss": 3.4001033348524907,
+      "tokens_seen": 2471231488
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002535708554004173,
+      "loss": 0.0662,
+      "theoretical_loss": 3.4000747415720336,
+      "tokens_seen": 2471493632
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025349061145883483,
+      "loss": 0.0673,
+      "theoretical_loss": 3.4000461521733025,
+      "tokens_seen": 2471755776
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025341036751725244,
+      "loss": 0.0655,
+      "theoretical_loss": 3.400017566655359,
+      "tokens_seen": 2472017920
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025333012357567006,
+      "loss": 0.0647,
+      "theoretical_loss": 3.3999889850172655,
+      "tokens_seen": 2472280064
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002532498796340876,
+      "loss": 0.0642,
+      "theoretical_loss": 3.399960407258083,
+      "tokens_seen": 2472542208
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025316963569250523,
+      "loss": 0.0647,
+      "theoretical_loss": 3.3999318333768747,
+      "tokens_seen": 2472804352
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025308939175092284,
+      "loss": 0.0619,
+      "theoretical_loss": 3.3999032633727024,
+      "tokens_seen": 2473066496
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002530091478093404,
+      "loss": 0.0675,
+      "theoretical_loss": 3.3998746972446305,
+      "tokens_seen": 2473328640
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025292890386775796,
+      "loss": 0.0673,
+      "theoretical_loss": 3.3998461349917215,
+      "tokens_seen": 2473590784
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025284865992617557,
+      "loss": 0.0644,
+      "theoretical_loss": 3.3998175766130396,
+      "tokens_seen": 2473852928
+    },
+    {
+      "epoch": 0.75,
+      "objective/train/advantage_avg": 0.0003516378055792302,
+      "objective/train/docs_used": 900350,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.250746250152588,
+      "objective/train/original_loss": 1.250746250152588,
+      "objective/train/theoretical_loss": 3.399803298876241,
+      "objective/train/tokens_used": 2494444000,
+      "objective/train/value_avg": -0.00687408447265625,
+      "objective/train/value_loss": 0.00016971146396826953,
+      "objective/train/value_max": -4.1961669921875e-05,
+      "objective/train/value_min": -0.227783203125,
+      "objective/train/value_reward_corr": 0.6703260986924842,
+      "objective/train/value_std": 0.012237548828125,
+      "objective/train/weight_avg": 1.0004278421401978,
+      "objective/train/weighted_lm_loss": 1.2503207921981812,
+      "objective/train/weights_max": 1.2307718992233276,
+      "objective/train/weights_min": 0.40160250663757324,
+      "theoretical_loss": 3.399803298876241,
+      "tokens_seen": 2473984000
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002527684159845932,
+      "loss": 0.0675,
+      "theoretical_loss": 3.399789022107648,
+      "tokens_seen": 2474115072
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025268817204301074,
+      "loss": 0.0657,
+      "theoretical_loss": 3.399760471474612,
+      "tokens_seen": 2474377216
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025260792810142835,
+      "loss": 0.0654,
+      "theoretical_loss": 3.399731924712996,
+      "tokens_seen": 2474639360
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025252768415984597,
+      "loss": 0.0645,
+      "theoretical_loss": 3.3997033818218645,
+      "tokens_seen": 2474901504
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002524474402182635,
+      "loss": 0.0676,
+      "theoretical_loss": 3.3996748428002834,
+      "tokens_seen": 2475163648
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025236719627668114,
+      "loss": 0.0666,
+      "theoretical_loss": 3.3996463076473185,
+      "tokens_seen": 2475425792
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002522869523350987,
+      "loss": 0.0659,
+      "theoretical_loss": 3.399617776362035,
+      "tokens_seen": 2475687936
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002522067083935163,
+      "loss": 0.0632,
+      "theoretical_loss": 3.3995892489435007,
+      "tokens_seen": 2475950080
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025212646445193387,
+      "loss": 0.065,
+      "theoretical_loss": 3.3995607253907805,
+      "tokens_seen": 2476212224
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002520462205103515,
+      "loss": 0.0658,
+      "theoretical_loss": 3.3995322057029425,
+      "tokens_seen": 2476474368
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002519659765687691,
+      "loss": 0.0639,
+      "theoretical_loss": 3.3995036898790536,
+      "tokens_seen": 2476736512
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025188573262718665,
+      "loss": 0.0641,
+      "theoretical_loss": 3.3994751779181813,
+      "tokens_seen": 2476998656
+    },
+    {
+      "epoch": 0.75,
+      "objective/train/advantage_avg": 0.0007612823974341154,
+      "objective/train/docs_used": 901501,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2392642498016357,
+      "objective/train/original_loss": 1.2392642498016357,
+      "objective/train/theoretical_loss": 3.399446669819394,
+      "objective/train/tokens_used": 2497720800,
+      "objective/train/value_avg": -0.00814056396484375,
+      "objective/train/value_loss": 0.00025297680986113846,
+      "objective/train/value_max": -8.416175842285156e-05,
+      "objective/train/value_min": -0.34326171875,
+      "objective/train/value_reward_corr": 0.6215869301992545,
+      "objective/train/value_std": 0.012786865234375,
+      "objective/train/weight_avg": 1.000868558883667,
+      "objective/train/weighted_lm_loss": 1.240895390510559,
+      "objective/train/weights_max": 1.3520135879516602,
+      "objective/train/weights_min": 0.3843126595020294,
+      "theoretical_loss": 3.399446669819394,
+      "tokens_seen": 2477260800
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025180548868560426,
+      "loss": 0.0639,
+      "theoretical_loss": 3.399446669819394,
+      "tokens_seen": 2477260800
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002517252447440218,
+      "loss": 0.066,
+      "theoretical_loss": 3.3994181655817592,
+      "tokens_seen": 2477522944
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002516450008024394,
+      "loss": 0.0634,
+      "theoretical_loss": 3.3993896652043465,
+      "tokens_seen": 2477785088
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.000251564756860857,
+      "loss": 0.0664,
+      "theoretical_loss": 3.399361168686224,
+      "tokens_seen": 2478047232
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002514845129192746,
+      "loss": 0.0654,
+      "theoretical_loss": 3.399332676026461,
+      "tokens_seen": 2478309376
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002514042689776922,
+      "loss": 0.0686,
+      "theoretical_loss": 3.399304187224127,
+      "tokens_seen": 2478571520
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002513240250361098,
+      "loss": 0.066,
+      "theoretical_loss": 3.3992757022782927,
+      "tokens_seen": 2478833664
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002512437810945274,
+      "loss": 0.0666,
+      "theoretical_loss": 3.3992472211880274,
+      "tokens_seen": 2479095808
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025116353715294495,
+      "loss": 0.0667,
+      "theoretical_loss": 3.3992187439524018,
+      "tokens_seen": 2479357952
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002510832932113625,
+      "loss": 0.0659,
+      "theoretical_loss": 3.399190270570487,
+      "tokens_seen": 2479620096
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002510030492697801,
+      "loss": 0.0666,
+      "theoretical_loss": 3.3991618010413536,
+      "tokens_seen": 2479882240
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025092280532819773,
+      "loss": 0.0676,
+      "theoretical_loss": 3.3991333353640742,
+      "tokens_seen": 2480144384
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025084256138661534,
+      "loss": 0.0689,
+      "theoretical_loss": 3.3991048735377194,
+      "tokens_seen": 2480406528
+    },
+    {
+      "epoch": 0.75,
+      "objective/train/advantage_avg": -0.00028219091473147273,
+      "objective/train/docs_used": 902720,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4025236368179321,
+      "objective/train/original_loss": 1.4025237560272217,
+      "objective/train/theoretical_loss": 3.399090644068349,
+      "objective/train/tokens_used": 2500997600,
+      "objective/train/value_avg": -0.00506591796875,
+      "objective/train/value_loss": 0.0002538026019465178,
+      "objective/train/value_max": -1.3828277587890625e-05,
+      "objective/train/value_min": -0.50634765625,
+      "objective/train/value_reward_corr": 0.6467943607471219,
+      "objective/train/value_std": 0.011993408203125,
+      "objective/train/weight_avg": 0.999829113483429,
+      "objective/train/weighted_lm_loss": 1.4015452861785889,
+      "objective/train/weights_max": 1.1559211015701294,
+      "objective/train/weights_min": 0.3705088496208191,
+      "theoretical_loss": 3.399090644068349,
+      "tokens_seen": 2480537600
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002507623174450329,
+      "loss": 0.0663,
+      "theoretical_loss": 3.399076415561362,
+      "tokens_seen": 2480668672
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002506820735034505,
+      "loss": 0.0669,
+      "theoretical_loss": 3.399047961434074,
+      "tokens_seen": 2480930816
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025060182956186813,
+      "loss": 0.0654,
+      "theoretical_loss": 3.399019511154929,
+      "tokens_seen": 2481192960
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025052158562028563,
+      "loss": 0.0647,
+      "theoretical_loss": 3.398991064722999,
+      "tokens_seen": 2481455104
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025044134167870324,
+      "loss": 0.0676,
+      "theoretical_loss": 3.3989626221373586,
+      "tokens_seen": 2481717248
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025036109773712086,
+      "loss": 0.0657,
+      "theoretical_loss": 3.3989341833970803,
+      "tokens_seen": 2481979392
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025028085379553847,
+      "loss": 0.0642,
+      "theoretical_loss": 3.398905748501239,
+      "tokens_seen": 2482241536
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025020060985395603,
+      "loss": 0.0657,
+      "theoretical_loss": 3.3988773174489086,
+      "tokens_seen": 2482503680
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025012036591237364,
+      "loss": 0.0667,
+      "theoretical_loss": 3.3988488902391643,
+      "tokens_seen": 2482765824
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00025004012197079125,
+      "loss": 0.0651,
+      "theoretical_loss": 3.3988204668710806,
+      "tokens_seen": 2483027968
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002499598780292088,
+      "loss": 0.0648,
+      "theoretical_loss": 3.398792047343733,
+      "tokens_seen": 2483290112
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002498796340876264,
+      "loss": 0.0669,
+      "theoretical_loss": 3.3987636316561978,
+      "tokens_seen": 2483552256
+    },
+    {
+      "epoch": 0.75,
+      "objective/train/advantage_avg": 0.0003041817108169198,
+      "objective/train/docs_used": 903880,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3639494180679321,
+      "objective/train/original_loss": 1.3639495372772217,
+      "objective/train/theoretical_loss": 3.3987352198075502,
+      "objective/train/tokens_used": 2504274400,
+      "objective/train/value_avg": -0.005992889404296875,
+      "objective/train/value_loss": 0.00011486706353025511,
+      "objective/train/value_max": -3.3736228942871094e-05,
+      "objective/train/value_min": -0.3359375,
+      "objective/train/value_reward_corr": 0.7736587933929978,
+      "objective/train/value_std": 0.01282501220703125,
+      "objective/train/weight_avg": 1.0003571510314941,
+      "objective/train/weighted_lm_loss": 1.36470627784729,
+      "objective/train/weights_max": 1.13873291015625,
+      "objective/train/weights_min": 0.3690531253814697,
+      "theoretical_loss": 3.3987352198075502,
+      "tokens_seen": 2483814400
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.000249799390146044,
+      "loss": 0.0663,
+      "theoretical_loss": 3.3987352198075502,
+      "tokens_seen": 2483814400
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024971914620446154,
+      "loss": 0.0667,
+      "theoretical_loss": 3.3987068117968664,
+      "tokens_seen": 2484076544
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024963890226287915,
+      "loss": 0.0635,
+      "theoretical_loss": 3.3986784076232235,
+      "tokens_seen": 2484338688
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024955865832129677,
+      "loss": 0.0625,
+      "theoretical_loss": 3.3986500072856987,
+      "tokens_seen": 2484600832
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002494784143797143,
+      "loss": 0.0642,
+      "theoretical_loss": 3.398621610783368,
+      "tokens_seen": 2484862976
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024939817043813194,
+      "loss": 0.0666,
+      "theoretical_loss": 3.3985932181153107,
+      "tokens_seen": 2485125120
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024931792649654955,
+      "loss": 0.0654,
+      "theoretical_loss": 3.398564829280603,
+      "tokens_seen": 2485387264
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002492376825549671,
+      "loss": 0.0648,
+      "theoretical_loss": 3.3985364442783244,
+      "tokens_seen": 2485649408
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024915743861338467,
+      "loss": 0.0654,
+      "theoretical_loss": 3.398508063107553,
+      "tokens_seen": 2485911552
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002490771946718023,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3984796857673674,
+      "tokens_seen": 2486173696
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002489969507302199,
+      "loss": 0.0653,
+      "theoretical_loss": 3.3984513122568467,
+      "tokens_seen": 2486435840
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024891670678863745,
+      "loss": 0.0677,
+      "theoretical_loss": 3.3984229425750714,
+      "tokens_seen": 2486697984
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024883646284705506,
+      "loss": 0.0654,
+      "theoretical_loss": 3.3983945767211203,
+      "tokens_seen": 2486960128
+    },
+    {
+      "epoch": 0.75,
+      "objective/train/advantage_avg": -8.536610403098166e-05,
+      "objective/train/docs_used": 905256,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3416714668273926,
+      "objective/train/original_loss": 1.3416712284088135,
+      "objective/train/theoretical_loss": 3.398380395229291,
+      "objective/train/tokens_used": 2507551200,
+      "objective/train/value_avg": -0.00720977783203125,
+      "objective/train/value_loss": 0.00013160381058696657,
+      "objective/train/value_max": -3.451108932495117e-05,
+      "objective/train/value_min": -0.469970703125,
+      "objective/train/value_reward_corr": 0.7648471489080964,
+      "objective/train/value_std": 0.01303863525390625,
+      "objective/train/weight_avg": 0.9999786019325256,
+      "objective/train/weighted_lm_loss": 1.3416770696640015,
+      "objective/train/weights_max": 1.2009410858154297,
+      "objective/train/weights_min": 0.5534672141075134,
+      "theoretical_loss": 3.398380395229291,
+      "tokens_seen": 2487091200
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002487562189054726,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3983662146940734,
+      "tokens_seen": 2487222272
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024867597496389023,
+      "loss": 0.0661,
+      "theoretical_loss": 3.398337856493012,
+      "tokens_seen": 2487484416
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002485957310223078,
+      "loss": 0.0663,
+      "theoretical_loss": 3.3983095021170158,
+      "tokens_seen": 2487746560
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002485154870807254,
+      "loss": 0.0664,
+      "theoretical_loss": 3.398281151565167,
+      "tokens_seen": 2488008704
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.000248435243139143,
+      "loss": 0.0695,
+      "theoretical_loss": 3.398252804836546,
+      "tokens_seen": 2488270848
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002483549991975606,
+      "loss": 0.0662,
+      "theoretical_loss": 3.3982244619302358,
+      "tokens_seen": 2488532992
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002482747552559782,
+      "loss": 0.0677,
+      "theoretical_loss": 3.398196122845317,
+      "tokens_seen": 2488795136
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024819451131439575,
+      "loss": 0.0673,
+      "theoretical_loss": 3.3981677875808725,
+      "tokens_seen": 2489057280
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024811426737281336,
+      "loss": 0.0649,
+      "theoretical_loss": 3.3981394561359854,
+      "tokens_seen": 2489319424
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002480340234312309,
+      "loss": 0.0682,
+      "theoretical_loss": 3.398111128509738,
+      "tokens_seen": 2489581568
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024795377948964853,
+      "loss": 0.0651,
+      "theoretical_loss": 3.3980828047012146,
+      "tokens_seen": 2489843712
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024787353554806614,
+      "loss": 0.0636,
+      "theoretical_loss": 3.3980544847094976,
+      "tokens_seen": 2490105856
+    },
+    {
+      "epoch": 0.75,
+      "objective/train/advantage_avg": 0.00030044748564250767,
+      "objective/train/docs_used": 906478,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2353644371032715,
+      "objective/train/original_loss": 1.2353644371032715,
+      "objective/train/theoretical_loss": 3.3980261685336712,
+      "objective/train/tokens_used": 2510828000,
+      "objective/train/value_avg": -0.0033092498779296875,
+      "objective/train/value_loss": 0.00019104246166534722,
+      "objective/train/value_max": -3.647804260253906e-05,
+      "objective/train/value_min": -0.28515625,
+      "objective/train/value_reward_corr": 0.4723460212474383,
+      "objective/train/value_std": 0.006641387939453125,
+      "objective/train/weight_avg": 1.0003799200057983,
+      "objective/train/weighted_lm_loss": 1.2356609106063843,
+      "objective/train/weights_max": 1.205199956893921,
+      "objective/train/weights_min": 0.3687315881252289,
+      "theoretical_loss": 3.3980261685336712,
+      "tokens_seen": 2490368000
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002477932916064837,
+      "loss": 0.0656,
+      "theoretical_loss": 3.3980261685336712,
+      "tokens_seen": 2490368000
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002477130476649013,
+      "loss": 0.0636,
+      "theoretical_loss": 3.39799785617282,
+      "tokens_seen": 2490630144
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.00024763280372331887,
+      "loss": 0.0663,
+      "theoretical_loss": 3.3979695476260288,
+      "tokens_seen": 2490892288
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 0.0002475525597817365,
+      "loss": 0.0668,
+      "theoretical_loss": 3.397941242892382,
+      "tokens_seen": 2491154432
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024747231584015404,
+      "loss": 0.0696,
+      "theoretical_loss": 3.397912941970965,
+      "tokens_seen": 2491416576
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024739207189857166,
+      "loss": 0.0677,
+      "theoretical_loss": 3.3978846448608633,
+      "tokens_seen": 2491678720
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024731182795698927,
+      "loss": 0.0668,
+      "theoretical_loss": 3.3978563515611624,
+      "tokens_seen": 2491940864
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024723158401540683,
+      "loss": 0.0676,
+      "theoretical_loss": 3.397828062070949,
+      "tokens_seen": 2492203008
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024715134007382444,
+      "loss": 0.0638,
+      "theoretical_loss": 3.3977997763893093,
+      "tokens_seen": 2492465152
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.000247071096132242,
+      "loss": 0.069,
+      "theoretical_loss": 3.3977714945153297,
+      "tokens_seen": 2492727296
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002469908521906596,
+      "loss": 0.0652,
+      "theoretical_loss": 3.397743216448098,
+      "tokens_seen": 2492989440
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002469106082490772,
+      "loss": 0.0668,
+      "theoretical_loss": 3.3977149421867012,
+      "tokens_seen": 2493251584
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002468303643074948,
+      "loss": 0.0684,
+      "theoretical_loss": 3.397686671730227,
+      "tokens_seen": 2493513728
+    },
+    {
+      "epoch": 0.76,
+      "objective/train/advantage_avg": 0.0004943067324347794,
+      "objective/train/docs_used": 907814,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2799855470657349,
+      "objective/train/original_loss": 1.2799855470657349,
+      "objective/train/theoretical_loss": 3.3976725379285506,
+      "objective/train/tokens_used": 2514104800,
+      "objective/train/value_avg": -0.0073394775390625,
+      "objective/train/value_loss": 0.00021812194609083235,
+      "objective/train/value_max": -1.4722347259521484e-05,
+      "objective/train/value_min": -0.232177734375,
+      "objective/train/value_reward_corr": 0.6116615900831515,
+      "objective/train/value_std": 0.01178741455078125,
+      "objective/train/weight_avg": 1.0005911588668823,
+      "objective/train/weighted_lm_loss": 1.2801631689071655,
+      "objective/train/weights_max": 1.1481863260269165,
+      "objective/train/weights_min": 0.3702644109725952,
+      "theoretical_loss": 3.3976725379285506,
+      "tokens_seen": 2493644800
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002467501203659124,
+      "loss": 0.0656,
+      "theoretical_loss": 3.3976584050777636,
+      "tokens_seen": 2493775872
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024666987642432995,
+      "loss": 0.0645,
+      "theoretical_loss": 3.397630142228399,
+      "tokens_seen": 2494038016
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024658963248274757,
+      "loss": 0.0657,
+      "theoretical_loss": 3.3976018831812222,
+      "tokens_seen": 2494300160
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002465093885411651,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3975736279353224,
+      "tokens_seen": 2494562304
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024642914459958274,
+      "loss": 0.063,
+      "theoretical_loss": 3.3975453764897883,
+      "tokens_seen": 2494824448
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024634890065800035,
+      "loss": 0.0651,
+      "theoretical_loss": 3.3975171288437096,
+      "tokens_seen": 2495086592
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002462686567164179,
+      "loss": 0.0644,
+      "theoretical_loss": 3.397488884996177,
+      "tokens_seen": 2495348736
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002461884127748355,
+      "loss": 0.0657,
+      "theoretical_loss": 3.3974606449462796,
+      "tokens_seen": 2495610880
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002461081688332531,
+      "loss": 0.0655,
+      "theoretical_loss": 3.3974324086931085,
+      "tokens_seen": 2495873024
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002460279248916707,
+      "loss": 0.0662,
+      "theoretical_loss": 3.3974041762357547,
+      "tokens_seen": 2496135168
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024594768095008825,
+      "loss": 0.0658,
+      "theoretical_loss": 3.397375947573309,
+      "tokens_seen": 2496397312
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024586743700850586,
+      "loss": 0.0647,
+      "theoretical_loss": 3.397347722704863,
+      "tokens_seen": 2496659456
+    },
+    {
+      "epoch": 0.76,
+      "objective/train/advantage_avg": 0.0017170101637020707,
+      "objective/train/docs_used": 908925,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3087087869644165,
+      "objective/train/original_loss": 1.308708667755127,
+      "objective/train/theoretical_loss": 3.397319501629509,
+      "objective/train/tokens_used": 2517381600,
+      "objective/train/value_avg": -0.011444091796875,
+      "objective/train/value_loss": 0.00022468052338808775,
+      "objective/train/value_max": -0.00010311603546142578,
+      "objective/train/value_min": -0.313720703125,
+      "objective/train/value_reward_corr": 0.7452039655829534,
+      "objective/train/value_std": 0.0175933837890625,
+      "objective/train/weight_avg": 1.0018271207809448,
+      "objective/train/weighted_lm_loss": 1.3122689723968506,
+      "objective/train/weights_max": 1.1817450523376465,
+      "objective/train/weights_min": 0.718687117099762,
+      "theoretical_loss": 3.397319501629509,
+      "tokens_seen": 2496921600
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002457871930669235,
+      "loss": 0.0665,
+      "theoretical_loss": 3.397319501629509,
+      "tokens_seen": 2496921600
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024570694912534103,
+      "loss": 0.0645,
+      "theoretical_loss": 3.397291284346338,
+      "tokens_seen": 2497183744
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024562670518375865,
+      "loss": 0.0652,
+      "theoretical_loss": 3.397263070854444,
+      "tokens_seen": 2497445888
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002455464612421762,
+      "loss": 0.066,
+      "theoretical_loss": 3.397234861152918,
+      "tokens_seen": 2497708032
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002454662173005938,
+      "loss": 0.0661,
+      "theoretical_loss": 3.397206655240854,
+      "tokens_seen": 2497970176
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002453859733590114,
+      "loss": 0.0638,
+      "theoretical_loss": 3.3971784531173457,
+      "tokens_seen": 2498232320
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.000245305729417429,
+      "loss": 0.0638,
+      "theoretical_loss": 3.397150254781486,
+      "tokens_seen": 2498494464
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002452254854758466,
+      "loss": 0.0672,
+      "theoretical_loss": 3.3971220602323693,
+      "tokens_seen": 2498756608
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024514524153426416,
+      "loss": 0.0666,
+      "theoretical_loss": 3.3970938694690895,
+      "tokens_seen": 2499018752
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024506499759268177,
+      "loss": 0.0636,
+      "theoretical_loss": 3.3970656824907417,
+      "tokens_seen": 2499280896
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024498475365109933,
+      "loss": 0.0647,
+      "theoretical_loss": 3.3970374992964207,
+      "tokens_seen": 2499543040
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024490450970951694,
+      "loss": 0.0664,
+      "theoretical_loss": 3.3970093198852216,
+      "tokens_seen": 2499805184
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024482426576793456,
+      "loss": 0.0669,
+      "theoretical_loss": 3.3969811442562396,
+      "tokens_seen": 2500067328
+    },
+    {
+      "epoch": 0.76,
+      "objective/train/advantage_avg": 3.514278068905696e-05,
+      "objective/train/docs_used": 910166,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2338531017303467,
+      "objective/train/original_loss": 1.233852744102478,
+      "objective/train/theoretical_loss": 3.396967057859798,
+      "objective/train/tokens_used": 2520658400,
+      "objective/train/value_avg": -0.00606536865234375,
+      "objective/train/value_loss": 0.00012258593051228672,
+      "objective/train/value_max": -2.8192996978759766e-05,
+      "objective/train/value_min": -0.167236328125,
+      "objective/train/value_reward_corr": 0.6512071554286463,
+      "objective/train/value_std": 0.00945281982421875,
+      "objective/train/weight_avg": 1.0000883340835571,
+      "objective/train/weighted_lm_loss": 1.234293818473816,
+      "objective/train/weights_max": 1.0883424282073975,
+      "objective/train/weights_min": 0.37021923065185547,
+      "theoretical_loss": 3.396967057859798,
+      "tokens_seen": 2500198400
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002447440218263521,
+      "loss": 0.0627,
+      "theoretical_loss": 3.396952972408571,
+      "tokens_seen": 2500329472
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002446637778847697,
+      "loss": 0.0656,
+      "theoretical_loss": 3.3969248043413125,
+      "tokens_seen": 2500591616
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002445835339431873,
+      "loss": 0.0641,
+      "theoretical_loss": 3.3968966400535594,
+      "tokens_seen": 2500853760
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002445032900016049,
+      "loss": 0.0647,
+      "theoretical_loss": 3.396868479544409,
+      "tokens_seen": 2501115904
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024442304606002246,
+      "loss": 0.0629,
+      "theoretical_loss": 3.3968403228129587,
+      "tokens_seen": 2501378048
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024434280211844007,
+      "loss": 0.0613,
+      "theoretical_loss": 3.396812169858306,
+      "tokens_seen": 2501640192
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002442625581768577,
+      "loss": 0.0648,
+      "theoretical_loss": 3.396784020679547,
+      "tokens_seen": 2501902336
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024418231423527524,
+      "loss": 0.0662,
+      "theoretical_loss": 3.3967558752757823,
+      "tokens_seen": 2502164480
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024410207029369282,
+      "loss": 0.0682,
+      "theoretical_loss": 3.396727733646108,
+      "tokens_seen": 2502426624
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024402182635211044,
+      "loss": 0.0652,
+      "theoretical_loss": 3.396699595789624,
+      "tokens_seen": 2502688768
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024394158241052802,
+      "loss": 0.0667,
+      "theoretical_loss": 3.396671461705429,
+      "tokens_seen": 2502950912
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024386133846894558,
+      "loss": 0.0652,
+      "theoretical_loss": 3.396643331392622,
+      "tokens_seen": 2503213056
+    },
+    {
+      "epoch": 0.76,
+      "objective/train/advantage_avg": -0.0003344157012179494,
+      "objective/train/docs_used": 911421,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.414215326309204,
+      "objective/train/original_loss": 1.414215326309204,
+      "objective/train/theoretical_loss": 3.3966152048503027,
+      "objective/train/tokens_used": 2523935200,
+      "objective/train/value_avg": -0.010101318359375,
+      "objective/train/value_loss": 0.0001809104287531227,
+      "objective/train/value_max": -5.739927291870117e-05,
+      "objective/train/value_min": -0.25341796875,
+      "objective/train/value_reward_corr": 0.7323826354438097,
+      "objective/train/value_std": 0.01412200927734375,
+      "objective/train/weight_avg": 0.999748945236206,
+      "objective/train/weighted_lm_loss": 1.4127529859542847,
+      "objective/train/weights_max": 1.0985801219940186,
+      "objective/train/weights_min": 0.38993605971336365,
+      "theoretical_loss": 3.3966152048503027,
+      "tokens_seen": 2503475200
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002437810945273632,
+      "loss": 0.0665,
+      "theoretical_loss": 3.3966152048503027,
+      "tokens_seen": 2503475200
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024370085058578078,
+      "loss": 0.0629,
+      "theoretical_loss": 3.396587082077571,
+      "tokens_seen": 2503737344
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024362060664419836,
+      "loss": 0.0672,
+      "theoretical_loss": 3.3965589630735273,
+      "tokens_seen": 2503999488
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024354036270261595,
+      "loss": 0.0623,
+      "theoretical_loss": 3.3965308478372718,
+      "tokens_seen": 2504261632
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024346011876103356,
+      "loss": 0.0652,
+      "theoretical_loss": 3.3965027363679052,
+      "tokens_seen": 2504523776
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024337987481945115,
+      "loss": 0.0652,
+      "theoretical_loss": 3.3964746286645293,
+      "tokens_seen": 2504785920
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002432996308778687,
+      "loss": 0.063,
+      "theoretical_loss": 3.3964465247262448,
+      "tokens_seen": 2505048064
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024321938693628632,
+      "loss": 0.0649,
+      "theoretical_loss": 3.3964184245521536,
+      "tokens_seen": 2505310208
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002431391429947039,
+      "loss": 0.0674,
+      "theoretical_loss": 3.396390328141358,
+      "tokens_seen": 2505572352
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024305889905312152,
+      "loss": 0.0655,
+      "theoretical_loss": 3.3963622354929597,
+      "tokens_seen": 2505834496
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024297865511153908,
+      "loss": 0.0653,
+      "theoretical_loss": 3.396334146606062,
+      "tokens_seen": 2506096640
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024289841116995666,
+      "loss": 0.0632,
+      "theoretical_loss": 3.396306061479768,
+      "tokens_seen": 2506358784
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024281816722837427,
+      "loss": 0.0671,
+      "theoretical_loss": 3.3962779801131804,
+      "tokens_seen": 2506620928
+    },
+    {
+      "epoch": 0.76,
+      "objective/train/advantage_avg": -7.992433529580012e-05,
+      "objective/train/docs_used": 912553,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.374509572982788,
+      "objective/train/original_loss": 1.374509334564209,
+      "objective/train/theoretical_loss": 3.396263940839497,
+      "objective/train/tokens_used": 2527212000,
+      "objective/train/value_avg": -0.0071868896484375,
+      "objective/train/value_loss": 0.0002604563778731972,
+      "objective/train/value_max": -2.5451183319091797e-05,
+      "objective/train/value_min": -0.70751953125,
+      "objective/train/value_reward_corr": 0.7040844074409639,
+      "objective/train/value_std": 0.017181396484375,
+      "objective/train/weight_avg": 1.0000439882278442,
+      "objective/train/weighted_lm_loss": 1.3741766214370728,
+      "objective/train/weights_max": 1.8518978357315063,
+      "objective/train/weights_min": 0.4019213020801544,
+      "theoretical_loss": 3.396263940839497,
+      "tokens_seen": 2506752000
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024273792328679183,
+      "loss": 0.0667,
+      "theoretical_loss": 3.3962499025054034,
+      "tokens_seen": 2506883072
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024265767934520945,
+      "loss": 0.0638,
+      "theoretical_loss": 3.3962218286555403,
+      "tokens_seen": 2507145216
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024257743540362703,
+      "loss": 0.0623,
+      "theoretical_loss": 3.3961937585626956,
+      "tokens_seen": 2507407360
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024249719146204464,
+      "loss": 0.0656,
+      "theoretical_loss": 3.3961656922259738,
+      "tokens_seen": 2507669504
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002424169475204622,
+      "loss": 0.0665,
+      "theoretical_loss": 3.3961376296444796,
+      "tokens_seen": 2507931648
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002423367035788798,
+      "loss": 0.0664,
+      "theoretical_loss": 3.396109570817318,
+      "tokens_seen": 2508193792
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002422564596372974,
+      "loss": 0.0654,
+      "theoretical_loss": 3.3960815157435946,
+      "tokens_seen": 2508455936
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024217621569571499,
+      "loss": 0.0652,
+      "theoretical_loss": 3.3960534644224154,
+      "tokens_seen": 2508718080
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024209597175413257,
+      "loss": 0.0671,
+      "theoretical_loss": 3.3960254168528863,
+      "tokens_seen": 2508980224
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024201572781255016,
+      "loss": 0.0622,
+      "theoretical_loss": 3.395997373034113,
+      "tokens_seen": 2509242368
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024193548387096774,
+      "loss": 0.0686,
+      "theoretical_loss": 3.395969332965203,
+      "tokens_seen": 2509504512
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024185523992938533,
+      "loss": 0.065,
+      "theoretical_loss": 3.3959412966452627,
+      "tokens_seen": 2509766656
+    },
+    {
+      "epoch": 0.76,
+      "objective/train/advantage_avg": -0.0007175038335844874,
+      "objective/train/docs_used": 913305,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4248614311218262,
+      "objective/train/original_loss": 1.4248614311218262,
+      "objective/train/theoretical_loss": 3.3959132640733998,
+      "objective/train/tokens_used": 2530488800,
+      "objective/train/value_avg": -0.00928497314453125,
+      "objective/train/value_loss": 0.000361848360626027,
+      "objective/train/value_max": -5.4776668548583984e-05,
+      "objective/train/value_min": -0.84033203125,
+      "objective/train/value_reward_corr": 0.7510871829901664,
+      "objective/train/value_std": 0.0186309814453125,
+      "objective/train/weight_avg": 0.9994414448738098,
+      "objective/train/weighted_lm_loss": 1.4235568046569824,
+      "objective/train/weights_max": 1.756426453590393,
+      "objective/train/weights_min": 0.37235790491104126,
+      "theoretical_loss": 3.3959132640733998,
+      "tokens_seen": 2510028800
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002417749959878029,
+      "loss": 0.0658,
+      "theoretical_loss": 3.3959132640733998,
+      "tokens_seen": 2510028800
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024169475204622053,
+      "loss": 0.0667,
+      "theoretical_loss": 3.395885235248721,
+      "tokens_seen": 2510290944
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002416145081046381,
+      "loss": 0.0672,
+      "theoretical_loss": 3.3958572101703353,
+      "tokens_seen": 2510553088
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002415342641630557,
+      "loss": 0.0627,
+      "theoretical_loss": 3.3958291888373506,
+      "tokens_seen": 2510815232
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024145402022147328,
+      "loss": 0.0654,
+      "theoretical_loss": 3.3958011712488747,
+      "tokens_seen": 2511077376
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024137377627989087,
+      "loss": 0.0667,
+      "theoretical_loss": 3.395773157404017,
+      "tokens_seen": 2511339520
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024129353233830848,
+      "loss": 0.069,
+      "theoretical_loss": 3.395745147301887,
+      "tokens_seen": 2511601664
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024121328839672604,
+      "loss": 0.066,
+      "theoretical_loss": 3.395717140941593,
+      "tokens_seen": 2511863808
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024113304445514365,
+      "loss": 0.0667,
+      "theoretical_loss": 3.395689138322245,
+      "tokens_seen": 2512125952
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024105280051356124,
+      "loss": 0.0656,
+      "theoretical_loss": 3.3956611394429537,
+      "tokens_seen": 2512388096
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002409725565719788,
+      "loss": 0.0639,
+      "theoretical_loss": 3.395633144302829,
+      "tokens_seen": 2512650240
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002408923126303964,
+      "loss": 0.0654,
+      "theoretical_loss": 3.3956051529009814,
+      "tokens_seen": 2512912384
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.000240812068688814,
+      "loss": 0.0654,
+      "theoretical_loss": 3.3955771652365216,
+      "tokens_seen": 2513174528
+    },
+    {
+      "epoch": 0.76,
+      "objective/train/advantage_avg": 0.0003988731768913567,
+      "objective/train/docs_used": 914593,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4558669328689575,
+      "objective/train/original_loss": 1.4558669328689575,
+      "objective/train/theoretical_loss": 3.395563172805535,
+      "objective/train/tokens_used": 2533765600,
+      "objective/train/value_avg": -0.004772186279296875,
+      "objective/train/value_loss": 8.095076191239059e-05,
+      "objective/train/value_max": -1.722574234008789e-05,
+      "objective/train/value_min": -0.1739501953125,
+      "objective/train/value_reward_corr": 0.6699456283607889,
+      "objective/train/value_std": 0.00800323486328125,
+      "objective/train/weight_avg": 1.0004359483718872,
+      "objective/train/weighted_lm_loss": 1.4565863609313965,
+      "objective/train/weights_max": 1.1899962425231934,
+      "objective/train/weights_min": 0.4030267000198364,
+      "theoretical_loss": 3.395563172805535,
+      "tokens_seen": 2513305600
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002407318247472316,
+      "loss": 0.0668,
+      "theoretical_loss": 3.395549181308562,
+      "tokens_seen": 2513436672
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024065158080564916,
+      "loss": 0.0684,
+      "theoretical_loss": 3.3955212011162126,
+      "tokens_seen": 2513698816
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024057133686406678,
+      "loss": 0.0654,
+      "theoretical_loss": 3.395493224658586,
+      "tokens_seen": 2513960960
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024049109292248436,
+      "loss": 0.0628,
+      "theoretical_loss": 3.395465251934794,
+      "tokens_seen": 2514223104
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024041084898090195,
+      "loss": 0.0636,
+      "theoretical_loss": 3.39543728294395,
+      "tokens_seen": 2514485248
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024033060503931953,
+      "loss": 0.0631,
+      "theoretical_loss": 3.395409317685165,
+      "tokens_seen": 2514747392
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024025036109773712,
+      "loss": 0.066,
+      "theoretical_loss": 3.395381356157554,
+      "tokens_seen": 2515009536
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024017011715615473,
+      "loss": 0.0675,
+      "theoretical_loss": 3.3953533983602293,
+      "tokens_seen": 2515271680
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024008987321457232,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3953254442923044,
+      "tokens_seen": 2515533824
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00024000962927298988,
+      "loss": 0.0672,
+      "theoretical_loss": 3.3952974939528935,
+      "tokens_seen": 2515795968
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002399293853314075,
+      "loss": 0.0639,
+      "theoretical_loss": 3.395269547341111,
+      "tokens_seen": 2516058112
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023984914138982507,
+      "loss": 0.0644,
+      "theoretical_loss": 3.3952416044560714,
+      "tokens_seen": 2516320256
+    },
+    {
+      "epoch": 0.76,
+      "objective/train/advantage_avg": -0.0004222899442538619,
+      "objective/train/docs_used": 916228,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4156543016433716,
+      "objective/train/original_loss": 1.4156540632247925,
+      "objective/train/theoretical_loss": 3.3952136652968896,
+      "objective/train/tokens_used": 2537042400,
+      "objective/train/value_avg": -0.0073394775390625,
+      "objective/train/value_loss": 0.00018496091070119292,
+      "objective/train/value_max": -5.02467155456543e-05,
+      "objective/train/value_min": -0.493408203125,
+      "objective/train/value_reward_corr": 0.6832515692663174,
+      "objective/train/value_std": 0.01263427734375,
+      "objective/train/weight_avg": 0.9996618032455444,
+      "objective/train/weighted_lm_loss": 1.4154560565948486,
+      "objective/train/weights_max": 1.4198497533798218,
+      "objective/train/weights_min": 0.3722045421600342,
+      "theoretical_loss": 3.3952136652968896,
+      "tokens_seen": 2516582400
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023976889744824266,
+      "loss": 0.0687,
+      "theoretical_loss": 3.3952136652968896,
+      "tokens_seen": 2516582400
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023968865350666024,
+      "loss": 0.065,
+      "theoretical_loss": 3.3951857298626806,
+      "tokens_seen": 2516844544
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023960840956507786,
+      "loss": 0.0648,
+      "theoretical_loss": 3.3951577981525602,
+      "tokens_seen": 2517106688
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023952816562349544,
+      "loss": 0.0655,
+      "theoretical_loss": 3.395129870165644,
+      "tokens_seen": 2517368832
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.000239447921681913,
+      "loss": 0.0666,
+      "theoretical_loss": 3.3951019459010476,
+      "tokens_seen": 2517630976
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002393676777403306,
+      "loss": 0.0636,
+      "theoretical_loss": 3.3950740253578875,
+      "tokens_seen": 2517893120
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002392874337987482,
+      "loss": 0.065,
+      "theoretical_loss": 3.3950461085352814,
+      "tokens_seen": 2518155264
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002392071898571658,
+      "loss": 0.0659,
+      "theoretical_loss": 3.3950181954323453,
+      "tokens_seen": 2518417408
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023912694591558337,
+      "loss": 0.0654,
+      "theoretical_loss": 3.3949902860481966,
+      "tokens_seen": 2518679552
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023904670197400096,
+      "loss": 0.0653,
+      "theoretical_loss": 3.394962380381953,
+      "tokens_seen": 2518941696
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023896645803241857,
+      "loss": 0.066,
+      "theoretical_loss": 3.3949344784327327,
+      "tokens_seen": 2519203840
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023888621409083613,
+      "loss": 0.0664,
+      "theoretical_loss": 3.3949065801996534,
+      "tokens_seen": 2519465984
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023880597014925374,
+      "loss": 0.0631,
+      "theoretical_loss": 3.3948786856818334,
+      "tokens_seen": 2519728128
+    },
+    {
+      "epoch": 0.76,
+      "objective/train/advantage_avg": 0.0014712244737893343,
+      "objective/train/docs_used": 917453,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2721487283706665,
+      "objective/train/original_loss": 1.2721487283706665,
+      "objective/train/theoretical_loss": 3.394864739815871,
+      "objective/train/tokens_used": 2540319200,
+      "objective/train/value_avg": -0.0135040283203125,
+      "objective/train/value_loss": 0.0003676135966088623,
+      "objective/train/value_max": -2.7120113372802734e-05,
+      "objective/train/value_min": -0.6787109375,
+      "objective/train/value_reward_corr": 0.78297990649481,
+      "objective/train/value_std": 0.025665283203125,
+      "objective/train/weight_avg": 1.0016452074050903,
+      "objective/train/weighted_lm_loss": 1.2733814716339111,
+      "objective/train/weights_max": 1.8618712425231934,
+      "objective/train/weights_min": 0.4023630917072296,
+      "theoretical_loss": 3.394864739815871,
+      "tokens_seen": 2519859200
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023872572620767132,
+      "loss": 0.0645,
+      "theoretical_loss": 3.3948507948783924,
+      "tokens_seen": 2519990272
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023864548226608894,
+      "loss": 0.0634,
+      "theoretical_loss": 3.3948229077884484,
+      "tokens_seen": 2520252416
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002385652383245065,
+      "loss": 0.0659,
+      "theoretical_loss": 3.394795024411122,
+      "tokens_seen": 2520514560
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023848499438292408,
+      "loss": 0.0665,
+      "theoretical_loss": 3.394767144745532,
+      "tokens_seen": 2520776704
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002384047504413417,
+      "loss": 0.0691,
+      "theoretical_loss": 3.394739268790798,
+      "tokens_seen": 2521038848
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023832450649975928,
+      "loss": 0.0644,
+      "theoretical_loss": 3.3947113965460414,
+      "tokens_seen": 2521300992
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023824426255817686,
+      "loss": 0.0675,
+      "theoretical_loss": 3.3946835280103826,
+      "tokens_seen": 2521563136
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023816401861659445,
+      "loss": 0.066,
+      "theoretical_loss": 3.3946556631829417,
+      "tokens_seen": 2521825280
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023808377467501204,
+      "loss": 0.0669,
+      "theoretical_loss": 3.3946278020628404,
+      "tokens_seen": 2522087424
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023800353073342962,
+      "loss": 0.0642,
+      "theoretical_loss": 3.3945999446492,
+      "tokens_seen": 2522349568
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002379232867918472,
+      "loss": 0.0653,
+      "theoretical_loss": 3.3945720909411428,
+      "tokens_seen": 2522611712
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023784304285026482,
+      "loss": 0.0654,
+      "theoretical_loss": 3.3945442409377904,
+      "tokens_seen": 2522873856
+    },
+    {
+      "epoch": 0.76,
+      "objective/train/advantage_avg": 0.0011498106177896261,
+      "objective/train/docs_used": 918647,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3918640613555908,
+      "objective/train/original_loss": 1.3918638229370117,
+      "objective/train/theoretical_loss": 3.394516394638265,
+      "objective/train/tokens_used": 2543596000,
+      "objective/train/value_avg": -0.00826263427734375,
+      "objective/train/value_loss": 0.0001964410621440038,
+      "objective/train/value_max": -8.547306060791016e-05,
+      "objective/train/value_min": -0.39306640625,
+      "objective/train/value_reward_corr": 0.7096400416725087,
+      "objective/train/value_std": 0.01409912109375,
+      "objective/train/weight_avg": 1.0012379884719849,
+      "objective/train/weighted_lm_loss": 1.3931630849838257,
+      "objective/train/weights_max": 1.191449761390686,
+      "objective/train/weights_min": 0.3797193765640259,
+      "theoretical_loss": 3.394516394638265,
+      "tokens_seen": 2523136000
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002377627989086824,
+      "loss": 0.0674,
+      "theoretical_loss": 3.394516394638265,
+      "tokens_seen": 2523136000
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.0002376825549671,
+      "loss": 0.0643,
+      "theoretical_loss": 3.3944885520416896,
+      "tokens_seen": 2523398144
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023760231102551758,
+      "loss": 0.0681,
+      "theoretical_loss": 3.3944607131471876,
+      "tokens_seen": 2523660288
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023752206708393516,
+      "loss": 0.0675,
+      "theoretical_loss": 3.3944328779538813,
+      "tokens_seen": 2523922432
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 0.00023744182314235277,
+      "loss": 0.0638,
+      "theoretical_loss": 3.3944050464608955,
+      "tokens_seen": 2524184576
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023736157920077033,
+      "loss": 0.0652,
+      "theoretical_loss": 3.394377218667353,
+      "tokens_seen": 2524446720
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023728133525918795,
+      "loss": 0.0628,
+      "theoretical_loss": 3.3943493945723784,
+      "tokens_seen": 2524708864
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023720109131760553,
+      "loss": 0.0677,
+      "theoretical_loss": 3.394321574175096,
+      "tokens_seen": 2524971008
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002371208473760231,
+      "loss": 0.0657,
+      "theoretical_loss": 3.394293757474631,
+      "tokens_seen": 2525233152
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002370406034344407,
+      "loss": 0.0682,
+      "theoretical_loss": 3.394265944470108,
+      "tokens_seen": 2525495296
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002369603594928583,
+      "loss": 0.063,
+      "theoretical_loss": 3.3942381351606525,
+      "tokens_seen": 2525757440
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002368801155512759,
+      "loss": 0.0669,
+      "theoretical_loss": 3.3942103295453903,
+      "tokens_seen": 2526019584
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023679987160969346,
+      "loss": 0.0635,
+      "theoretical_loss": 3.394182527623448,
+      "tokens_seen": 2526281728
+    },
+    {
+      "epoch": 0.77,
+      "objective/train/advantage_avg": 0.0006156957242637873,
+      "objective/train/docs_used": 919806,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.17476487159729,
+      "objective/train/original_loss": 1.17476487159729,
+      "objective/train/theoretical_loss": 3.394168628047198,
+      "objective/train/tokens_used": 2546872800,
+      "objective/train/value_avg": -0.01107025146484375,
+      "objective/train/value_loss": 0.0007344778859987855,
+      "objective/train/value_max": -5.519390106201172e-05,
+      "objective/train/value_min": -0.95849609375,
+      "objective/train/value_reward_corr": 0.7828262709025253,
+      "objective/train/value_std": 0.035430908203125,
+      "objective/train/weight_avg": 1.0009746551513672,
+      "objective/train/weighted_lm_loss": 1.1768461465835571,
+      "objective/train/weights_max": 2.2943544387817383,
+      "objective/train/weights_min": 0.38488778471946716,
+      "theoretical_loss": 3.394168628047198,
+      "tokens_seen": 2526412800
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023671962766811107,
+      "loss": 0.0677,
+      "theoretical_loss": 3.39415472939395,
+      "tokens_seen": 2526543872
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023663938372652866,
+      "loss": 0.0645,
+      "theoretical_loss": 3.3941269348560246,
+      "tokens_seen": 2526806016
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023655913978494624,
+      "loss": 0.0664,
+      "theoretical_loss": 3.3940991440087984,
+      "tokens_seen": 2527068160
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023647889584336383,
+      "loss": 0.0682,
+      "theoretical_loss": 3.3940713568513976,
+      "tokens_seen": 2527330304
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002363986519017814,
+      "loss": 0.0647,
+      "theoretical_loss": 3.394043573382951,
+      "tokens_seen": 2527592448
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023631840796019903,
+      "loss": 0.0654,
+      "theoretical_loss": 3.394015793602585,
+      "tokens_seen": 2527854592
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023623816401861658,
+      "loss": 0.0654,
+      "theoretical_loss": 3.393988017509429,
+      "tokens_seen": 2528116736
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002361579200770342,
+      "loss": 0.0666,
+      "theoretical_loss": 3.3939602451026096,
+      "tokens_seen": 2528378880
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023607767613545178,
+      "loss": 0.0646,
+      "theoretical_loss": 3.393932476381257,
+      "tokens_seen": 2528641024
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023599743219386937,
+      "loss": 0.066,
+      "theoretical_loss": 3.3939047113445,
+      "tokens_seen": 2528903168
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023591718825228695,
+      "loss": 0.0658,
+      "theoretical_loss": 3.393876949991467,
+      "tokens_seen": 2529165312
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023583694431070454,
+      "loss": 0.0659,
+      "theoretical_loss": 3.393849192321288,
+      "tokens_seen": 2529427456
+    },
+    {
+      "epoch": 0.77,
+      "objective/train/advantage_avg": 0.00048636639257892966,
+      "objective/train/docs_used": 920541,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2705464363098145,
+      "objective/train/original_loss": 1.270546555519104,
+      "objective/train/theoretical_loss": 3.3938214383330925,
+      "objective/train/tokens_used": 2550149600,
+      "objective/train/value_avg": -0.00516510009765625,
+      "objective/train/value_loss": 0.00013838338782079518,
+      "objective/train/value_max": -5.346536636352539e-05,
+      "objective/train/value_min": -0.260498046875,
+      "objective/train/value_reward_corr": 0.5679147804729487,
+      "objective/train/value_std": 0.0089569091796875,
+      "objective/train/weight_avg": 1.0005491971969604,
+      "objective/train/weighted_lm_loss": 1.2707606554031372,
+      "objective/train/weights_max": 1.1276289224624634,
+      "objective/train/weights_min": 0.3921318054199219,
+      "theoretical_loss": 3.3938214383330925,
+      "tokens_seen": 2529689600
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023575670036912215,
+      "loss": 0.0677,
+      "theoretical_loss": 3.3938214383330925,
+      "tokens_seen": 2529689600
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023567645642753974,
+      "loss": 0.0665,
+      "theoretical_loss": 3.3937936880260113,
+      "tokens_seen": 2529951744
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002355962124859573,
+      "loss": 0.0675,
+      "theoretical_loss": 3.3937659413991743,
+      "tokens_seen": 2530213888
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002355159685443749,
+      "loss": 0.066,
+      "theoretical_loss": 3.3937381984517123,
+      "tokens_seen": 2530476032
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002354357246027925,
+      "loss": 0.0634,
+      "theoretical_loss": 3.3937104591827563,
+      "tokens_seen": 2530738176
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002353554806612101,
+      "loss": 0.0658,
+      "theoretical_loss": 3.3936827235914375,
+      "tokens_seen": 2531000320
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023527523671962766,
+      "loss": 0.0665,
+      "theoretical_loss": 3.3936549916768874,
+      "tokens_seen": 2531262464
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023519499277804528,
+      "loss": 0.0683,
+      "theoretical_loss": 3.3936272634382387,
+      "tokens_seen": 2531524608
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023511474883646286,
+      "loss": 0.0663,
+      "theoretical_loss": 3.3935995388746227,
+      "tokens_seen": 2531786752
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023503450489488042,
+      "loss": 0.0633,
+      "theoretical_loss": 3.393571817985172,
+      "tokens_seen": 2532048896
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023495426095329803,
+      "loss": 0.0673,
+      "theoretical_loss": 3.3935441007690197,
+      "tokens_seen": 2532311040
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023487401701171562,
+      "loss": 0.0654,
+      "theoretical_loss": 3.393516387225299,
+      "tokens_seen": 2532573184
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023479377307013323,
+      "loss": 0.0653,
+      "theoretical_loss": 3.393488677353142,
+      "tokens_seen": 2532835328
+    },
+    {
+      "epoch": 0.77,
+      "objective/train/advantage_avg": -0.0022209736052900553,
+      "objective/train/docs_used": 921695,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3328808546066284,
+      "objective/train/original_loss": 1.3328807353973389,
+      "objective/train/theoretical_loss": 3.39347482379363,
+      "objective/train/tokens_used": 2553426400,
+      "objective/train/value_avg": -0.00995635986328125,
+      "objective/train/value_loss": 0.00040688071749173105,
+      "objective/train/value_max": -3.147125244140625e-05,
+      "objective/train/value_min": -0.552734375,
+      "objective/train/value_reward_corr": 0.7400871618874371,
+      "objective/train/value_std": 0.0183563232421875,
+      "objective/train/weight_avg": 0.9979639649391174,
+      "objective/train/weighted_lm_loss": 1.330152988433838,
+      "objective/train/weights_max": 1.2169888019561768,
+      "objective/train/weights_min": 0.3680844008922577,
+      "theoretical_loss": 3.39347482379363,
+      "tokens_seen": 2532966400
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002347135291285508,
+      "loss": 0.0655,
+      "theoretical_loss": 3.3934609711516845,
+      "tokens_seen": 2533097472
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023463328518696838,
+      "loss": 0.0638,
+      "theoretical_loss": 3.3934332686200586,
+      "tokens_seen": 2533359616
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.000234553041245386,
+      "loss": 0.0665,
+      "theoretical_loss": 3.3934055697573995,
+      "tokens_seen": 2533621760
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023447279730380357,
+      "loss": 0.0663,
+      "theoretical_loss": 3.393377874562841,
+      "tokens_seen": 2533883904
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023439255336222116,
+      "loss": 0.0656,
+      "theoretical_loss": 3.393350183035519,
+      "tokens_seen": 2534146048
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023431230942063874,
+      "loss": 0.0645,
+      "theoretical_loss": 3.3933224951745675,
+      "tokens_seen": 2534408192
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023423206547905636,
+      "loss": 0.0656,
+      "theoretical_loss": 3.3932948109791226,
+      "tokens_seen": 2534670336
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023415182153747392,
+      "loss": 0.0666,
+      "theoretical_loss": 3.39326713044832,
+      "tokens_seen": 2534932480
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002340715775958915,
+      "loss": 0.0669,
+      "theoretical_loss": 3.393239453581295,
+      "tokens_seen": 2535194624
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002339913336543091,
+      "loss": 0.067,
+      "theoretical_loss": 3.393211780377185,
+      "tokens_seen": 2535456768
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002339110897127267,
+      "loss": 0.0645,
+      "theoretical_loss": 3.3931841108351257,
+      "tokens_seen": 2535718912
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023383084577114428,
+      "loss": 0.0646,
+      "theoretical_loss": 3.393156444954255,
+      "tokens_seen": 2535981056
+    },
+    {
+      "epoch": 0.77,
+      "objective/train/advantage_avg": 0.00019107981643173844,
+      "objective/train/docs_used": 922895,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.402829885482788,
+      "objective/train/original_loss": 1.402829647064209,
+      "objective/train/theoretical_loss": 3.393128782733709,
+      "objective/train/tokens_used": 2556703200,
+      "objective/train/value_avg": -0.007198333740234375,
+      "objective/train/value_loss": 0.0003119884931948036,
+      "objective/train/value_max": -3.6776065826416016e-05,
+      "objective/train/value_min": -0.7333984375,
+      "objective/train/value_reward_corr": 0.7111351334598468,
+      "objective/train/value_std": 0.0170745849609375,
+      "objective/train/weight_avg": 1.000330924987793,
+      "objective/train/weighted_lm_loss": 1.4031528234481812,
+      "objective/train/weights_max": 1.3954987525939941,
+      "objective/train/weights_min": 0.3701740503311157,
+      "theoretical_loss": 3.393128782733709,
+      "tokens_seen": 2536243200
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023375060182956187,
+      "loss": 0.0671,
+      "theoretical_loss": 3.393128782733709,
+      "tokens_seen": 2536243200
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023367035788797946,
+      "loss": 0.0677,
+      "theoretical_loss": 3.3931011241726248,
+      "tokens_seen": 2536505344
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023359011394639707,
+      "loss": 0.0642,
+      "theoretical_loss": 3.393073469270142,
+      "tokens_seen": 2536767488
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023350987000481463,
+      "loss": 0.0653,
+      "theoretical_loss": 3.393045818025397,
+      "tokens_seen": 2537029632
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023342962606323224,
+      "loss": 0.0664,
+      "theoretical_loss": 3.393018170437529,
+      "tokens_seen": 2537291776
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023334938212164982,
+      "loss": 0.0671,
+      "theoretical_loss": 3.392990526505676,
+      "tokens_seen": 2537553920
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002332691381800674,
+      "loss": 0.0659,
+      "theoretical_loss": 3.3929628862289776,
+      "tokens_seen": 2537816064
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.000233188894238485,
+      "loss": 0.0662,
+      "theoretical_loss": 3.392935249606573,
+      "tokens_seen": 2538078208
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023310865029690258,
+      "loss": 0.0673,
+      "theoretical_loss": 3.3929076166376007,
+      "tokens_seen": 2538340352
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002330284063553202,
+      "loss": 0.0633,
+      "theoretical_loss": 3.3928799873212014,
+      "tokens_seen": 2538602496
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023294816241373775,
+      "loss": 0.065,
+      "theoretical_loss": 3.3928523616565154,
+      "tokens_seen": 2538864640
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023286791847215536,
+      "loss": 0.0637,
+      "theoretical_loss": 3.3928247396426827,
+      "tokens_seen": 2539126784
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023278767453057295,
+      "loss": 0.0629,
+      "theoretical_loss": 3.3927971212788437,
+      "tokens_seen": 2539388928
+    },
+    {
+      "epoch": 0.77,
+      "objective/train/advantage_avg": -3.936679786420427e-05,
+      "objective/train/docs_used": 924078,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2824890613555908,
+      "objective/train/original_loss": 1.2824890613555908,
+      "objective/train/theoretical_loss": 3.3927833134654035,
+      "objective/train/tokens_used": 2559980000,
+      "objective/train/value_avg": -0.00939178466796875,
+      "objective/train/value_loss": 0.000269672047579661,
+      "objective/train/value_max": -4.267692565917969e-05,
+      "objective/train/value_min": -0.66162109375,
+      "objective/train/value_reward_corr": 0.7390932968979256,
+      "objective/train/value_std": 0.0201263427734375,
+      "objective/train/weight_avg": 1.0000852346420288,
+      "objective/train/weighted_lm_loss": 1.2810968160629272,
+      "objective/train/weights_max": 1.523683786392212,
+      "objective/train/weights_min": 0.3727017939090729,
+      "theoretical_loss": 3.3927833134654035,
+      "tokens_seen": 2539520000
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023270743058899054,
+      "loss": 0.0658,
+      "theoretical_loss": 3.3927695065641394,
+      "tokens_seen": 2539651072
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023262718664740812,
+      "loss": 0.0643,
+      "theoretical_loss": 3.392741895497712,
+      "tokens_seen": 2539913216
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002325469427058257,
+      "loss": 0.0628,
+      "theoretical_loss": 3.392714288078702,
+      "tokens_seen": 2540175360
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023246669876424332,
+      "loss": 0.0647,
+      "theoretical_loss": 3.3926866843062515,
+      "tokens_seen": 2540437504
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023238645482266088,
+      "loss": 0.0643,
+      "theoretical_loss": 3.392659084179503,
+      "tokens_seen": 2540699648
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002323062108810785,
+      "loss": 0.0633,
+      "theoretical_loss": 3.3926314876975985,
+      "tokens_seen": 2540961792
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023222596693949608,
+      "loss": 0.0643,
+      "theoretical_loss": 3.3926038948596813,
+      "tokens_seen": 2541223936
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023214572299791366,
+      "loss": 0.0648,
+      "theoretical_loss": 3.3925763056648934,
+      "tokens_seen": 2541486080
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023206547905633125,
+      "loss": 0.0648,
+      "theoretical_loss": 3.3925487201123787,
+      "tokens_seen": 2541748224
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023198523511474883,
+      "loss": 0.0671,
+      "theoretical_loss": 3.392521138201281,
+      "tokens_seen": 2542010368
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023190499117316645,
+      "loss": 0.065,
+      "theoretical_loss": 3.392493559930744,
+      "tokens_seen": 2542272512
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023182474723158403,
+      "loss": 0.0668,
+      "theoretical_loss": 3.3924659852999115,
+      "tokens_seen": 2542534656
+    },
+    {
+      "epoch": 0.77,
+      "objective/train/advantage_avg": -0.0005489328177645802,
+      "objective/train/docs_used": 925229,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1024796962738037,
+      "objective/train/original_loss": 1.1024794578552246,
+      "objective/train/theoretical_loss": 3.3924384143079283,
+      "objective/train/tokens_used": 2563256800,
+      "objective/train/value_avg": -0.00791168212890625,
+      "objective/train/value_loss": 0.00014381870278157294,
+      "objective/train/value_max": -3.24249267578125e-05,
+      "objective/train/value_min": -0.222900390625,
+      "objective/train/value_reward_corr": 0.77419179227656,
+      "objective/train/value_std": 0.013671875,
+      "objective/train/weight_avg": 0.999518632888794,
+      "objective/train/weighted_lm_loss": 1.1019738912582397,
+      "objective/train/weights_max": 1.1265925168991089,
+      "objective/train/weights_min": 0.37364134192466736,
+      "theoretical_loss": 3.3924384143079283,
+      "tokens_seen": 2542796800
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002317445032900016,
+      "loss": 0.0647,
+      "theoretical_loss": 3.3924384143079283,
+      "tokens_seen": 2542796800
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002316642593484192,
+      "loss": 0.0684,
+      "theoretical_loss": 3.392410846953939,
+      "tokens_seen": 2543058944
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002315840154068368,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3923832832370886,
+      "tokens_seen": 2543321088
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023150377146525437,
+      "loss": 0.0672,
+      "theoretical_loss": 3.392355723156523,
+      "tokens_seen": 2543583232
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023142352752367196,
+      "loss": 0.0663,
+      "theoretical_loss": 3.392328166711387,
+      "tokens_seen": 2543845376
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023134328358208957,
+      "loss": 0.0652,
+      "theoretical_loss": 3.3923006139008267,
+      "tokens_seen": 2544107520
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023126303964050716,
+      "loss": 0.0642,
+      "theoretical_loss": 3.3922730647239887,
+      "tokens_seen": 2544369664
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023118279569892471,
+      "loss": 0.0632,
+      "theoretical_loss": 3.392245519180019,
+      "tokens_seen": 2544631808
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023110255175734233,
+      "loss": 0.0642,
+      "theoretical_loss": 3.3922179772680643,
+      "tokens_seen": 2544893952
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002310223078157599,
+      "loss": 0.0658,
+      "theoretical_loss": 3.392190438987272,
+      "tokens_seen": 2545156096
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023094206387417753,
+      "loss": 0.0667,
+      "theoretical_loss": 3.39216290433679,
+      "tokens_seen": 2545418240
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023086181993259508,
+      "loss": 0.0641,
+      "theoretical_loss": 3.392135373315764,
+      "tokens_seen": 2545680384
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023078157599101267,
+      "loss": 0.0634,
+      "theoretical_loss": 3.392107845923344,
+      "tokens_seen": 2545942528
+    },
+    {
+      "epoch": 0.77,
+      "objective/train/advantage_avg": 0.001058534486219287,
+      "objective/train/docs_used": 926243,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2137850522994995,
+      "objective/train/original_loss": 1.21378493309021,
+      "objective/train/theoretical_loss": 3.392094083587595,
+      "objective/train/tokens_used": 2566533600,
+      "objective/train/value_avg": -0.007965087890625,
+      "objective/train/value_loss": 0.0001501610822742805,
+      "objective/train/value_max": -2.2649765014648438e-05,
+      "objective/train/value_min": -0.30517578125,
+      "objective/train/value_reward_corr": 0.6901371006082144,
+      "objective/train/value_std": 0.01177215576171875,
+      "objective/train/weight_avg": 1.0011311769485474,
+      "objective/train/weighted_lm_loss": 1.2146296501159668,
+      "objective/train/weights_max": 1.2128745317459106,
+      "objective/train/weights_min": 0.6144814491271973,
+      "theoretical_loss": 3.392094083587595,
+      "tokens_seen": 2546073600
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023070133204943028,
+      "loss": 0.0637,
+      "theoretical_loss": 3.3920803221586775,
+      "tokens_seen": 2546204672
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023062108810784784,
+      "loss": 0.0636,
+      "theoretical_loss": 3.3920528020209124,
+      "tokens_seen": 2546466816
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023054084416626545,
+      "loss": 0.0617,
+      "theoretical_loss": 3.392025285509198,
+      "tokens_seen": 2546728960
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023046060022468304,
+      "loss": 0.0632,
+      "theoretical_loss": 3.391997772622683,
+      "tokens_seen": 2546991104
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00023038035628310065,
+      "loss": 0.0621,
+      "theoretical_loss": 3.3919702633605175,
+      "tokens_seen": 2547253248
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002303001123415182,
+      "loss": 0.0622,
+      "theoretical_loss": 3.391942757721851,
+      "tokens_seen": 2547515392
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002302198683999358,
+      "loss": 0.065,
+      "theoretical_loss": 3.3919152557058325,
+      "tokens_seen": 2547777536
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002301396244583534,
+      "loss": 0.0638,
+      "theoretical_loss": 3.3918877573116126,
+      "tokens_seen": 2548039680
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.000230059380516771,
+      "loss": 0.066,
+      "theoretical_loss": 3.3918602625383425,
+      "tokens_seen": 2548301824
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022997913657518858,
+      "loss": 0.0647,
+      "theoretical_loss": 3.391832771385172,
+      "tokens_seen": 2548563968
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022989889263360616,
+      "loss": 0.0633,
+      "theoretical_loss": 3.3918052838512533,
+      "tokens_seen": 2548826112
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022981864869202375,
+      "loss": 0.0642,
+      "theoretical_loss": 3.3917777999357366,
+      "tokens_seen": 2549088256
+    },
+    {
+      "epoch": 0.77,
+      "objective/train/advantage_avg": -0.0001033979351632297,
+      "objective/train/docs_used": 927515,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2924425601959229,
+      "objective/train/original_loss": 1.2924423217773438,
+      "objective/train/theoretical_loss": 3.3917503196377745,
+      "objective/train/tokens_used": 2569810400,
+      "objective/train/value_avg": -0.00595855712890625,
+      "objective/train/value_loss": 0.0001676402462180704,
+      "objective/train/value_max": -4.00543212890625e-05,
+      "objective/train/value_min": -0.206787109375,
+      "objective/train/value_reward_corr": 0.6296488844205017,
+      "objective/train/value_std": 0.00970458984375,
+      "objective/train/weight_avg": 0.9999687671661377,
+      "objective/train/weighted_lm_loss": 1.292542815208435,
+      "objective/train/weights_max": 1.1142876148223877,
+      "objective/train/weights_min": 0.36936861276626587,
+      "theoretical_loss": 3.3917503196377745,
+      "tokens_seen": 2549350400
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022973840475044136,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3917503196377745,
+      "tokens_seen": 2549350400
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022965816080885892,
+      "loss": 0.0647,
+      "theoretical_loss": 3.391722842956518,
+      "tokens_seen": 2549612544
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022957791686727653,
+      "loss": 0.0627,
+      "theoretical_loss": 3.3916953698911203,
+      "tokens_seen": 2549874688
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022949767292569412,
+      "loss": 0.0639,
+      "theoretical_loss": 3.391667900440733,
+      "tokens_seen": 2550136832
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002294174289841117,
+      "loss": 0.0647,
+      "theoretical_loss": 3.3916404346045095,
+      "tokens_seen": 2550398976
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002293371850425293,
+      "loss": 0.0641,
+      "theoretical_loss": 3.3916129723816026,
+      "tokens_seen": 2550661120
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022925694110094688,
+      "loss": 0.0654,
+      "theoretical_loss": 3.3915855137711657,
+      "tokens_seen": 2550923264
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002291766971593645,
+      "loss": 0.0619,
+      "theoretical_loss": 3.3915580587723526,
+      "tokens_seen": 2551185408
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022909645321778205,
+      "loss": 0.0623,
+      "theoretical_loss": 3.391530607384317,
+      "tokens_seen": 2551447552
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022901620927619966,
+      "loss": 0.0633,
+      "theoretical_loss": 3.3915031596062133,
+      "tokens_seen": 2551709696
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022893596533461724,
+      "loss": 0.0634,
+      "theoretical_loss": 3.391475715437196,
+      "tokens_seen": 2551971840
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022885572139303486,
+      "loss": 0.0623,
+      "theoretical_loss": 3.39144827487642,
+      "tokens_seen": 2552233984
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022877547745145242,
+      "loss": 0.0657,
+      "theoretical_loss": 3.3914208379230395,
+      "tokens_seen": 2552496128
+    },
+    {
+      "epoch": 0.77,
+      "objective/train/advantage_avg": 0.0006462376331910491,
+      "objective/train/docs_used": 928605,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2336928844451904,
+      "objective/train/original_loss": 1.2336928844451904,
+      "objective/train/theoretical_loss": 3.391407120798859,
+      "objective/train/tokens_used": 2573087200,
+      "objective/train/value_avg": -0.01276397705078125,
+      "objective/train/value_loss": 0.00046865135664120317,
+      "objective/train/value_max": -6.204843521118164e-05,
+      "objective/train/value_min": -0.7060546875,
+      "objective/train/value_reward_corr": 0.6902025219331749,
+      "objective/train/value_std": 0.0211639404296875,
+      "objective/train/weight_avg": 1.000848650932312,
+      "objective/train/weighted_lm_loss": 1.2347334623336792,
+      "objective/train/weights_max": 1.2664811611175537,
+      "objective/train/weights_min": 0.36937421560287476,
+      "theoretical_loss": 3.391407120798859,
+      "tokens_seen": 2552627200
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022869523350987,
+      "loss": 0.0656,
+      "theoretical_loss": 3.3913934045762106,
+      "tokens_seen": 2552758272
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022861498956828761,
+      "loss": 0.0624,
+      "theoretical_loss": 3.3913659748350895,
+      "tokens_seen": 2553020416
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022853474562670517,
+      "loss": 0.0659,
+      "theoretical_loss": 3.391338548698831,
+      "tokens_seen": 2553282560
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022845450168512278,
+      "loss": 0.0644,
+      "theoretical_loss": 3.391311126166592,
+      "tokens_seen": 2553544704
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022837425774354037,
+      "loss": 0.0623,
+      "theoretical_loss": 3.3912837072375286,
+      "tokens_seen": 2553806848
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022829401380195796,
+      "loss": 0.0648,
+      "theoretical_loss": 3.391256291910798,
+      "tokens_seen": 2554068992
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022821376986037554,
+      "loss": 0.0642,
+      "theoretical_loss": 3.391228880185557,
+      "tokens_seen": 2554331136
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022813352591879313,
+      "loss": 0.0625,
+      "theoretical_loss": 3.3912014720609625,
+      "tokens_seen": 2554593280
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022805328197721074,
+      "loss": 0.0659,
+      "theoretical_loss": 3.391174067536173,
+      "tokens_seen": 2554855424
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022797303803562832,
+      "loss": 0.0665,
+      "theoretical_loss": 3.391146666610346,
+      "tokens_seen": 2555117568
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022789279409404588,
+      "loss": 0.0665,
+      "theoretical_loss": 3.3911192692826395,
+      "tokens_seen": 2555379712
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.0002278125501524635,
+      "loss": 0.0647,
+      "theoretical_loss": 3.3910918755522124,
+      "tokens_seen": 2555641856
+    },
+    {
+      "epoch": 0.77,
+      "objective/train/advantage_avg": 0.0010113369207829237,
+      "objective/train/docs_used": 929757,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3162387609481812,
+      "objective/train/original_loss": 1.3162386417388916,
+      "objective/train/theoretical_loss": 3.391064485418223,
+      "objective/train/tokens_used": 2576364000,
+      "objective/train/value_avg": -0.005802154541015625,
+      "objective/train/value_loss": 0.0001210497721331194,
+      "objective/train/value_max": -5.829334259033203e-05,
+      "objective/train/value_min": -0.424072265625,
+      "objective/train/value_reward_corr": 0.7426082308917776,
+      "objective/train/value_std": 0.01232147216796875,
+      "objective/train/weight_avg": 1.0010701417922974,
+      "objective/train/weighted_lm_loss": 1.3174530267715454,
+      "objective/train/weights_max": 1.1952372789382935,
+      "objective/train/weights_min": 0.7097556591033936,
+      "theoretical_loss": 3.391064485418223,
+      "tokens_seen": 2555904000
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022773230621088108,
+      "loss": 0.0644,
+      "theoretical_loss": 3.391064485418223,
+      "tokens_seen": 2555904000
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022765206226929867,
+      "loss": 0.0667,
+      "theoretical_loss": 3.3910370988798304,
+      "tokens_seen": 2556166144
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022757181832771625,
+      "loss": 0.0666,
+      "theoretical_loss": 3.3910097159361943,
+      "tokens_seen": 2556428288
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022749157438613386,
+      "loss": 0.0635,
+      "theoretical_loss": 3.3909823365864744,
+      "tokens_seen": 2556690432
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.00022741133044455145,
+      "loss": 0.0646,
+      "theoretical_loss": 3.3909549608298297,
+      "tokens_seen": 2556952576
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 0.000227331086502969,
+      "loss": 0.0646,
+      "theoretical_loss": 3.3909275886654213,
+      "tokens_seen": 2557214720
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022725084256138662,
+      "loss": 0.0653,
+      "theoretical_loss": 3.3909002200924094,
+      "tokens_seen": 2557476864
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002271705986198042,
+      "loss": 0.0655,
+      "theoretical_loss": 3.3908728551099543,
+      "tokens_seen": 2557739008
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022709035467822182,
+      "loss": 0.0656,
+      "theoretical_loss": 3.390845493717218,
+      "tokens_seen": 2558001152
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022701011073663938,
+      "loss": 0.067,
+      "theoretical_loss": 3.3908181359133605,
+      "tokens_seen": 2558263296
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.000226929866795057,
+      "loss": 0.0642,
+      "theoretical_loss": 3.390790781697544,
+      "tokens_seen": 2558525440
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022684962285347458,
+      "loss": 0.0665,
+      "theoretical_loss": 3.390763431068931,
+      "tokens_seen": 2558787584
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022676937891189213,
+      "loss": 0.0641,
+      "theoretical_loss": 3.390736084026683,
+      "tokens_seen": 2559049728
+    },
+    {
+      "epoch": 0.78,
+      "objective/train/advantage_avg": 1.652149876463227e-06,
+      "objective/train/docs_used": 930868,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.139220952987671,
+      "objective/train/original_loss": 1.139220952987671,
+      "objective/train/theoretical_loss": 3.390722411850184,
+      "objective/train/tokens_used": 2579640800,
+      "objective/train/value_avg": -0.0068511962890625,
+      "objective/train/value_loss": 0.0003664409159682691,
+      "objective/train/value_max": -4.035234451293945e-05,
+      "objective/train/value_min": -0.71240234375,
+      "objective/train/value_reward_corr": 0.6045150013411714,
+      "objective/train/value_std": 0.0150909423828125,
+      "objective/train/weight_avg": 1.0001568794250488,
+      "objective/train/weighted_lm_loss": 1.1387513875961304,
+      "objective/train/weights_max": 1.5181612968444824,
+      "objective/train/weights_min": 0.3759317100048065,
+      "theoretical_loss": 3.390722411850184,
+      "tokens_seen": 2559180800
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022668913497030975,
+      "loss": 0.0609,
+      "theoretical_loss": 3.390708740569962,
+      "tokens_seen": 2559311872
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022660889102872733,
+      "loss": 0.0665,
+      "theoretical_loss": 3.3906814006979316,
+      "tokens_seen": 2559574016
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022652864708714495,
+      "loss": 0.0641,
+      "theoretical_loss": 3.390654064409754,
+      "tokens_seen": 2559836160
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002264484031455625,
+      "loss": 0.0602,
+      "theoretical_loss": 3.3906267317045935,
+      "tokens_seen": 2560098304
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002263681592039801,
+      "loss": 0.064,
+      "theoretical_loss": 3.3905994025816124,
+      "tokens_seen": 2560360448
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002262879152623977,
+      "loss": 0.0643,
+      "theoretical_loss": 3.3905720770399754,
+      "tokens_seen": 2560622592
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002262076713208153,
+      "loss": 0.0616,
+      "theoretical_loss": 3.3905447550788463,
+      "tokens_seen": 2560884736
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022612742737923287,
+      "loss": 0.0636,
+      "theoretical_loss": 3.39051743669739,
+      "tokens_seen": 2561146880
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022604718343765046,
+      "loss": 0.0641,
+      "theoretical_loss": 3.39049012189477,
+      "tokens_seen": 2561409024
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022596693949606807,
+      "loss": 0.0626,
+      "theoretical_loss": 3.3904628106701526,
+      "tokens_seen": 2561671168
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022588669555448563,
+      "loss": 0.066,
+      "theoretical_loss": 3.3904355030227022,
+      "tokens_seen": 2561933312
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022580645161290321,
+      "loss": 0.0643,
+      "theoretical_loss": 3.3904081989515844,
+      "tokens_seen": 2562195456
+    },
+    {
+      "epoch": 0.78,
+      "objective/train/advantage_avg": 0.0004914909368380904,
+      "objective/train/docs_used": 931999,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2367290258407593,
+      "objective/train/original_loss": 1.2367291450500488,
+      "objective/train/theoretical_loss": 3.390380898455965,
+      "objective/train/tokens_used": 2582917600,
+      "objective/train/value_avg": -0.006000518798828125,
+      "objective/train/value_loss": 8.66991831571795e-05,
+      "objective/train/value_max": -4.267692565917969e-05,
+      "objective/train/value_min": -0.195556640625,
+      "objective/train/value_reward_corr": 0.74524198172894,
+      "objective/train/value_std": 0.01154327392578125,
+      "objective/train/weight_avg": 1.0005344152450562,
+      "objective/train/weighted_lm_loss": 1.2368214130401611,
+      "objective/train/weights_max": 1.1457269191741943,
+      "objective/train/weights_min": 0.822311520576477,
+      "theoretical_loss": 3.390380898455965,
+      "tokens_seen": 2562457600
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022572620767132083,
+      "loss": 0.0631,
+      "theoretical_loss": 3.390380898455965,
+      "tokens_seen": 2562457600
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002256459637297384,
+      "loss": 0.0665,
+      "theoretical_loss": 3.390353601535011,
+      "tokens_seen": 2562719744
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.000225565719788156,
+      "loss": 0.0642,
+      "theoretical_loss": 3.3903263081878876,
+      "tokens_seen": 2562981888
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022548547584657358,
+      "loss": 0.0635,
+      "theoretical_loss": 3.390299018413762,
+      "tokens_seen": 2563244032
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022540523190499117,
+      "loss": 0.0636,
+      "theoretical_loss": 3.3902717322118,
+      "tokens_seen": 2563506176
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022532498796340878,
+      "loss": 0.0659,
+      "theoretical_loss": 3.3902444495811705,
+      "tokens_seen": 2563768320
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022524474402182634,
+      "loss": 0.065,
+      "theoretical_loss": 3.3902171705210398,
+      "tokens_seen": 2564030464
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022516450008024395,
+      "loss": 0.0626,
+      "theoretical_loss": 3.390189895030576,
+      "tokens_seen": 2564292608
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022508425613866154,
+      "loss": 0.0636,
+      "theoretical_loss": 3.390162623108948,
+      "tokens_seen": 2564554752
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022500401219707915,
+      "loss": 0.0652,
+      "theoretical_loss": 3.390135354755323,
+      "tokens_seen": 2564816896
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002249237682554967,
+      "loss": 0.0643,
+      "theoretical_loss": 3.3901080899688694,
+      "tokens_seen": 2565079040
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002248435243139143,
+      "loss": 0.0638,
+      "theoretical_loss": 3.390080828748757,
+      "tokens_seen": 2565341184
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002247632803723319,
+      "loss": 0.0644,
+      "theoretical_loss": 3.390053571094154,
+      "tokens_seen": 2565603328
+    },
+    {
+      "epoch": 0.78,
+      "objective/train/advantage_avg": 6.488848612207221e-06,
+      "objective/train/docs_used": 933221,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.4858256578445435,
+      "objective/train/original_loss": 1.485825538635254,
+      "objective/train/theoretical_loss": 3.3900399436036595,
+      "objective/train/tokens_used": 2586194400,
+      "objective/train/value_avg": -0.00689697265625,
+      "objective/train/value_loss": 0.00017969694454222918,
+      "objective/train/value_max": -3.170967102050781e-05,
+      "objective/train/value_min": -0.291748046875,
+      "objective/train/value_reward_corr": 0.7551056464542931,
+      "objective/train/value_std": 0.01503753662109375,
+      "objective/train/weight_avg": 1.0000879764556885,
+      "objective/train/weighted_lm_loss": 1.4855372905731201,
+      "objective/train/weights_max": 1.12247896194458,
+      "objective/train/weights_min": 0.3718355596065521,
+      "theoretical_loss": 3.3900399436036595,
+      "tokens_seen": 2565734400
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022468303643074947,
+      "loss": 0.0655,
+      "theoretical_loss": 3.3900263170042306,
+      "tokens_seen": 2565865472
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022460279248916708,
+      "loss": 0.0633,
+      "theoretical_loss": 3.3899990664781563,
+      "tokens_seen": 2566127616
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022452254854758466,
+      "loss": 0.0659,
+      "theoretical_loss": 3.389971819515101,
+      "tokens_seen": 2566389760
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022444230460600225,
+      "loss": 0.0661,
+      "theoretical_loss": 3.389944576114235,
+      "tokens_seen": 2566651904
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022436206066441984,
+      "loss": 0.0624,
+      "theoretical_loss": 3.3899173362747286,
+      "tokens_seen": 2566914048
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022428181672283742,
+      "loss": 0.0648,
+      "theoretical_loss": 3.389890099995753,
+      "tokens_seen": 2567176192
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022420157278125503,
+      "loss": 0.0661,
+      "theoretical_loss": 3.389862867276479,
+      "tokens_seen": 2567438336
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022412132883967262,
+      "loss": 0.0628,
+      "theoretical_loss": 3.3898356381160784,
+      "tokens_seen": 2567700480
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002240410848980902,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3898084125137222,
+      "tokens_seen": 2567962624
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002239608409565078,
+      "loss": 0.0645,
+      "theoretical_loss": 3.389781190468583,
+      "tokens_seen": 2568224768
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022388059701492538,
+      "loss": 0.0651,
+      "theoretical_loss": 3.3897539719798324,
+      "tokens_seen": 2568486912
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022380035307334296,
+      "loss": 0.0657,
+      "theoretical_loss": 3.389726757046643,
+      "tokens_seen": 2568749056
+    },
+    {
+      "epoch": 0.78,
+      "objective/train/advantage_avg": 0.0008476021466776729,
+      "objective/train/docs_used": 934338,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1809134483337402,
+      "objective/train/original_loss": 1.1809133291244507,
+      "objective/train/theoretical_loss": 3.3896995456681878,
+      "objective/train/tokens_used": 2589471200,
+      "objective/train/value_avg": -0.00931549072265625,
+      "objective/train/value_loss": 0.00018791698676068336,
+      "objective/train/value_max": -3.045797348022461e-05,
+      "objective/train/value_min": -0.2313232421875,
+      "objective/train/value_reward_corr": 0.8049078212678962,
+      "objective/train/value_std": 0.018310546875,
+      "objective/train/weight_avg": 1.0009363889694214,
+      "objective/train/weighted_lm_loss": 1.1823976039886475,
+      "objective/train/weights_max": 1.1745542287826538,
+      "objective/train/weights_min": 0.37418052554130554,
+      "theoretical_loss": 3.3896995456681878,
+      "tokens_seen": 2569011200
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022372010913176055,
+      "loss": 0.0633,
+      "theoretical_loss": 3.3896995456681878,
+      "tokens_seen": 2569011200
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022363986519017816,
+      "loss": 0.069,
+      "theoretical_loss": 3.38967233784364,
+      "tokens_seen": 2569273344
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022355962124859574,
+      "loss": 0.0636,
+      "theoretical_loss": 3.3896451335721727,
+      "tokens_seen": 2569535488
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002234793773070133,
+      "loss": 0.0646,
+      "theoretical_loss": 3.3896179328529588,
+      "tokens_seen": 2569797632
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022339913336543092,
+      "loss": 0.0665,
+      "theoretical_loss": 3.3895907356851733,
+      "tokens_seen": 2570059776
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002233188894238485,
+      "loss": 0.0652,
+      "theoretical_loss": 3.38956354206799,
+      "tokens_seen": 2570321920
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022323864548226611,
+      "loss": 0.0667,
+      "theoretical_loss": 3.389536352000583,
+      "tokens_seen": 2570584064
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022315840154068367,
+      "loss": 0.0686,
+      "theoretical_loss": 3.389509165482127,
+      "tokens_seen": 2570846208
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022307815759910128,
+      "loss": 0.0663,
+      "theoretical_loss": 3.389481982511797,
+      "tokens_seen": 2571108352
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022299791365751887,
+      "loss": 0.0669,
+      "theoretical_loss": 3.3894548030887686,
+      "tokens_seen": 2571370496
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022291766971593643,
+      "loss": 0.0644,
+      "theoretical_loss": 3.3894276272122172,
+      "tokens_seen": 2571632640
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022283742577435404,
+      "loss": 0.0641,
+      "theoretical_loss": 3.389400454881318,
+      "tokens_seen": 2571894784
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022275718183277163,
+      "loss": 0.0669,
+      "theoretical_loss": 3.389373286095248,
+      "tokens_seen": 2572156928
+    },
+    {
+      "epoch": 0.78,
+      "objective/train/advantage_avg": 0.0004840648907702416,
+      "objective/train/docs_used": 935480,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2403749227523804,
+      "objective/train/original_loss": 1.24037504196167,
+      "objective/train/theoretical_loss": 3.3893597030312663,
+      "objective/train/tokens_used": 2592748000,
+      "objective/train/value_avg": -0.007007598876953125,
+      "objective/train/value_loss": 0.0001475304743507877,
+      "objective/train/value_max": -4.506111145019531e-05,
+      "objective/train/value_min": -0.305908203125,
+      "objective/train/value_reward_corr": 0.7053234026415041,
+      "objective/train/value_std": 0.0128021240234375,
+      "objective/train/weight_avg": 1.0005536079406738,
+      "objective/train/weighted_lm_loss": 1.2408409118652344,
+      "objective/train/weights_max": 1.186949610710144,
+      "objective/train/weights_min": 0.39292243123054504,
+      "theoretical_loss": 3.3893597030312663,
+      "tokens_seen": 2572288000
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022267693789118924,
+      "loss": 0.0707,
+      "theoretical_loss": 3.389346120853183,
+      "tokens_seen": 2572419072
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002225966939496068,
+      "loss": 0.0681,
+      "theoretical_loss": 3.3893189591543,
+      "tokens_seen": 2572681216
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022251645000802438,
+      "loss": 0.0686,
+      "theoretical_loss": 3.3892918009977753,
+      "tokens_seen": 2572943360
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.000222436206066442,
+      "loss": 0.0642,
+      "theoretical_loss": 3.3892646463827862,
+      "tokens_seen": 2573205504
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022235596212485958,
+      "loss": 0.0634,
+      "theoretical_loss": 3.3892374953085107,
+      "tokens_seen": 2573467648
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022227571818327717,
+      "loss": 0.0701,
+      "theoretical_loss": 3.389210347774126,
+      "tokens_seen": 2573729792
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022219547424169475,
+      "loss": 0.0646,
+      "theoretical_loss": 3.3891832037788103,
+      "tokens_seen": 2573991936
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022211523030011236,
+      "loss": 0.0631,
+      "theoretical_loss": 3.389156063321742,
+      "tokens_seen": 2574254080
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022203498635852992,
+      "loss": 0.065,
+      "theoretical_loss": 3.3891289264020994,
+      "tokens_seen": 2574516224
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002219547424169475,
+      "loss": 0.066,
+      "theoretical_loss": 3.3891017930190612,
+      "tokens_seen": 2574778368
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022187449847536512,
+      "loss": 0.0673,
+      "theoretical_loss": 3.3890746631718067,
+      "tokens_seen": 2575040512
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002217942545337827,
+      "loss": 0.0611,
+      "theoretical_loss": 3.3890475368595157,
+      "tokens_seen": 2575302656
+    },
+    {
+      "epoch": 0.78,
+      "objective/train/advantage_avg": 0.000912030111066997,
+      "objective/train/docs_used": 936685,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3602075576782227,
+      "objective/train/original_loss": 1.360207438468933,
+      "objective/train/theoretical_loss": 3.389020414081367,
+      "objective/train/tokens_used": 2596024800,
+      "objective/train/value_avg": -0.00836181640625,
+      "objective/train/value_loss": 0.0001833103597164154,
+      "objective/train/value_max": -6.759166717529297e-05,
+      "objective/train/value_min": -0.2305908203125,
+      "objective/train/value_reward_corr": 0.6511028326974286,
+      "objective/train/value_std": 0.01253509521484375,
+      "objective/train/weight_avg": 1.0009958744049072,
+      "objective/train/weighted_lm_loss": 1.3608472347259521,
+      "objective/train/weights_max": 1.1629977226257324,
+      "objective/train/weights_min": 0.3708355128765106,
+      "theoretical_loss": 3.389020414081367,
+      "tokens_seen": 2575564800
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002217140105922003,
+      "loss": 0.0665,
+      "theoretical_loss": 3.389020414081367,
+      "tokens_seen": 2575564800
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022163376665061788,
+      "loss": 0.066,
+      "theoretical_loss": 3.3889932948365407,
+      "tokens_seen": 2575826944
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022155352270903546,
+      "loss": 0.0632,
+      "theoretical_loss": 3.3889661791242176,
+      "tokens_seen": 2576089088
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022147327876745308,
+      "loss": 0.0678,
+      "theoretical_loss": 3.3889390669435775,
+      "tokens_seen": 2576351232
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022139303482587063,
+      "loss": 0.0681,
+      "theoretical_loss": 3.388911958293802,
+      "tokens_seen": 2576613376
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022131279088428825,
+      "loss": 0.0615,
+      "theoretical_loss": 3.388884853174071,
+      "tokens_seen": 2576875520
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022123254694270583,
+      "loss": 0.0688,
+      "theoretical_loss": 3.3888577515835663,
+      "tokens_seen": 2577137664
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022115230300112342,
+      "loss": 0.0615,
+      "theoretical_loss": 3.3888306535214694,
+      "tokens_seen": 2577399808
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.000221072059059541,
+      "loss": 0.0657,
+      "theoretical_loss": 3.3888035589869627,
+      "tokens_seen": 2577661952
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002209918151179586,
+      "loss": 0.0653,
+      "theoretical_loss": 3.3887764679792274,
+      "tokens_seen": 2577924096
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002209115711763762,
+      "loss": 0.0656,
+      "theoretical_loss": 3.3887493804974462,
+      "tokens_seen": 2578186240
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022083132723479376,
+      "loss": 0.0691,
+      "theoretical_loss": 3.388722296540802,
+      "tokens_seen": 2578448384
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022075108329321137,
+      "loss": 0.0673,
+      "theoretical_loss": 3.3886952161084776,
+      "tokens_seen": 2578710528
+    },
+    {
+      "epoch": 0.78,
+      "objective/train/advantage_avg": 0.00034838804276660085,
+      "objective/train/docs_used": 937774,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.201945185661316,
+      "objective/train/original_loss": 1.2019450664520264,
+      "objective/train/theoretical_loss": 3.38868167721368,
+      "objective/train/tokens_used": 2599301600,
+      "objective/train/value_avg": -0.00756072998046875,
+      "objective/train/value_loss": 0.00030759855872020125,
+      "objective/train/value_max": -3.218650817871094e-05,
+      "objective/train/value_min": -0.5537109375,
+      "objective/train/value_reward_corr": 0.70862187347612,
+      "objective/train/value_std": 0.01560211181640625,
+      "objective/train/weight_avg": 1.0004799365997314,
+      "objective/train/weighted_lm_loss": 1.2030587196350098,
+      "objective/train/weights_max": 1.176781177520752,
+      "objective/train/weights_min": 0.3896505832672119,
+      "theoretical_loss": 3.38868167721368,
+      "tokens_seen": 2578841600
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022067083935162896,
+      "loss": 0.065,
+      "theoretical_loss": 3.3886681391996563,
+      "tokens_seen": 2578972672
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022059059541004654,
+      "loss": 0.0677,
+      "theoretical_loss": 3.3886410658135206,
+      "tokens_seen": 2579234816
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022051035146846413,
+      "loss": 0.068,
+      "theoretical_loss": 3.388613995949256,
+      "tokens_seen": 2579496960
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022043010752688171,
+      "loss": 0.0675,
+      "theoretical_loss": 3.388586929606045,
+      "tokens_seen": 2579759104
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022034986358529933,
+      "loss": 0.0643,
+      "theoretical_loss": 3.3885598667830727,
+      "tokens_seen": 2580021248
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002202696196437169,
+      "loss": 0.0619,
+      "theoretical_loss": 3.3885328074795233,
+      "tokens_seen": 2580283392
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002201893757021345,
+      "loss": 0.0647,
+      "theoretical_loss": 3.3885057516945816,
+      "tokens_seen": 2580545536
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022010913176055208,
+      "loss": 0.0673,
+      "theoretical_loss": 3.388478699427433,
+      "tokens_seen": 2580807680
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00022002888781896967,
+      "loss": 0.0639,
+      "theoretical_loss": 3.388451650677262,
+      "tokens_seen": 2581069824
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021994864387738725,
+      "loss": 0.0669,
+      "theoretical_loss": 3.388424605443256,
+      "tokens_seen": 2581331968
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021986839993580484,
+      "loss": 0.0655,
+      "theoretical_loss": 3.3883975637245993,
+      "tokens_seen": 2581594112
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021978815599422245,
+      "loss": 0.0639,
+      "theoretical_loss": 3.388370525520479,
+      "tokens_seen": 2581856256
+    },
+    {
+      "epoch": 0.78,
+      "objective/train/advantage_avg": 1.741459891491104e-05,
+      "objective/train/docs_used": 938961,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.39012610912323,
+      "objective/train/original_loss": 1.3901259899139404,
+      "objective/train/theoretical_loss": 3.3883434908300805,
+      "objective/train/tokens_used": 2602578400,
+      "objective/train/value_avg": -0.01126861572265625,
+      "objective/train/value_loss": 0.0005996196996420622,
+      "objective/train/value_max": -3.975629806518555e-05,
+      "objective/train/value_min": -0.6435546875,
+      "objective/train/value_reward_corr": 0.7908500869694258,
+      "objective/train/value_std": 0.028778076171875,
+      "objective/train/weight_avg": 1.0002272129058838,
+      "objective/train/weighted_lm_loss": 1.3898470401763916,
+      "objective/train/weights_max": 1.3806082010269165,
+      "objective/train/weights_min": 0.05692806467413902,
+      "theoretical_loss": 3.3883434908300805,
+      "tokens_seen": 2582118400
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021970791205264004,
+      "loss": 0.0654,
+      "theoretical_loss": 3.3883434908300805,
+      "tokens_seen": 2582118400
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002196276681110576,
+      "loss": 0.0675,
+      "theoretical_loss": 3.388316459652591,
+      "tokens_seen": 2582380544
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002195474241694752,
+      "loss": 0.0639,
+      "theoretical_loss": 3.3882894319871983,
+      "tokens_seen": 2582642688
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002194671802278928,
+      "loss": 0.0654,
+      "theoretical_loss": 3.388262407833089,
+      "tokens_seen": 2582904832
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002193869362863104,
+      "loss": 0.0673,
+      "theoretical_loss": 3.3882353871894506,
+      "tokens_seen": 2583166976
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021930669234472797,
+      "loss": 0.0657,
+      "theoretical_loss": 3.388208370055471,
+      "tokens_seen": 2583429120
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021922644840314558,
+      "loss": 0.065,
+      "theoretical_loss": 3.388181356430338,
+      "tokens_seen": 2583691264
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021914620446156316,
+      "loss": 0.0641,
+      "theoretical_loss": 3.3881543463132404,
+      "tokens_seen": 2583953408
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021906596051998072,
+      "loss": 0.0644,
+      "theoretical_loss": 3.3881273397033667,
+      "tokens_seen": 2584215552
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021898571657839834,
+      "loss": 0.0657,
+      "theoretical_loss": 3.388100336599906,
+      "tokens_seen": 2584477696
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021890547263681592,
+      "loss": 0.0649,
+      "theoretical_loss": 3.388073337002047,
+      "tokens_seen": 2584739840
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021882522869523353,
+      "loss": 0.065,
+      "theoretical_loss": 3.388046340908979,
+      "tokens_seen": 2585001984
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002187449847536511,
+      "loss": 0.0653,
+      "theoretical_loss": 3.388019348319892,
+      "tokens_seen": 2585264128
+    },
+    {
+      "epoch": 0.78,
+      "objective/train/advantage_avg": 0.0002849264128599316,
+      "objective/train/docs_used": 939988,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.224392056465149,
+      "objective/train/original_loss": 1.224392056465149,
+      "objective/train/theoretical_loss": 3.3880058533390884,
+      "objective/train/tokens_used": 2605855200,
+      "objective/train/value_avg": -0.006206512451171875,
+      "objective/train/value_loss": 0.0002516733657103032,
+      "objective/train/value_max": -3.2961368560791016e-05,
+      "objective/train/value_min": -0.38623046875,
+      "objective/train/value_reward_corr": 0.6863537063306748,
+      "objective/train/value_std": 0.01476287841796875,
+      "objective/train/weight_avg": 1.0003920793533325,
+      "objective/train/weighted_lm_loss": 1.2246785163879395,
+      "objective/train/weights_max": 1.432533860206604,
+      "objective/train/weights_min": 0.38275596499443054,
+      "theoretical_loss": 3.3880058533390884,
+      "tokens_seen": 2585395200
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002186647408120687,
+      "loss": 0.0624,
+      "theoretical_loss": 3.3879923592339765,
+      "tokens_seen": 2585526272
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002185844968704863,
+      "loss": 0.0641,
+      "theoretical_loss": 3.3879653736504216,
+      "tokens_seen": 2585788416
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021850425292890388,
+      "loss": 0.0626,
+      "theoretical_loss": 3.3879383915684187,
+      "tokens_seen": 2586050560
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021842400898732146,
+      "loss": 0.0696,
+      "theoretical_loss": 3.387911412987158,
+      "tokens_seen": 2586312704
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021834376504573905,
+      "loss": 0.0653,
+      "theoretical_loss": 3.3878844379058313,
+      "tokens_seen": 2586574848
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021826352110415666,
+      "loss": 0.0677,
+      "theoretical_loss": 3.387857466323629,
+      "tokens_seen": 2586836992
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021818327716257422,
+      "loss": 0.0676,
+      "theoretical_loss": 3.3878304982397434,
+      "tokens_seen": 2587099136
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002181030332209918,
+      "loss": 0.0646,
+      "theoretical_loss": 3.387803533653366,
+      "tokens_seen": 2587361280
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021802278927940942,
+      "loss": 0.0683,
+      "theoretical_loss": 3.3877765725636886,
+      "tokens_seen": 2587623424
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.000217942545337827,
+      "loss": 0.0657,
+      "theoretical_loss": 3.3877496149699047,
+      "tokens_seen": 2587885568
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002178623013962446,
+      "loss": 0.0638,
+      "theoretical_loss": 3.387722660871206,
+      "tokens_seen": 2588147712
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021778205745466217,
+      "loss": 0.0662,
+      "theoretical_loss": 3.387695710266785,
+      "tokens_seen": 2588409856
+    },
+    {
+      "epoch": 0.78,
+      "objective/train/advantage_avg": -8.60083382576704e-05,
+      "objective/train/docs_used": 941211,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.307102918624878,
+      "objective/train/original_loss": 1.3071026802062988,
+      "objective/train/theoretical_loss": 3.387668763155836,
+      "objective/train/tokens_used": 2609132000,
+      "objective/train/value_avg": -0.007694244384765625,
+      "objective/train/value_loss": 0.00019316507678013295,
+      "objective/train/value_max": -4.7206878662109375e-05,
+      "objective/train/value_min": -0.380126953125,
+      "objective/train/value_reward_corr": 0.6662835411165384,
+      "objective/train/value_std": 0.0128936767578125,
+      "objective/train/weight_avg": 1.000002145767212,
+      "objective/train/weighted_lm_loss": 1.3065834045410156,
+      "objective/train/weights_max": 1.1861990690231323,
+      "objective/train/weights_min": 0.3745060861110687,
+      "theoretical_loss": 3.387668763155836,
+      "tokens_seen": 2588672000
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021770181351307978,
+      "loss": 0.0681,
+      "theoretical_loss": 3.387668763155836,
+      "tokens_seen": 2588672000
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021762156957149737,
+      "loss": 0.066,
+      "theoretical_loss": 3.387641819537552,
+      "tokens_seen": 2588934144
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021754132562991493,
+      "loss": 0.0682,
+      "theoretical_loss": 3.3876148794111267,
+      "tokens_seen": 2589196288
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021746108168833254,
+      "loss": 0.064,
+      "theoretical_loss": 3.3875879427757543,
+      "tokens_seen": 2589458432
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.00021738083774675013,
+      "loss": 0.0653,
+      "theoretical_loss": 3.3875610096306286,
+      "tokens_seen": 2589720576
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002173005938051677,
+      "loss": 0.0659,
+      "theoretical_loss": 3.3875340799749445,
+      "tokens_seen": 2589982720
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 0.0002172203498635853,
+      "loss": 0.0612,
+      "theoretical_loss": 3.3875071538078965,
+      "tokens_seen": 2590244864
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021714010592200288,
+      "loss": 0.0623,
+      "theoretical_loss": 3.38748023112868,
+      "tokens_seen": 2590507008
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002170598619804205,
+      "loss": 0.0641,
+      "theoretical_loss": 3.3874533119364902,
+      "tokens_seen": 2590769152
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021697961803883805,
+      "loss": 0.0669,
+      "theoretical_loss": 3.3874263962305227,
+      "tokens_seen": 2591031296
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021689937409725567,
+      "loss": 0.0635,
+      "theoretical_loss": 3.3873994840099733,
+      "tokens_seen": 2591293440
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021681913015567325,
+      "loss": 0.0663,
+      "theoretical_loss": 3.3873725752740382,
+      "tokens_seen": 2591555584
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021673888621409087,
+      "loss": 0.0637,
+      "theoretical_loss": 3.387345670021914,
+      "tokens_seen": 2591817728
+    },
+    {
+      "epoch": 0.79,
+      "objective/train/advantage_avg": 0.000786102085839957,
+      "objective/train/docs_used": 942495,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2474442720413208,
+      "objective/train/original_loss": 1.2474443912506104,
+      "objective/train/theoretical_loss": 3.3873322187020296,
+      "objective/train/tokens_used": 2612408800,
+      "objective/train/value_avg": -0.006381988525390625,
+      "objective/train/value_loss": 0.00012210979184601456,
+      "objective/train/value_max": -3.11732292175293e-05,
+      "objective/train/value_min": -0.37255859375,
+      "objective/train/value_reward_corr": 0.750205768261117,
+      "objective/train/value_std": 0.0131072998046875,
+      "objective/train/weight_avg": 1.000842809677124,
+      "objective/train/weighted_lm_loss": 1.248508095741272,
+      "objective/train/weights_max": 1.166409969329834,
+      "objective/train/weights_min": 0.3774433732032776,
+      "theoretical_loss": 3.3873322187020296,
+      "tokens_seen": 2591948800
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021665864227250842,
+      "loss": 0.0657,
+      "theoretical_loss": 3.387318768252797,
+      "tokens_seen": 2592079872
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.000216578398330926,
+      "loss": 0.0652,
+      "theoretical_loss": 3.387291869965884,
+      "tokens_seen": 2592342016
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021649815438934362,
+      "loss": 0.0676,
+      "theoretical_loss": 3.3872649751603725,
+      "tokens_seen": 2592604160
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021641791044776118,
+      "loss": 0.0663,
+      "theoretical_loss": 3.3872380838354603,
+      "tokens_seen": 2592866304
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002163376665061788,
+      "loss": 0.0618,
+      "theoretical_loss": 3.3872111959903446,
+      "tokens_seen": 2593128448
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021625742256459638,
+      "loss": 0.0633,
+      "theoretical_loss": 3.3871843116242237,
+      "tokens_seen": 2593390592
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021617717862301396,
+      "loss": 0.0654,
+      "theoretical_loss": 3.3871574307362957,
+      "tokens_seen": 2593652736
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021609693468143155,
+      "loss": 0.0651,
+      "theoretical_loss": 3.387130553325759,
+      "tokens_seen": 2593914880
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021601669073984913,
+      "loss": 0.0659,
+      "theoretical_loss": 3.3871036793918123,
+      "tokens_seen": 2594177024
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021593644679826675,
+      "loss": 0.0659,
+      "theoretical_loss": 3.3870768089336556,
+      "tokens_seen": 2594439168
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021585620285668433,
+      "loss": 0.0663,
+      "theoretical_loss": 3.3870499419504867,
+      "tokens_seen": 2594701312
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021577595891510192,
+      "loss": 0.0659,
+      "theoretical_loss": 3.3870230784415067,
+      "tokens_seen": 2594963456
+    },
+    {
+      "epoch": 0.79,
+      "objective/train/advantage_avg": -0.0007075011963024735,
+      "objective/train/docs_used": 943650,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3366597890853882,
+      "objective/train/original_loss": 1.3366597890853882,
+      "objective/train/theoretical_loss": 3.3869962184059146,
+      "objective/train/tokens_used": 2615685600,
+      "objective/train/value_avg": -0.00949859619140625,
+      "objective/train/value_loss": 0.0006140853511169553,
+      "objective/train/value_max": -1.9252300262451172e-05,
+      "objective/train/value_min": -0.75830078125,
+      "objective/train/value_reward_corr": 0.7244639267372566,
+      "objective/train/value_std": 0.0239410400390625,
+      "objective/train/weight_avg": 0.9995501637458801,
+      "objective/train/weighted_lm_loss": 1.3356695175170898,
+      "objective/train/weights_max": 1.4631335735321045,
+      "objective/train/weights_min": 0.23861059546470642,
+      "theoretical_loss": 3.3869962184059146,
+      "tokens_seen": 2595225600
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002156957149735195,
+      "loss": 0.0685,
+      "theoretical_loss": 3.3869962184059146,
+      "tokens_seen": 2595225600
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002156154710319371,
+      "loss": 0.0695,
+      "theoretical_loss": 3.386969361842911,
+      "tokens_seen": 2595487744
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002155352270903547,
+      "loss": 0.0671,
+      "theoretical_loss": 3.3869425087516953,
+      "tokens_seen": 2595749888
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021545498314877226,
+      "loss": 0.0673,
+      "theoretical_loss": 3.386915659131469,
+      "tokens_seen": 2596012032
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021537473920718987,
+      "loss": 0.0681,
+      "theoretical_loss": 3.386888812981433,
+      "tokens_seen": 2596274176
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021529449526560746,
+      "loss": 0.0642,
+      "theoretical_loss": 3.3868619703007883,
+      "tokens_seen": 2596536320
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021521425132402502,
+      "loss": 0.0641,
+      "theoretical_loss": 3.386835131088737,
+      "tokens_seen": 2596798464
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021513400738244263,
+      "loss": 0.0648,
+      "theoretical_loss": 3.386808295344479,
+      "tokens_seen": 2597060608
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021505376344086021,
+      "loss": 0.0654,
+      "theoretical_loss": 3.386781463067218,
+      "tokens_seen": 2597322752
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021497351949927783,
+      "loss": 0.0655,
+      "theoretical_loss": 3.386754634256156,
+      "tokens_seen": 2597584896
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021489327555769539,
+      "loss": 0.0644,
+      "theoretical_loss": 3.3867278089104946,
+      "tokens_seen": 2597847040
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.000214813031616113,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3867009870294376,
+      "tokens_seen": 2598109184
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021473278767453058,
+      "loss": 0.0686,
+      "theoretical_loss": 3.3866741686121875,
+      "tokens_seen": 2598371328
+    },
+    {
+      "epoch": 0.79,
+      "objective/train/advantage_avg": -0.00011098467803094536,
+      "objective/train/docs_used": 944930,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2284715175628662,
+      "objective/train/original_loss": 1.2284715175628662,
+      "objective/train/theoretical_loss": 3.3866607607022408,
+      "objective/train/tokens_used": 2618962400,
+      "objective/train/value_avg": -0.00836181640625,
+      "objective/train/value_loss": 0.0007939133793115616,
+      "objective/train/value_max": -3.916025161743164e-05,
+      "objective/train/value_min": -0.68017578125,
+      "objective/train/value_reward_corr": 0.5531434795339926,
+      "objective/train/value_std": 0.0196075439453125,
+      "objective/train/weight_avg": 1.0001929998397827,
+      "objective/train/weighted_lm_loss": 1.2284572124481201,
+      "objective/train/weights_max": 1.8591458797454834,
+      "objective/train/weights_min": 0.06395908445119858,
+      "theoretical_loss": 3.3866607607022408,
+      "tokens_seen": 2598502400
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021465254373294817,
+      "loss": 0.0646,
+      "theoretical_loss": 3.3866473536579473,
+      "tokens_seen": 2598633472
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021457229979136576,
+      "loss": 0.0684,
+      "theoretical_loss": 3.3866205421659217,
+      "tokens_seen": 2598895616
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021449205584978334,
+      "loss": 0.0674,
+      "theoretical_loss": 3.386593734135313,
+      "tokens_seen": 2599157760
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021441181190820095,
+      "loss": 0.0674,
+      "theoretical_loss": 3.3865669295653262,
+      "tokens_seen": 2599419904
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002143315679666185,
+      "loss": 0.0626,
+      "theoretical_loss": 3.3865401284551657,
+      "tokens_seen": 2599682048
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002142513240250361,
+      "loss": 0.0662,
+      "theoretical_loss": 3.3865133308040356,
+      "tokens_seen": 2599944192
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002141710800834537,
+      "loss": 0.0674,
+      "theoretical_loss": 3.3864865366111414,
+      "tokens_seen": 2600206336
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002140908361418713,
+      "loss": 0.0626,
+      "theoretical_loss": 3.3864597458756878,
+      "tokens_seen": 2600468480
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021401059220028888,
+      "loss": 0.0675,
+      "theoretical_loss": 3.38643295859688,
+      "tokens_seen": 2600730624
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021393034825870647,
+      "loss": 0.0695,
+      "theoretical_loss": 3.3864061747739242,
+      "tokens_seen": 2600992768
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021385010431712408,
+      "loss": 0.0696,
+      "theoretical_loss": 3.386379394406026,
+      "tokens_seen": 2601254912
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021376986037554166,
+      "loss": 0.0649,
+      "theoretical_loss": 3.3863526174923915,
+      "tokens_seen": 2601517056
+    },
+    {
+      "epoch": 0.79,
+      "objective/train/advantage_avg": 0.000880547275301069,
+      "objective/train/docs_used": 946148,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.5105973482131958,
+      "objective/train/original_loss": 1.5105971097946167,
+      "objective/train/theoretical_loss": 3.3863258440322275,
+      "objective/train/tokens_used": 2622239200,
+      "objective/train/value_avg": -0.006900787353515625,
+      "objective/train/value_loss": 0.00013182000839151442,
+      "objective/train/value_max": -2.956390380859375e-05,
+      "objective/train/value_min": -0.343017578125,
+      "objective/train/value_reward_corr": 0.8042078950290413,
+      "objective/train/value_std": 0.01509857177734375,
+      "objective/train/weight_avg": 1.0009452104568481,
+      "objective/train/weighted_lm_loss": 1.5125845670700073,
+      "objective/train/weights_max": 1.2113949060440063,
+      "objective/train/weights_min": 0.6516265273094177,
+      "theoretical_loss": 3.3863258440322275,
+      "tokens_seen": 2601779200
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021368961643395922,
+      "loss": 0.0669,
+      "theoretical_loss": 3.3863258440322275,
+      "tokens_seen": 2601779200
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021360937249237684,
+      "loss": 0.0677,
+      "theoretical_loss": 3.3862990740247403,
+      "tokens_seen": 2602041344
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021352912855079442,
+      "loss": 0.0653,
+      "theoretical_loss": 3.3862723074691377,
+      "tokens_seen": 2602303488
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.000213448884609212,
+      "loss": 0.0705,
+      "theoretical_loss": 3.3862455443646255,
+      "tokens_seen": 2602565632
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002133686406676296,
+      "loss": 0.0683,
+      "theoretical_loss": 3.3862187847104126,
+      "tokens_seen": 2602827776
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021328839672604718,
+      "loss": 0.0652,
+      "theoretical_loss": 3.3861920285057057,
+      "tokens_seen": 2603089920
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002132081527844648,
+      "loss": 0.0661,
+      "theoretical_loss": 3.386165275749714,
+      "tokens_seen": 2603352064
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021312790884288235,
+      "loss": 0.0652,
+      "theoretical_loss": 3.3861385264416444,
+      "tokens_seen": 2603614208
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021304766490129996,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3861117805807064,
+      "tokens_seen": 2603876352
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021296742095971755,
+      "loss": 0.0663,
+      "theoretical_loss": 3.3860850381661085,
+      "tokens_seen": 2604138496
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021288717701813516,
+      "loss": 0.0674,
+      "theoretical_loss": 3.3860582991970594,
+      "tokens_seen": 2604400640
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021280693307655272,
+      "loss": 0.0668,
+      "theoretical_loss": 3.3860315636727694,
+      "tokens_seen": 2604662784
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002127266891349703,
+      "loss": 0.0667,
+      "theoretical_loss": 3.386004831592447,
+      "tokens_seen": 2604924928
+    },
+    {
+      "epoch": 0.79,
+      "objective/train/advantage_avg": -0.00013600836973637342,
+      "objective/train/docs_used": 947237,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.1746832132339478,
+      "objective/train/original_loss": 1.1746830940246582,
+      "objective/train/theoretical_loss": 3.385991466843527,
+      "objective/train/tokens_used": 2625516000,
+      "objective/train/value_avg": -0.00673675537109375,
+      "objective/train/value_loss": 0.00014905710122548044,
+      "objective/train/value_max": -3.534555435180664e-05,
+      "objective/train/value_min": -0.218994140625,
+      "objective/train/value_reward_corr": 0.7247144381162496,
+      "objective/train/value_std": 0.0120391845703125,
+      "objective/train/weight_avg": 0.9999330639839172,
+      "objective/train/weighted_lm_loss": 1.1741994619369507,
+      "objective/train/weights_max": 1.1198102235794067,
+      "objective/train/weights_min": 0.3691932260990143,
+      "theoretical_loss": 3.385991466843527,
+      "tokens_seen": 2605056000
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021264644519338792,
+      "loss": 0.0638,
+      "theoretical_loss": 3.385978102955303,
+      "tokens_seen": 2605187072
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021256620125180547,
+      "loss": 0.0647,
+      "theoretical_loss": 3.3859513777605468,
+      "tokens_seen": 2605449216
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002124859573102231,
+      "loss": 0.065,
+      "theoretical_loss": 3.385924656007389,
+      "tokens_seen": 2605711360
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021240571336864067,
+      "loss": 0.0658,
+      "theoretical_loss": 3.38589793769504,
+      "tokens_seen": 2605973504
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021232546942705826,
+      "loss": 0.0668,
+      "theoretical_loss": 3.3858712228227117,
+      "tokens_seen": 2606235648
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021224522548547584,
+      "loss": 0.0632,
+      "theoretical_loss": 3.385844511389614,
+      "tokens_seen": 2606497792
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021216498154389343,
+      "loss": 0.0693,
+      "theoretical_loss": 3.385817803394959,
+      "tokens_seen": 2606759936
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021208473760231104,
+      "loss": 0.0651,
+      "theoretical_loss": 3.3857910988379576,
+      "tokens_seen": 2607022080
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021200449366072863,
+      "loss": 0.0654,
+      "theoretical_loss": 3.385764397717823,
+      "tokens_seen": 2607284224
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002119242497191462,
+      "loss": 0.0659,
+      "theoretical_loss": 3.3857377000337663,
+      "tokens_seen": 2607546368
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002118440057775638,
+      "loss": 0.0631,
+      "theoretical_loss": 3.3857110057850006,
+      "tokens_seen": 2607808512
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021176376183598138,
+      "loss": 0.0662,
+      "theoretical_loss": 3.3856843149707383,
+      "tokens_seen": 2608070656
+    },
+    {
+      "epoch": 0.79,
+      "objective/train/advantage_avg": 0.001057982794009149,
+      "objective/train/docs_used": 948381,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.307920217514038,
+      "objective/train/original_loss": 1.307920217514038,
+      "objective/train/theoretical_loss": 3.3856576275901924,
+      "objective/train/tokens_used": 2628792800,
+      "objective/train/value_avg": -0.0081939697265625,
+      "objective/train/value_loss": 0.0002661271137185395,
+      "objective/train/value_max": -2.7120113372802734e-05,
+      "objective/train/value_min": -0.9365234375,
+      "objective/train/value_reward_corr": 0.7017694738472946,
+      "objective/train/value_std": 0.018798828125,
+      "objective/train/weight_avg": 1.001183032989502,
+      "objective/train/weighted_lm_loss": 1.3090044260025024,
+      "objective/train/weights_max": 1.9032342433929443,
+      "objective/train/weights_min": 0.37690240144729614,
+      "theoretical_loss": 3.3856576275901924,
+      "tokens_seen": 2608332800
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021168351789439897,
+      "loss": 0.0635,
+      "theoretical_loss": 3.3856576275901924,
+      "tokens_seen": 2608332800
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021160327395281655,
+      "loss": 0.0665,
+      "theoretical_loss": 3.385630943642576,
+      "tokens_seen": 2608594944
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021152303001123417,
+      "loss": 0.0646,
+      "theoretical_loss": 3.3856042631271026,
+      "tokens_seen": 2608857088
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021144278606965175,
+      "loss": 0.0659,
+      "theoretical_loss": 3.385577586042986,
+      "tokens_seen": 2609119232
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002113625421280693,
+      "loss": 0.0661,
+      "theoretical_loss": 3.385550912389441,
+      "tokens_seen": 2609381376
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021128229818648692,
+      "loss": 0.0671,
+      "theoretical_loss": 3.385524242165681,
+      "tokens_seen": 2609643520
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002112020542449045,
+      "loss": 0.0635,
+      "theoretical_loss": 3.3854975753709207,
+      "tokens_seen": 2609905664
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021112181030332212,
+      "loss": 0.0672,
+      "theoretical_loss": 3.3854709120043744,
+      "tokens_seen": 2610167808
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021104156636173968,
+      "loss": 0.0653,
+      "theoretical_loss": 3.3854442520652577,
+      "tokens_seen": 2610429952
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002109613224201573,
+      "loss": 0.0653,
+      "theoretical_loss": 3.385417595552786,
+      "tokens_seen": 2610692096
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021088107847857488,
+      "loss": 0.0674,
+      "theoretical_loss": 3.3853909424661746,
+      "tokens_seen": 2610954240
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021080083453699246,
+      "loss": 0.0675,
+      "theoretical_loss": 3.3853642928046397,
+      "tokens_seen": 2611216384
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021072059059541005,
+      "loss": 0.0668,
+      "theoretical_loss": 3.3853376465673968,
+      "tokens_seen": 2611478528
+    },
+    {
+      "epoch": 0.79,
+      "objective/train/advantage_avg": 0.0015518314903602004,
+      "objective/train/docs_used": 949550,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.2171560525894165,
+      "objective/train/original_loss": 1.217156171798706,
+      "objective/train/theoretical_loss": 3.38532432473264,
+      "objective/train/tokens_used": 2632069600,
+      "objective/train/value_avg": -0.007183074951171875,
+      "objective/train/value_loss": 0.00020713287813123316,
+      "objective/train/value_max": -3.24249267578125e-05,
+      "objective/train/value_min": -0.37255859375,
+      "objective/train/value_reward_corr": 0.6996092584226039,
+      "objective/train/value_std": 0.01509857177734375,
+      "objective/train/weight_avg": 1.0016437768936157,
+      "objective/train/weighted_lm_loss": 1.218762993812561,
+      "objective/train/weights_max": 1.308711290359497,
+      "objective/train/weights_min": 0.37102511525154114,
+      "theoretical_loss": 3.38532432473264,
+      "tokens_seen": 2611609600
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021064034665382763,
+      "loss": 0.07,
+      "theoretical_loss": 3.3853110037536625,
+      "tokens_seen": 2611740672
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021056010271224525,
+      "loss": 0.0681,
+      "theoretical_loss": 3.3852843643626533,
+      "tokens_seen": 2612002816
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002104798587706628,
+      "loss": 0.069,
+      "theoretical_loss": 3.3852577283935865,
+      "tokens_seen": 2612264960
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002103996148290804,
+      "loss": 0.0681,
+      "theoretical_loss": 3.385231095845678,
+      "tokens_seen": 2612527104
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.000210319370887498,
+      "loss": 0.0624,
+      "theoretical_loss": 3.385204466718147,
+      "tokens_seen": 2612789248
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002102391269459156,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3851778410102096,
+      "tokens_seen": 2613051392
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021015888300433317,
+      "loss": 0.0673,
+      "theoretical_loss": 3.3851512187210844,
+      "tokens_seen": 2613313536
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00021007863906275076,
+      "loss": 0.066,
+      "theoretical_loss": 3.3851245998499895,
+      "tokens_seen": 2613575680
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020999839512116837,
+      "loss": 0.0669,
+      "theoretical_loss": 3.385097984396143,
+      "tokens_seen": 2613837824
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020991815117958596,
+      "loss": 0.0656,
+      "theoretical_loss": 3.3850713723587633,
+      "tokens_seen": 2614099968
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020983790723800352,
+      "loss": 0.065,
+      "theoretical_loss": 3.38504476373707,
+      "tokens_seen": 2614362112
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020975766329642113,
+      "loss": 0.0652,
+      "theoretical_loss": 3.3850181585302823,
+      "tokens_seen": 2614624256
+    },
+    {
+      "epoch": 0.79,
+      "objective/train/advantage_avg": 2.870370735763572e-05,
+      "objective/train/docs_used": 950739,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3297094106674194,
+      "objective/train/original_loss": 1.3297092914581299,
+      "objective/train/theoretical_loss": 3.384991556737619,
+      "objective/train/tokens_used": 2635346400,
+      "objective/train/value_avg": -0.00719451904296875,
+      "objective/train/value_loss": 0.00031871211831457913,
+      "objective/train/value_max": -4.7206878662109375e-05,
+      "objective/train/value_min": -0.94091796875,
+      "objective/train/value_reward_corr": 0.7650113923665571,
+      "objective/train/value_std": 0.0180816650390625,
+      "objective/train/weight_avg": 1.0001702308654785,
+      "objective/train/weighted_lm_loss": 1.3302415609359741,
+      "objective/train/weights_max": 1.8773905038833618,
+      "objective/train/weights_min": 0.3694460988044739,
+      "theoretical_loss": 3.384991556737619,
+      "tokens_seen": 2614886400
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020967741935483871,
+      "loss": 0.0634,
+      "theoretical_loss": 3.384991556737619,
+      "tokens_seen": 2614886400
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002095971754132563,
+      "loss": 0.0666,
+      "theoretical_loss": 3.3849649583583004,
+      "tokens_seen": 2615148544
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020951693147167389,
+      "loss": 0.0657,
+      "theoretical_loss": 3.3849383633915457,
+      "tokens_seen": 2615410688
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002094366875300915,
+      "loss": 0.065,
+      "theoretical_loss": 3.3849117718365758,
+      "tokens_seen": 2615672832
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020935644358850908,
+      "loss": 0.067,
+      "theoretical_loss": 3.38488518369261,
+      "tokens_seen": 2615934976
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020927619964692664,
+      "loss": 0.0653,
+      "theoretical_loss": 3.384858598958871,
+      "tokens_seen": 2616197120
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020919595570534426,
+      "loss": 0.0655,
+      "theoretical_loss": 3.384832017634578,
+      "tokens_seen": 2616459264
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020911571176376184,
+      "loss": 0.0651,
+      "theoretical_loss": 3.384805439718953,
+      "tokens_seen": 2616721408
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020903546782217945,
+      "loss": 0.0656,
+      "theoretical_loss": 3.384778865211217,
+      "tokens_seen": 2616983552
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.000208955223880597,
+      "loss": 0.0661,
+      "theoretical_loss": 3.3847522941105925,
+      "tokens_seen": 2617245696
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002088749799390146,
+      "loss": 0.0643,
+      "theoretical_loss": 3.384725726416301,
+      "tokens_seen": 2617507840
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002087947359974322,
+      "loss": 0.0646,
+      "theoretical_loss": 3.384699162127564,
+      "tokens_seen": 2617769984
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020871449205584977,
+      "loss": 0.0673,
+      "theoretical_loss": 3.3846726012436057,
+      "tokens_seen": 2618032128
+    },
+    {
+      "epoch": 0.79,
+      "objective/train/advantage_avg": 0.00038211545324884355,
+      "objective/train/docs_used": 951960,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.40494704246521,
+      "objective/train/original_loss": 1.40494704246521,
+      "objective/train/theoretical_loss": 3.384659322078175,
+      "objective/train/tokens_used": 2638623200,
+      "objective/train/value_avg": -0.01016998291015625,
+      "objective/train/value_loss": 0.00022523697407450527,
+      "objective/train/value_max": -4.297494888305664e-05,
+      "objective/train/value_min": -0.306640625,
+      "objective/train/value_reward_corr": 0.7207486398264816,
+      "objective/train/value_std": 0.0157012939453125,
+      "objective/train/weight_avg": 1.0004863739013672,
+      "objective/train/weighted_lm_loss": 1.4053072929382324,
+      "objective/train/weights_max": 1.181889295578003,
+      "objective/train/weights_min": 0.38233569264411926,
+      "theoretical_loss": 3.384659322078175,
+      "tokens_seen": 2618163200
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020863424811426738,
+      "loss": 0.0636,
+      "theoretical_loss": 3.3846460437636474,
+      "tokens_seen": 2618294272
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020855400417268497,
+      "loss": 0.0635,
+      "theoretical_loss": 3.3846194896869126,
+      "tokens_seen": 2618556416
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020847376023110258,
+      "loss": 0.0683,
+      "theoretical_loss": 3.3845929390126246,
+      "tokens_seen": 2618818560
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020839351628952014,
+      "loss": 0.0672,
+      "theoretical_loss": 3.384566391740007,
+      "tokens_seen": 2619080704
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020831327234793772,
+      "loss": 0.0674,
+      "theoretical_loss": 3.3845398478682833,
+      "tokens_seen": 2619342848
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020823302840635534,
+      "loss": 0.0637,
+      "theoretical_loss": 3.384513307396678,
+      "tokens_seen": 2619604992
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020815278446477292,
+      "loss": 0.0647,
+      "theoretical_loss": 3.384486770324415,
+      "tokens_seen": 2619867136
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002080725405231905,
+      "loss": 0.0657,
+      "theoretical_loss": 3.3844602366507184,
+      "tokens_seen": 2620129280
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.0002079922965816081,
+      "loss": 0.0667,
+      "theoretical_loss": 3.3844337063748138,
+      "tokens_seen": 2620391424
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020791205264002568,
+      "loss": 0.0683,
+      "theoretical_loss": 3.384407179495926,
+      "tokens_seen": 2620653568
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020783180869844326,
+      "loss": 0.0634,
+      "theoretical_loss": 3.38438065601328,
+      "tokens_seen": 2620915712
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020775156475686085,
+      "loss": 0.0627,
+      "theoretical_loss": 3.3843541359261016,
+      "tokens_seen": 2621177856
+    },
+    {
+      "debugging/Compilability": 1.0,
+      "debugging/distinct-1-grams": 0.7183788677974512,
+      "debugging/entropy-1-grams": 5.311092611595205,
+      "debugging/length": 475.35714285714283,
+      "debugging/num_segments": 14,
+      "debugging/raw_token_scores_avg": 0.007209094241261482,
+      "debugging/raw_token_scores_std": 0.01626458764076233,
+      "debugging/score": 0.005309415865714915,
+      "debugging/score_std": 0.0067471201632161425,
+      "epoch": 0.79,
+      "objective/train/advantage_avg": -0.00040069257374852896,
+      "objective/train/docs_used": 953088,
+      "objective/train/instantaneous_batch_size": 32,
+      "objective/train/instantaneous_microbatch_size": 32768,
+      "objective/train/lm_loss": 1.3306788206100464,
+      "objective/train/original_loss": 1.3306788206100464,
+      "objective/train/theoretical_loss": 3.3843276192336167,
+      "objective/train/tokens_used": 2641900000,
+      "objective/train/value_avg": -0.006809234619140625,
+      "objective/train/value_loss": 0.00010478802141733468,
+      "objective/train/value_max": -2.86102294921875e-05,
+      "objective/train/value_min": -0.298095703125,
+      "objective/train/value_reward_corr": 0.778743762677434,
+      "objective/train/value_std": 0.01197052001953125,
+      "objective/train/weight_avg": 0.9996510744094849,
+      "objective/train/weighted_lm_loss": 1.3306902647018433,
+      "objective/train/weights_max": 1.1428725719451904,
+      "objective/train/weights_min": 0.8198684453964233,
+      "theoretical_loss": 3.3843276192336167,
+      "tokens_seen": 2621440000
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 0.00020767132081527846,
+      "loss": 0.0664,
+      "theoretical_loss": 3.3843276192336167,
+      "tokens_seen": 2621440000
+    }
+  ],
+  "max_steps": 12588,
+  "num_train_epochs": 9223372036854775807,
+  "total_flos": 1.33782728343552e+18,
+  "trial_name": null,
+  "trial_params": null
+}