| { | |
| "best_metric": 0.4387560784816742, | |
| "best_model_checkpoint": "model_output/e2e_opentable_5_way__19684-shot__seed-42__lstm/checkpoint-1250", | |
| "epoch": 8.116883116883116, | |
| "global_step": 1250, | |
| "is_hyper_param_search": false, | |
| "is_local_process_zero": true, | |
| "is_world_process_zero": true, | |
| "log_history": [ | |
| { | |
| "epoch": 0.13, | |
| "learning_rate": 4.9783362218370885e-05, | |
| "loss": 20.6562, | |
| "step": 20 | |
| }, | |
| { | |
| "epoch": 0.26, | |
| "learning_rate": 4.956672443674177e-05, | |
| "loss": 18.0703, | |
| "step": 40 | |
| }, | |
| { | |
| "epoch": 0.32, | |
| "eval_distillation_accuracy_counterfactual": 0.48717290918419703, | |
| "eval_distillation_accuracy_factual": 0.9012314007183171, | |
| "eval_distillation_f1_counterfactual": 0.4535140666639451, | |
| "eval_distillation_f1_factual": 0.8957696500942338, | |
| "eval_groundtruth_accuracy_counterfactual": 0.38558234992303747, | |
| "eval_groundtruth_f1_counterfactual": 0.3510147662827972, | |
| "eval_groundtruth_f1_factual": 0.5809634631420686, | |
| "eval_icace_cosine": 0.5996042490005493, | |
| "eval_icace_l2": 0.6879788637161255, | |
| "eval_icace_normdiff": 0.6130202412605286, | |
| "eval_loss": 18.372575759887695, | |
| "eval_runtime": 4.1379, | |
| "eval_samples_per_second": 942.017, | |
| "eval_steps_per_second": 7.492, | |
| "step": 50 | |
| }, | |
| { | |
| "epoch": 0.39, | |
| "learning_rate": 4.935008665511265e-05, | |
| "loss": 16.8521, | |
| "step": 60 | |
| }, | |
| { | |
| "epoch": 0.52, | |
| "learning_rate": 4.913344887348354e-05, | |
| "loss": 15.1477, | |
| "step": 80 | |
| }, | |
| { | |
| "epoch": 0.65, | |
| "learning_rate": 4.891681109185442e-05, | |
| "loss": 13.5861, | |
| "step": 100 | |
| }, | |
| { | |
| "epoch": 0.65, | |
| "eval_distillation_accuracy_counterfactual": 0.5200102616726526, | |
| "eval_distillation_accuracy_factual": 0.8473576192919446, | |
| "eval_distillation_f1_counterfactual": 0.4895334324691376, | |
| "eval_distillation_f1_factual": 0.8392241931747526, | |
| "eval_groundtruth_accuracy_counterfactual": 0.4238070805541303, | |
| "eval_groundtruth_f1_counterfactual": 0.3924731874181019, | |
| "eval_groundtruth_f1_factual": 0.5673025956236939, | |
| "eval_icace_cosine": 0.5384137630462646, | |
| "eval_icace_l2": 0.6406373977661133, | |
| "eval_icace_normdiff": 0.5294525623321533, | |
| "eval_loss": 13.979488372802734, | |
| "eval_runtime": 4.4493, | |
| "eval_samples_per_second": 876.093, | |
| "eval_steps_per_second": 6.967, | |
| "step": 100 | |
| }, | |
| { | |
| "epoch": 0.78, | |
| "learning_rate": 4.8700173310225307e-05, | |
| "loss": 12.2237, | |
| "step": 120 | |
| }, | |
| { | |
| "epoch": 0.91, | |
| "learning_rate": 4.848353552859619e-05, | |
| "loss": 11.2675, | |
| "step": 140 | |
| }, | |
| { | |
| "epoch": 0.97, | |
| "eval_distillation_accuracy_counterfactual": 0.5726013340174448, | |
| "eval_distillation_accuracy_factual": 0.8022062596203181, | |
| "eval_distillation_f1_counterfactual": 0.5314200417069295, | |
| "eval_distillation_f1_factual": 0.7853514931295182, | |
| "eval_groundtruth_accuracy_counterfactual": 0.4917906618778861, | |
| "eval_groundtruth_f1_counterfactual": 0.4494045856838918, | |
| "eval_groundtruth_f1_factual": 0.5299931970242554, | |
| "eval_icace_cosine": 0.4923843741416931, | |
| "eval_icace_l2": 0.561776340007782, | |
| "eval_icace_normdiff": 0.4008543789386749, | |
| "eval_loss": 10.179123878479004, | |
| "eval_runtime": 4.1977, | |
| "eval_samples_per_second": 928.603, | |
| "eval_steps_per_second": 7.385, | |
| "step": 150 | |
| }, | |
| { | |
| "epoch": 1.04, | |
| "learning_rate": 4.826689774696707e-05, | |
| "loss": 9.9178, | |
| "step": 160 | |
| }, | |
| { | |
| "epoch": 1.17, | |
| "learning_rate": 4.8050259965337955e-05, | |
| "loss": 9.1541, | |
| "step": 180 | |
| }, | |
| { | |
| "epoch": 1.3, | |
| "learning_rate": 4.7833622183708845e-05, | |
| "loss": 8.7158, | |
| "step": 200 | |
| }, | |
| { | |
| "epoch": 1.3, | |
| "eval_distillation_accuracy_counterfactual": 0.5826064648537712, | |
| "eval_distillation_accuracy_factual": 0.8129810159055926, | |
| "eval_distillation_f1_counterfactual": 0.5429038638893491, | |
| "eval_distillation_f1_factual": 0.7982170975645125, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5082093381221139, | |
| "eval_groundtruth_f1_counterfactual": 0.4622598550088801, | |
| "eval_groundtruth_f1_factual": 0.53329740892104, | |
| "eval_icace_cosine": 0.4688790440559387, | |
| "eval_icace_l2": 0.538577139377594, | |
| "eval_icace_normdiff": 0.3778749108314514, | |
| "eval_loss": 9.010956764221191, | |
| "eval_runtime": 4.0943, | |
| "eval_samples_per_second": 952.058, | |
| "eval_steps_per_second": 7.572, | |
| "step": 200 | |
| }, | |
| { | |
| "epoch": 1.43, | |
| "learning_rate": 4.761698440207972e-05, | |
| "loss": 8.5692, | |
| "step": 220 | |
| }, | |
| { | |
| "epoch": 1.56, | |
| "learning_rate": 4.740034662045061e-05, | |
| "loss": 8.0816, | |
| "step": 240 | |
| }, | |
| { | |
| "epoch": 1.62, | |
| "eval_distillation_accuracy_counterfactual": 0.5928681375064135, | |
| "eval_distillation_accuracy_factual": 0.8070805541303232, | |
| "eval_distillation_f1_counterfactual": 0.5529708874649586, | |
| "eval_distillation_f1_factual": 0.79399343360367, | |
| "eval_groundtruth_accuracy_counterfactual": 0.517444843509492, | |
| "eval_groundtruth_f1_counterfactual": 0.4725321895461628, | |
| "eval_groundtruth_f1_factual": 0.5345191071930586, | |
| "eval_icace_cosine": 0.46347475051879883, | |
| "eval_icace_l2": 0.527389407157898, | |
| "eval_icace_normdiff": 0.37002304196357727, | |
| "eval_loss": 8.498320579528809, | |
| "eval_runtime": 4.3399, | |
| "eval_samples_per_second": 898.168, | |
| "eval_steps_per_second": 7.143, | |
| "step": 250 | |
| }, | |
| { | |
| "epoch": 1.69, | |
| "learning_rate": 4.7183708838821494e-05, | |
| "loss": 8.4685, | |
| "step": 260 | |
| }, | |
| { | |
| "epoch": 1.82, | |
| "learning_rate": 4.6967071057192376e-05, | |
| "loss": 7.8803, | |
| "step": 280 | |
| }, | |
| { | |
| "epoch": 1.95, | |
| "learning_rate": 4.675043327556326e-05, | |
| "loss": 7.6873, | |
| "step": 300 | |
| }, | |
| { | |
| "epoch": 1.95, | |
| "eval_distillation_accuracy_counterfactual": 0.6064648537711647, | |
| "eval_distillation_accuracy_factual": 0.8178553104155978, | |
| "eval_distillation_f1_counterfactual": 0.5629061661390706, | |
| "eval_distillation_f1_factual": 0.8023836086835677, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5279630579784504, | |
| "eval_groundtruth_f1_counterfactual": 0.47789130421114673, | |
| "eval_groundtruth_f1_factual": 0.5344257070862026, | |
| "eval_icace_cosine": 0.45601892471313477, | |
| "eval_icace_l2": 0.5144051909446716, | |
| "eval_icace_normdiff": 0.3617022633552551, | |
| "eval_loss": 8.192581176757812, | |
| "eval_runtime": 5.0158, | |
| "eval_samples_per_second": 777.137, | |
| "eval_steps_per_second": 6.18, | |
| "step": 300 | |
| }, | |
| { | |
| "epoch": 2.08, | |
| "learning_rate": 4.653379549393415e-05, | |
| "loss": 7.5662, | |
| "step": 320 | |
| }, | |
| { | |
| "epoch": 2.21, | |
| "learning_rate": 4.6317157712305025e-05, | |
| "loss": 7.6061, | |
| "step": 340 | |
| }, | |
| { | |
| "epoch": 2.27, | |
| "eval_distillation_accuracy_counterfactual": 0.6095433555669574, | |
| "eval_distillation_accuracy_factual": 0.8217034376603386, | |
| "eval_distillation_f1_counterfactual": 0.5684665160653969, | |
| "eval_distillation_f1_factual": 0.8081531276248081, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5356593124679323, | |
| "eval_groundtruth_f1_counterfactual": 0.4879170474066199, | |
| "eval_groundtruth_f1_factual": 0.5428429092164265, | |
| "eval_icace_cosine": 0.45691362023353577, | |
| "eval_icace_l2": 0.5149694681167603, | |
| "eval_icace_normdiff": 0.3580142557621002, | |
| "eval_loss": 8.089337348937988, | |
| "eval_runtime": 4.1368, | |
| "eval_samples_per_second": 942.268, | |
| "eval_steps_per_second": 7.494, | |
| "step": 350 | |
| }, | |
| { | |
| "epoch": 2.34, | |
| "learning_rate": 4.6100519930675915e-05, | |
| "loss": 7.4725, | |
| "step": 360 | |
| }, | |
| { | |
| "epoch": 2.47, | |
| "learning_rate": 4.58838821490468e-05, | |
| "loss": 7.3287, | |
| "step": 380 | |
| }, | |
| { | |
| "epoch": 2.6, | |
| "learning_rate": 4.566724436741768e-05, | |
| "loss": 7.5374, | |
| "step": 400 | |
| }, | |
| { | |
| "epoch": 2.6, | |
| "eval_distillation_accuracy_counterfactual": 0.6159569009748589, | |
| "eval_distillation_accuracy_factual": 0.8309389430477168, | |
| "eval_distillation_f1_counterfactual": 0.5691376817104438, | |
| "eval_distillation_f1_factual": 0.8148479579609189, | |
| "eval_groundtruth_accuracy_counterfactual": 0.539250897896357, | |
| "eval_groundtruth_f1_counterfactual": 0.485360382747431, | |
| "eval_groundtruth_f1_factual": 0.5317474369735481, | |
| "eval_icace_cosine": 0.45355308055877686, | |
| "eval_icace_l2": 0.5047734379768372, | |
| "eval_icace_normdiff": 0.3515755236148834, | |
| "eval_loss": 7.922365188598633, | |
| "eval_runtime": 5.8925, | |
| "eval_samples_per_second": 661.514, | |
| "eval_steps_per_second": 5.261, | |
| "step": 400 | |
| }, | |
| { | |
| "epoch": 2.73, | |
| "learning_rate": 4.5450606585788563e-05, | |
| "loss": 7.1126, | |
| "step": 420 | |
| }, | |
| { | |
| "epoch": 2.86, | |
| "learning_rate": 4.5233968804159446e-05, | |
| "loss": 7.3885, | |
| "step": 440 | |
| }, | |
| { | |
| "epoch": 2.92, | |
| "eval_distillation_accuracy_counterfactual": 0.6144176500769626, | |
| "eval_distillation_accuracy_factual": 0.8327347357619292, | |
| "eval_distillation_f1_counterfactual": 0.5729661409695295, | |
| "eval_distillation_f1_factual": 0.8192940502978703, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5407901487942535, | |
| "eval_groundtruth_f1_counterfactual": 0.4946759059049534, | |
| "eval_groundtruth_f1_factual": 0.5452799385658833, | |
| "eval_icace_cosine": 0.45755502581596375, | |
| "eval_icace_l2": 0.506329357624054, | |
| "eval_icace_normdiff": 0.35191622376441956, | |
| "eval_loss": 7.8961567878723145, | |
| "eval_runtime": 4.1665, | |
| "eval_samples_per_second": 935.55, | |
| "eval_steps_per_second": 7.44, | |
| "step": 450 | |
| }, | |
| { | |
| "epoch": 2.99, | |
| "learning_rate": 4.501733102253033e-05, | |
| "loss": 7.0656, | |
| "step": 460 | |
| }, | |
| { | |
| "epoch": 3.12, | |
| "learning_rate": 4.480069324090121e-05, | |
| "loss": 7.16, | |
| "step": 480 | |
| }, | |
| { | |
| "epoch": 3.25, | |
| "learning_rate": 4.45840554592721e-05, | |
| "loss": 6.7453, | |
| "step": 500 | |
| }, | |
| { | |
| "epoch": 3.25, | |
| "eval_distillation_accuracy_counterfactual": 0.612365315546434, | |
| "eval_distillation_accuracy_factual": 0.8273473576192919, | |
| "eval_distillation_f1_counterfactual": 0.5700286878042866, | |
| "eval_distillation_f1_factual": 0.8116751283042352, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5461775269368907, | |
| "eval_groundtruth_f1_counterfactual": 0.4982594429639036, | |
| "eval_groundtruth_f1_factual": 0.5426844401130816, | |
| "eval_icace_cosine": 0.4582451283931732, | |
| "eval_icace_l2": 0.5082744359970093, | |
| "eval_icace_normdiff": 0.3508155047893524, | |
| "eval_loss": 7.861159801483154, | |
| "eval_runtime": 6.4679, | |
| "eval_samples_per_second": 602.668, | |
| "eval_steps_per_second": 4.793, | |
| "step": 500 | |
| }, | |
| { | |
| "epoch": 3.38, | |
| "learning_rate": 4.436741767764298e-05, | |
| "loss": 6.9515, | |
| "step": 520 | |
| }, | |
| { | |
| "epoch": 3.51, | |
| "learning_rate": 4.415077989601387e-05, | |
| "loss": 6.6913, | |
| "step": 540 | |
| }, | |
| { | |
| "epoch": 3.57, | |
| "eval_distillation_accuracy_counterfactual": 0.6228835300153925, | |
| "eval_distillation_accuracy_factual": 0.823242688558235, | |
| "eval_distillation_f1_counterfactual": 0.5797100592542744, | |
| "eval_distillation_f1_factual": 0.8054169495093865, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5515649050795279, | |
| "eval_groundtruth_f1_counterfactual": 0.500957403465055, | |
| "eval_groundtruth_f1_factual": 0.5372992623096391, | |
| "eval_icace_cosine": 0.4541945159435272, | |
| "eval_icace_l2": 0.5000237226486206, | |
| "eval_icace_normdiff": 0.3448871374130249, | |
| "eval_loss": 7.762178897857666, | |
| "eval_runtime": 4.1547, | |
| "eval_samples_per_second": 938.211, | |
| "eval_steps_per_second": 7.461, | |
| "step": 550 | |
| }, | |
| { | |
| "epoch": 3.64, | |
| "learning_rate": 4.393414211438475e-05, | |
| "loss": 6.8003, | |
| "step": 560 | |
| }, | |
| { | |
| "epoch": 3.77, | |
| "learning_rate": 4.371750433275563e-05, | |
| "loss": 7.0336, | |
| "step": 580 | |
| }, | |
| { | |
| "epoch": 3.9, | |
| "learning_rate": 4.3500866551126516e-05, | |
| "loss": 6.8497, | |
| "step": 600 | |
| }, | |
| { | |
| "epoch": 3.9, | |
| "eval_distillation_accuracy_counterfactual": 0.6177526936890713, | |
| "eval_distillation_accuracy_factual": 0.8150333504361211, | |
| "eval_distillation_f1_counterfactual": 0.5812759881530707, | |
| "eval_distillation_f1_factual": 0.8048351370604768, | |
| "eval_groundtruth_accuracy_counterfactual": 0.543099025141098, | |
| "eval_groundtruth_f1_counterfactual": 0.502597773992042, | |
| "eval_groundtruth_f1_factual": 0.5413490240682356, | |
| "eval_icace_cosine": 0.4559078812599182, | |
| "eval_icace_l2": 0.5052775740623474, | |
| "eval_icace_normdiff": 0.34785109758377075, | |
| "eval_loss": 7.774956703186035, | |
| "eval_runtime": 5.8188, | |
| "eval_samples_per_second": 669.896, | |
| "eval_steps_per_second": 5.328, | |
| "step": 600 | |
| }, | |
| { | |
| "epoch": 4.03, | |
| "learning_rate": 4.3284228769497406e-05, | |
| "loss": 6.4975, | |
| "step": 620 | |
| }, | |
| { | |
| "epoch": 4.16, | |
| "learning_rate": 4.306759098786828e-05, | |
| "loss": 6.5642, | |
| "step": 640 | |
| }, | |
| { | |
| "epoch": 4.22, | |
| "eval_distillation_accuracy_counterfactual": 0.6136480246280144, | |
| "eval_distillation_accuracy_factual": 0.8183683940482298, | |
| "eval_distillation_f1_counterfactual": 0.580211267327247, | |
| "eval_distillation_f1_factual": 0.8068432506573439, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5454079014879425, | |
| "eval_groundtruth_f1_counterfactual": 0.5053418375187448, | |
| "eval_groundtruth_f1_factual": 0.5441340640231574, | |
| "eval_icace_cosine": 0.4580763578414917, | |
| "eval_icace_l2": 0.5093169212341309, | |
| "eval_icace_normdiff": 0.3487666845321655, | |
| "eval_loss": 7.786745548248291, | |
| "eval_runtime": 4.0712, | |
| "eval_samples_per_second": 957.454, | |
| "eval_steps_per_second": 7.614, | |
| "step": 650 | |
| }, | |
| { | |
| "epoch": 4.29, | |
| "learning_rate": 4.285095320623917e-05, | |
| "loss": 6.4044, | |
| "step": 660 | |
| }, | |
| { | |
| "epoch": 4.42, | |
| "learning_rate": 4.2634315424610055e-05, | |
| "loss": 6.4254, | |
| "step": 680 | |
| }, | |
| { | |
| "epoch": 4.55, | |
| "learning_rate": 4.241767764298094e-05, | |
| "loss": 6.393, | |
| "step": 700 | |
| }, | |
| { | |
| "epoch": 4.55, | |
| "eval_distillation_accuracy_counterfactual": 0.6182657773217034, | |
| "eval_distillation_accuracy_factual": 0.8160595177013853, | |
| "eval_distillation_f1_counterfactual": 0.5808522807878681, | |
| "eval_distillation_f1_factual": 0.8040768770971043, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5454079014879425, | |
| "eval_groundtruth_f1_counterfactual": 0.5006294773113318, | |
| "eval_groundtruth_f1_factual": 0.5459691019650267, | |
| "eval_icace_cosine": 0.4525621235370636, | |
| "eval_icace_l2": 0.5032183527946472, | |
| "eval_icace_normdiff": 0.3443147540092468, | |
| "eval_loss": 7.749052047729492, | |
| "eval_runtime": 5.3047, | |
| "eval_samples_per_second": 734.818, | |
| "eval_steps_per_second": 5.844, | |
| "step": 700 | |
| }, | |
| { | |
| "epoch": 4.68, | |
| "learning_rate": 4.220103986135182e-05, | |
| "loss": 6.3202, | |
| "step": 720 | |
| }, | |
| { | |
| "epoch": 4.81, | |
| "learning_rate": 4.198440207972271e-05, | |
| "loss": 6.6167, | |
| "step": 740 | |
| }, | |
| { | |
| "epoch": 4.87, | |
| "eval_distillation_accuracy_counterfactual": 0.6182657773217034, | |
| "eval_distillation_accuracy_factual": 0.8122113904566445, | |
| "eval_distillation_f1_counterfactual": 0.5818063896260541, | |
| "eval_distillation_f1_factual": 0.7988404215810908, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5484864032837352, | |
| "eval_groundtruth_f1_counterfactual": 0.5052584515576525, | |
| "eval_groundtruth_f1_factual": 0.5394581551303677, | |
| "eval_icace_cosine": 0.4518362283706665, | |
| "eval_icace_l2": 0.501703679561615, | |
| "eval_icace_normdiff": 0.3422965407371521, | |
| "eval_loss": 7.7101054191589355, | |
| "eval_runtime": 6.8348, | |
| "eval_samples_per_second": 570.314, | |
| "eval_steps_per_second": 4.536, | |
| "step": 750 | |
| }, | |
| { | |
| "epoch": 4.94, | |
| "learning_rate": 4.1767764298093586e-05, | |
| "loss": 6.4981, | |
| "step": 760 | |
| }, | |
| { | |
| "epoch": 5.06, | |
| "learning_rate": 4.1551126516464476e-05, | |
| "loss": 6.2109, | |
| "step": 780 | |
| }, | |
| { | |
| "epoch": 5.19, | |
| "learning_rate": 4.133448873483536e-05, | |
| "loss": 6.5075, | |
| "step": 800 | |
| }, | |
| { | |
| "epoch": 5.19, | |
| "eval_distillation_accuracy_counterfactual": 0.6180092355053873, | |
| "eval_distillation_accuracy_factual": 0.8142637249871729, | |
| "eval_distillation_f1_counterfactual": 0.5811576499624775, | |
| "eval_distillation_f1_factual": 0.8023242774635648, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5443817342226783, | |
| "eval_groundtruth_f1_counterfactual": 0.5018080158011544, | |
| "eval_groundtruth_f1_factual": 0.5428002033480686, | |
| "eval_icace_cosine": 0.45876333117485046, | |
| "eval_icace_l2": 0.505964457988739, | |
| "eval_icace_normdiff": 0.34437206387519836, | |
| "eval_loss": 7.784791469573975, | |
| "eval_runtime": 4.1015, | |
| "eval_samples_per_second": 950.379, | |
| "eval_steps_per_second": 7.558, | |
| "step": 800 | |
| }, | |
| { | |
| "epoch": 5.32, | |
| "learning_rate": 4.111785095320624e-05, | |
| "loss": 6.1528, | |
| "step": 820 | |
| }, | |
| { | |
| "epoch": 5.45, | |
| "learning_rate": 4.0901213171577124e-05, | |
| "loss": 6.2869, | |
| "step": 840 | |
| }, | |
| { | |
| "epoch": 5.52, | |
| "eval_distillation_accuracy_counterfactual": 0.6185223191380195, | |
| "eval_distillation_accuracy_factual": 0.8193945613134941, | |
| "eval_distillation_f1_counterfactual": 0.5836247405075541, | |
| "eval_distillation_f1_factual": 0.8091365156198146, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5515649050795279, | |
| "eval_groundtruth_f1_counterfactual": 0.510664962216415, | |
| "eval_groundtruth_f1_factual": 0.5423106601987745, | |
| "eval_icace_cosine": 0.45199427008628845, | |
| "eval_icace_l2": 0.5019307732582092, | |
| "eval_icace_normdiff": 0.3425298035144806, | |
| "eval_loss": 7.68693733215332, | |
| "eval_runtime": 4.883, | |
| "eval_samples_per_second": 798.276, | |
| "eval_steps_per_second": 6.349, | |
| "step": 850 | |
| }, | |
| { | |
| "epoch": 5.58, | |
| "learning_rate": 4.068457538994801e-05, | |
| "loss": 6.0441, | |
| "step": 860 | |
| }, | |
| { | |
| "epoch": 5.71, | |
| "learning_rate": 4.04679376083189e-05, | |
| "loss": 6.4014, | |
| "step": 880 | |
| }, | |
| { | |
| "epoch": 5.84, | |
| "learning_rate": 4.025129982668977e-05, | |
| "loss": 6.0569, | |
| "step": 900 | |
| }, | |
| { | |
| "epoch": 5.84, | |
| "eval_distillation_accuracy_counterfactual": 0.6226269881990765, | |
| "eval_distillation_accuracy_factual": 0.8114417650076963, | |
| "eval_distillation_f1_counterfactual": 0.5854768643927275, | |
| "eval_distillation_f1_factual": 0.7973005853765658, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5461775269368907, | |
| "eval_groundtruth_f1_counterfactual": 0.502247129912884, | |
| "eval_groundtruth_f1_factual": 0.5450537518673406, | |
| "eval_icace_cosine": 0.44710850715637207, | |
| "eval_icace_l2": 0.4970768690109253, | |
| "eval_icace_normdiff": 0.34042122960090637, | |
| "eval_loss": 7.644977569580078, | |
| "eval_runtime": 4.2109, | |
| "eval_samples_per_second": 925.687, | |
| "eval_steps_per_second": 7.362, | |
| "step": 900 | |
| }, | |
| { | |
| "epoch": 5.97, | |
| "learning_rate": 4.003466204506066e-05, | |
| "loss": 5.9713, | |
| "step": 920 | |
| }, | |
| { | |
| "epoch": 6.1, | |
| "learning_rate": 3.9818024263431546e-05, | |
| "loss": 6.0174, | |
| "step": 940 | |
| }, | |
| { | |
| "epoch": 6.17, | |
| "eval_distillation_accuracy_counterfactual": 0.6180092355053873, | |
| "eval_distillation_accuracy_factual": 0.8065674704976912, | |
| "eval_distillation_f1_counterfactual": 0.5841502017308554, | |
| "eval_distillation_f1_factual": 0.7977275175828281, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5461775269368907, | |
| "eval_groundtruth_f1_counterfactual": 0.506203917901584, | |
| "eval_groundtruth_f1_factual": 0.5465942326831105, | |
| "eval_icace_cosine": 0.45468205213546753, | |
| "eval_icace_l2": 0.5059431791305542, | |
| "eval_icace_normdiff": 0.3447358012199402, | |
| "eval_loss": 7.76542854309082, | |
| "eval_runtime": 4.3013, | |
| "eval_samples_per_second": 906.232, | |
| "eval_steps_per_second": 7.207, | |
| "step": 950 | |
| }, | |
| { | |
| "epoch": 6.23, | |
| "learning_rate": 3.960138648180243e-05, | |
| "loss": 6.0261, | |
| "step": 960 | |
| }, | |
| { | |
| "epoch": 6.36, | |
| "learning_rate": 3.938474870017331e-05, | |
| "loss": 6.0645, | |
| "step": 980 | |
| }, | |
| { | |
| "epoch": 6.49, | |
| "learning_rate": 3.91681109185442e-05, | |
| "loss": 6.0191, | |
| "step": 1000 | |
| }, | |
| { | |
| "epoch": 6.49, | |
| "eval_distillation_accuracy_counterfactual": 0.620318111852232, | |
| "eval_distillation_accuracy_factual": 0.8083632632119035, | |
| "eval_distillation_f1_counterfactual": 0.5830477001223874, | |
| "eval_distillation_f1_factual": 0.7963583724748586, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5531041559774243, | |
| "eval_groundtruth_f1_counterfactual": 0.5079499434851045, | |
| "eval_groundtruth_f1_factual": 0.5433906032493001, | |
| "eval_icace_cosine": 0.44436362385749817, | |
| "eval_icace_l2": 0.49828970432281494, | |
| "eval_icace_normdiff": 0.33891886472702026, | |
| "eval_loss": 7.664845943450928, | |
| "eval_runtime": 4.1469, | |
| "eval_samples_per_second": 939.989, | |
| "eval_steps_per_second": 7.476, | |
| "step": 1000 | |
| }, | |
| { | |
| "epoch": 6.62, | |
| "learning_rate": 3.895147313691508e-05, | |
| "loss": 5.9454, | |
| "step": 1020 | |
| }, | |
| { | |
| "epoch": 6.75, | |
| "learning_rate": 3.873483535528597e-05, | |
| "loss": 6.0488, | |
| "step": 1040 | |
| }, | |
| { | |
| "epoch": 6.82, | |
| "eval_distillation_accuracy_counterfactual": 0.6228835300153925, | |
| "eval_distillation_accuracy_factual": 0.810415597742432, | |
| "eval_distillation_f1_counterfactual": 0.5845336847422479, | |
| "eval_distillation_f1_factual": 0.7976470172715806, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5525910723447922, | |
| "eval_groundtruth_f1_counterfactual": 0.5059005142035983, | |
| "eval_groundtruth_f1_factual": 0.540178117547913, | |
| "eval_icace_cosine": 0.44367682933807373, | |
| "eval_icace_l2": 0.49686795473098755, | |
| "eval_icace_normdiff": 0.3366642892360687, | |
| "eval_loss": 7.668118476867676, | |
| "eval_runtime": 4.3685, | |
| "eval_samples_per_second": 892.297, | |
| "eval_steps_per_second": 7.096, | |
| "step": 1050 | |
| }, | |
| { | |
| "epoch": 6.88, | |
| "learning_rate": 3.851819757365685e-05, | |
| "loss": 5.9853, | |
| "step": 1060 | |
| }, | |
| { | |
| "epoch": 7.01, | |
| "learning_rate": 3.830155979202773e-05, | |
| "loss": 5.7619, | |
| "step": 1080 | |
| }, | |
| { | |
| "epoch": 7.14, | |
| "learning_rate": 3.8084922010398616e-05, | |
| "loss": 5.5946, | |
| "step": 1100 | |
| }, | |
| { | |
| "epoch": 7.14, | |
| "eval_distillation_accuracy_counterfactual": 0.6200615700359159, | |
| "eval_distillation_accuracy_factual": 0.8006670087224218, | |
| "eval_distillation_f1_counterfactual": 0.5850719981039058, | |
| "eval_distillation_f1_factual": 0.7887892378887658, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5500256541816316, | |
| "eval_groundtruth_f1_counterfactual": 0.50798102067776, | |
| "eval_groundtruth_f1_factual": 0.5514350312905094, | |
| "eval_icace_cosine": 0.445857971906662, | |
| "eval_icace_l2": 0.5012825131416321, | |
| "eval_icace_normdiff": 0.34079626202583313, | |
| "eval_loss": 7.738803863525391, | |
| "eval_runtime": 4.1651, | |
| "eval_samples_per_second": 935.867, | |
| "eval_steps_per_second": 7.443, | |
| "step": 1100 | |
| }, | |
| { | |
| "epoch": 7.27, | |
| "learning_rate": 3.78682842287695e-05, | |
| "loss": 5.8864, | |
| "step": 1120 | |
| }, | |
| { | |
| "epoch": 7.4, | |
| "learning_rate": 3.765164644714038e-05, | |
| "loss": 5.8432, | |
| "step": 1140 | |
| }, | |
| { | |
| "epoch": 7.47, | |
| "eval_distillation_accuracy_counterfactual": 0.6213442791174961, | |
| "eval_distillation_accuracy_factual": 0.7993842996408415, | |
| "eval_distillation_f1_counterfactual": 0.5819867386627711, | |
| "eval_distillation_f1_factual": 0.7847982271637772, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5515649050795279, | |
| "eval_groundtruth_f1_counterfactual": 0.5040470994435106, | |
| "eval_groundtruth_f1_factual": 0.5443687788777846, | |
| "eval_icace_cosine": 0.4456865191459656, | |
| "eval_icace_l2": 0.4975723624229431, | |
| "eval_icace_normdiff": 0.33811891078948975, | |
| "eval_loss": 7.686666011810303, | |
| "eval_runtime": 4.2581, | |
| "eval_samples_per_second": 915.424, | |
| "eval_steps_per_second": 7.28, | |
| "step": 1150 | |
| }, | |
| { | |
| "epoch": 7.53, | |
| "learning_rate": 3.7435008665511264e-05, | |
| "loss": 5.9408, | |
| "step": 1160 | |
| }, | |
| { | |
| "epoch": 7.66, | |
| "learning_rate": 3.7218370883882154e-05, | |
| "loss": 5.9549, | |
| "step": 1180 | |
| }, | |
| { | |
| "epoch": 7.79, | |
| "learning_rate": 3.700173310225303e-05, | |
| "loss": 5.5485, | |
| "step": 1200 | |
| }, | |
| { | |
| "epoch": 7.79, | |
| "eval_distillation_accuracy_counterfactual": 0.6187788609543355, | |
| "eval_distillation_accuracy_factual": 0.7981015905592611, | |
| "eval_distillation_f1_counterfactual": 0.5834765365950793, | |
| "eval_distillation_f1_factual": 0.7884169448594356, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5502821959979477, | |
| "eval_groundtruth_f1_counterfactual": 0.5082397669474391, | |
| "eval_groundtruth_f1_factual": 0.5480615734783056, | |
| "eval_icace_cosine": 0.45296600461006165, | |
| "eval_icace_l2": 0.5046243667602539, | |
| "eval_icace_normdiff": 0.34213510155677795, | |
| "eval_loss": 7.746112823486328, | |
| "eval_runtime": 4.1823, | |
| "eval_samples_per_second": 932.018, | |
| "eval_steps_per_second": 7.412, | |
| "step": 1200 | |
| }, | |
| { | |
| "epoch": 7.92, | |
| "learning_rate": 3.678509532062392e-05, | |
| "loss": 5.851, | |
| "step": 1220 | |
| }, | |
| { | |
| "epoch": 8.05, | |
| "learning_rate": 3.65684575389948e-05, | |
| "loss": 5.6669, | |
| "step": 1240 | |
| }, | |
| { | |
| "epoch": 8.12, | |
| "eval_distillation_accuracy_counterfactual": 0.6236531554643406, | |
| "eval_distillation_accuracy_factual": 0.8022062596203181, | |
| "eval_distillation_f1_counterfactual": 0.5840941091107561, | |
| "eval_distillation_f1_factual": 0.7870917202572578, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5556695741405849, | |
| "eval_groundtruth_f1_counterfactual": 0.5089342422895711, | |
| "eval_groundtruth_f1_factual": 0.5380085766933975, | |
| "eval_icace_cosine": 0.4387560784816742, | |
| "eval_icace_l2": 0.49367281794548035, | |
| "eval_icace_normdiff": 0.33387720584869385, | |
| "eval_loss": 7.6664533615112305, | |
| "eval_runtime": 4.3535, | |
| "eval_samples_per_second": 895.377, | |
| "eval_steps_per_second": 7.121, | |
| "step": 1250 | |
| } | |
| ], | |
| "max_steps": 4616, | |
| "num_train_epochs": 30, | |
| "total_flos": 188664237047808.0, | |
| "trial_name": null, | |
| "trial_params": null | |
| } | |