{ "best_metric": 83.6764, "best_model_checkpoint": "models/one_alignment_vs_summary/flan-t5-large/flan_t5_large_alignment_first/checkpoint-6500", "epoch": 0.9742206235011991, "eval_steps": 100, "global_step": 6500, "is_hyper_param_search": false, "is_local_process_zero": true, "is_world_process_zero": true, "log_history": [ { "epoch": 0.01, "learning_rate": 4.992505995203837e-05, "loss": 0.5753, "step": 100 }, { "epoch": 0.01, "eval_F1": 46.8005, "eval_Precision": 65.9011, "eval_Recall": 36.284, "eval_accuracy": 58.536, "eval_accuracy_negative": 80.7879, "eval_accuracy_positive": 36.284, "eval_gen_len": 3.0, "eval_loss": 0.3608146607875824, "eval_runtime": 77.8641, "eval_samples_per_second": 52.81, "eval_steps_per_second": 3.301, "step": 100 }, { "epoch": 0.03, "learning_rate": 4.9850119904076745e-05, "loss": 0.3529, "step": 200 }, { "epoch": 0.03, "eval_F1": 75.56, "eval_Precision": 63.7671, "eval_Recall": 92.7043, "eval_accuracy": 69.9903, "eval_accuracy_negative": 47.2763, "eval_accuracy_positive": 92.7043, "eval_gen_len": 3.0, "eval_loss": 0.3245851397514343, "eval_runtime": 78.1082, "eval_samples_per_second": 52.645, "eval_steps_per_second": 3.29, "step": 200 }, { "epoch": 0.04, "learning_rate": 4.977517985611511e-05, "loss": 0.2974, "step": 300 }, { "epoch": 0.04, "eval_F1": 78.279, "eval_Precision": 68.5944, "eval_Recall": 91.1479, "eval_accuracy": 74.392, "eval_accuracy_negative": 57.6362, "eval_accuracy_positive": 91.1479, "eval_gen_len": 3.0, "eval_loss": 0.41329512000083923, "eval_runtime": 78.0252, "eval_samples_per_second": 52.701, "eval_steps_per_second": 3.294, "step": 300 }, { "epoch": 0.06, "learning_rate": 4.970023980815348e-05, "loss": 0.2854, "step": 400 }, { "epoch": 0.06, "eval_F1": 79.6539, "eval_Precision": 73.2354, "eval_Recall": 87.3054, "eval_accuracy": 77.3346, "eval_accuracy_negative": 67.3638, "eval_accuracy_positive": 87.3054, "eval_gen_len": 3.0, "eval_loss": 0.2850329875946045, "eval_runtime": 77.9552, "eval_samples_per_second": 52.748, "eval_steps_per_second": 3.297, "step": 400 }, { "epoch": 0.07, "learning_rate": 4.9625299760191855e-05, "loss": 0.2569, "step": 500 }, { "epoch": 0.07, "eval_F1": 79.9727, "eval_Precision": 75.225, "eval_Recall": 85.3599, "eval_accuracy": 78.2588, "eval_accuracy_negative": 71.1576, "eval_accuracy_positive": 85.3599, "eval_gen_len": 3.0, "eval_loss": 0.31440889835357666, "eval_runtime": 77.9127, "eval_samples_per_second": 52.777, "eval_steps_per_second": 3.299, "step": 500 }, { "epoch": 0.09, "learning_rate": 4.9550359712230215e-05, "loss": 0.273, "step": 600 }, { "epoch": 0.09, "eval_F1": 80.6036, "eval_Precision": 74.1224, "eval_Recall": 88.3268, "eval_accuracy": 78.6722, "eval_accuracy_negative": 69.0175, "eval_accuracy_positive": 88.3268, "eval_gen_len": 3.0, "eval_loss": 0.2862711548805237, "eval_runtime": 77.9247, "eval_samples_per_second": 52.769, "eval_steps_per_second": 3.298, "step": 600 }, { "epoch": 0.1, "learning_rate": 4.947541966426858e-05, "loss": 0.2769, "step": 700 }, { "epoch": 0.1, "eval_F1": 79.2214, "eval_Precision": 79.26, "eval_Recall": 79.1829, "eval_accuracy": 78.4776, "eval_accuracy_negative": 77.7724, "eval_accuracy_positive": 79.1829, "eval_gen_len": 3.0, "eval_loss": 0.26657262444496155, "eval_runtime": 77.9599, "eval_samples_per_second": 52.745, "eval_steps_per_second": 3.297, "step": 700 }, { "epoch": 0.12, "learning_rate": 4.940047961630696e-05, "loss": 0.288, "step": 800 }, { "epoch": 0.12, "eval_F1": 80.125, "eval_Precision": 79.1924, "eval_Recall": 81.0798, "eval_accuracy": 79.7909, "eval_accuracy_negative": 78.5019, "eval_accuracy_positive": 81.0798, "eval_gen_len": 3.0, "eval_loss": 0.23273229598999023, "eval_runtime": 77.9321, "eval_samples_per_second": 52.764, "eval_steps_per_second": 3.298, "step": 800 }, { "epoch": 0.13, "learning_rate": 4.9325539568345325e-05, "loss": 0.2597, "step": 900 }, { "epoch": 0.13, "eval_F1": 79.5494, "eval_Precision": 80.1184, "eval_Recall": 78.9883, "eval_accuracy": 78.7451, "eval_accuracy_negative": 78.5019, "eval_accuracy_positive": 78.9883, "eval_gen_len": 3.0, "eval_loss": 0.22254109382629395, "eval_runtime": 77.9631, "eval_samples_per_second": 52.743, "eval_steps_per_second": 3.296, "step": 900 }, { "epoch": 0.15, "learning_rate": 4.925059952038369e-05, "loss": 0.2735, "step": 1000 }, { "epoch": 0.15, "eval_F1": 81.095, "eval_Precision": 78.1335, "eval_Recall": 84.2899, "eval_accuracy": 79.9368, "eval_accuracy_negative": 75.5837, "eval_accuracy_positive": 84.2899, "eval_gen_len": 3.0, "eval_loss": 0.23284350335597992, "eval_runtime": 77.9328, "eval_samples_per_second": 52.763, "eval_steps_per_second": 3.298, "step": 1000 }, { "epoch": 0.16, "learning_rate": 4.917565947242207e-05, "loss": 0.2591, "step": 1100 }, { "epoch": 0.16, "eval_F1": 81.521, "eval_Precision": 77.891, "eval_Recall": 85.5058, "eval_accuracy": 80.5934, "eval_accuracy_negative": 75.6809, "eval_accuracy_positive": 85.5058, "eval_gen_len": 3.0, "eval_loss": 0.26482418179512024, "eval_runtime": 78.0339, "eval_samples_per_second": 52.695, "eval_steps_per_second": 3.293, "step": 1100 }, { "epoch": 0.18, "learning_rate": 4.9100719424460435e-05, "loss": 0.2675, "step": 1200 }, { "epoch": 0.18, "eval_F1": 80.1508, "eval_Precision": 82.9002, "eval_Recall": 77.5778, "eval_accuracy": 80.7879, "eval_accuracy_negative": 83.9981, "eval_accuracy_positive": 77.5778, "eval_gen_len": 3.0, "eval_loss": 0.23910629749298096, "eval_runtime": 78.0093, "eval_samples_per_second": 52.712, "eval_steps_per_second": 3.294, "step": 1200 }, { "epoch": 0.19, "learning_rate": 4.90257793764988e-05, "loss": 0.2549, "step": 1300 }, { "epoch": 0.19, "eval_F1": 81.6754, "eval_Precision": 79.9627, "eval_Recall": 83.463, "eval_accuracy": 81.1041, "eval_accuracy_negative": 78.7451, "eval_accuracy_positive": 83.463, "eval_gen_len": 3.0, "eval_loss": 0.22716127336025238, "eval_runtime": 78.0476, "eval_samples_per_second": 52.686, "eval_steps_per_second": 3.293, "step": 1300 }, { "epoch": 0.21, "learning_rate": 4.895083932853717e-05, "loss": 0.2338, "step": 1400 }, { "epoch": 0.21, "eval_F1": 80.1995, "eval_Precision": 82.2927, "eval_Recall": 78.2101, "eval_accuracy": 80.4475, "eval_accuracy_negative": 82.6848, "eval_accuracy_positive": 78.2101, "eval_gen_len": 3.0, "eval_loss": 0.24518640339374542, "eval_runtime": 77.9981, "eval_samples_per_second": 52.719, "eval_steps_per_second": 3.295, "step": 1400 }, { "epoch": 0.22, "learning_rate": 4.8875899280575545e-05, "loss": 0.2747, "step": 1500 }, { "epoch": 0.22, "eval_F1": 81.9309, "eval_Precision": 75.2033, "eval_Recall": 89.9805, "eval_accuracy": 79.4504, "eval_accuracy_negative": 68.9202, "eval_accuracy_positive": 89.9805, "eval_gen_len": 3.0, "eval_loss": 0.2351878583431244, "eval_runtime": 77.8581, "eval_samples_per_second": 52.814, "eval_steps_per_second": 3.301, "step": 1500 }, { "epoch": 0.24, "learning_rate": 4.880095923261391e-05, "loss": 0.2389, "step": 1600 }, { "epoch": 0.24, "eval_F1": 77.4047, "eval_Precision": 80.4724, "eval_Recall": 74.5623, "eval_accuracy": 76.143, "eval_accuracy_negative": 77.7237, "eval_accuracy_positive": 74.5623, "eval_gen_len": 3.0, "eval_loss": 0.240137979388237, "eval_runtime": 77.9704, "eval_samples_per_second": 52.738, "eval_steps_per_second": 3.296, "step": 1600 }, { "epoch": 0.25, "learning_rate": 4.872601918465228e-05, "loss": 0.2486, "step": 1700 }, { "epoch": 0.25, "eval_F1": 74.3833, "eval_Precision": 84.0171, "eval_Recall": 66.7315, "eval_accuracy": 73.5165, "eval_accuracy_negative": 80.3016, "eval_accuracy_positive": 66.7315, "eval_gen_len": 3.0, "eval_loss": 0.23113004863262177, "eval_runtime": 77.9471, "eval_samples_per_second": 52.754, "eval_steps_per_second": 3.297, "step": 1700 }, { "epoch": 0.27, "learning_rate": 4.865107913669065e-05, "loss": 0.2558, "step": 1800 }, { "epoch": 0.27, "eval_F1": 75.5311, "eval_Precision": 81.9579, "eval_Recall": 70.0389, "eval_accuracy": 71.644, "eval_accuracy_negative": 73.249, "eval_accuracy_positive": 70.0389, "eval_gen_len": 3.0, "eval_loss": 0.2416938990354538, "eval_runtime": 77.8981, "eval_samples_per_second": 52.787, "eval_steps_per_second": 3.299, "step": 1800 }, { "epoch": 0.28, "learning_rate": 4.8576139088729016e-05, "loss": 0.2521, "step": 1900 }, { "epoch": 0.28, "eval_F1": 75.8988, "eval_Precision": 83.8729, "eval_Recall": 69.3093, "eval_accuracy": 72.0574, "eval_accuracy_negative": 74.8054, "eval_accuracy_positive": 69.3093, "eval_gen_len": 3.0, "eval_loss": 0.2378227710723877, "eval_runtime": 78.1302, "eval_samples_per_second": 52.63, "eval_steps_per_second": 3.289, "step": 1900 }, { "epoch": 0.3, "learning_rate": 4.8501199040767384e-05, "loss": 0.2487, "step": 2000 }, { "epoch": 0.3, "eval_F1": 82.5998, "eval_Precision": 77.7587, "eval_Recall": 88.0837, "eval_accuracy": 72.5438, "eval_accuracy_negative": 57.0039, "eval_accuracy_positive": 88.0837, "eval_gen_len": 3.0, "eval_loss": 0.31695932149887085, "eval_runtime": 78.0522, "eval_samples_per_second": 52.683, "eval_steps_per_second": 3.293, "step": 2000 }, { "epoch": 0.31, "learning_rate": 4.842625899280576e-05, "loss": 0.246, "step": 2100 }, { "epoch": 0.31, "eval_F1": 81.1168, "eval_Precision": 81.6971, "eval_Recall": 80.5447, "eval_accuracy": 66.3911, "eval_accuracy_negative": 52.2374, "eval_accuracy_positive": 80.5447, "eval_gen_len": 3.0, "eval_loss": 0.25484344363212585, "eval_runtime": 78.0548, "eval_samples_per_second": 52.681, "eval_steps_per_second": 3.293, "step": 2100 }, { "epoch": 0.33, "learning_rate": 4.8351318944844126e-05, "loss": 0.2407, "step": 2200 }, { "epoch": 0.33, "eval_F1": 74.0905, "eval_Precision": 86.343, "eval_Recall": 64.8833, "eval_accuracy": 71.7656, "eval_accuracy_negative": 78.6479, "eval_accuracy_positive": 64.8833, "eval_gen_len": 3.0, "eval_loss": 0.26473525166511536, "eval_runtime": 77.9999, "eval_samples_per_second": 52.718, "eval_steps_per_second": 3.295, "step": 2200 }, { "epoch": 0.34, "learning_rate": 4.8276378896882494e-05, "loss": 0.2293, "step": 2300 }, { "epoch": 0.34, "eval_F1": 76.173, "eval_Precision": 82.6038, "eval_Recall": 70.6712, "eval_accuracy": 72.7383, "eval_accuracy_negative": 74.8054, "eval_accuracy_positive": 70.6712, "eval_gen_len": 3.0, "eval_loss": 0.23481938242912292, "eval_runtime": 78.0051, "eval_samples_per_second": 52.714, "eval_steps_per_second": 3.295, "step": 2300 }, { "epoch": 0.36, "learning_rate": 4.820143884892087e-05, "loss": 0.199, "step": 2400 }, { "epoch": 0.36, "eval_F1": 81.2854, "eval_Precision": 79.0441, "eval_Recall": 83.6576, "eval_accuracy": 77.8453, "eval_accuracy_negative": 72.0331, "eval_accuracy_positive": 83.6576, "eval_gen_len": 3.0, "eval_loss": 0.29514577984809875, "eval_runtime": 77.9919, "eval_samples_per_second": 52.723, "eval_steps_per_second": 3.295, "step": 2400 }, { "epoch": 0.37, "learning_rate": 4.8126498800959236e-05, "loss": 0.247, "step": 2500 }, { "epoch": 0.37, "eval_F1": 67.5548, "eval_Precision": 86.3086, "eval_Recall": 55.4961, "eval_accuracy": 70.3551, "eval_accuracy_negative": 85.214, "eval_accuracy_positive": 55.4961, "eval_gen_len": 3.0, "eval_loss": 0.23327696323394775, "eval_runtime": 78.0614, "eval_samples_per_second": 52.677, "eval_steps_per_second": 3.292, "step": 2500 }, { "epoch": 0.39, "learning_rate": 4.8051558752997604e-05, "loss": 0.2659, "step": 2600 }, { "epoch": 0.39, "eval_F1": 77.5335, "eval_Precision": 80.6944, "eval_Recall": 74.6109, "eval_accuracy": 68.6284, "eval_accuracy_negative": 62.6459, "eval_accuracy_positive": 74.6109, "eval_gen_len": 3.0, "eval_loss": 0.26498734951019287, "eval_runtime": 77.9945, "eval_samples_per_second": 52.722, "eval_steps_per_second": 3.295, "step": 2600 }, { "epoch": 0.4, "learning_rate": 4.797661870503598e-05, "loss": 0.2445, "step": 2700 }, { "epoch": 0.4, "eval_F1": 81.6411, "eval_Precision": 77.1998, "eval_Recall": 86.6245, "eval_accuracy": 78.0156, "eval_accuracy_negative": 69.4066, "eval_accuracy_positive": 86.6245, "eval_gen_len": 3.0, "eval_loss": 0.25022661685943604, "eval_runtime": 78.0413, "eval_samples_per_second": 52.69, "eval_steps_per_second": 3.293, "step": 2700 }, { "epoch": 0.42, "learning_rate": 4.7901678657074346e-05, "loss": 0.2397, "step": 2800 }, { "epoch": 0.42, "eval_F1": 81.1614, "eval_Precision": 77.679, "eval_Recall": 84.9708, "eval_accuracy": 75.9241, "eval_accuracy_negative": 66.8774, "eval_accuracy_positive": 84.9708, "eval_gen_len": 3.0, "eval_loss": 0.2765290439128876, "eval_runtime": 77.9454, "eval_samples_per_second": 52.755, "eval_steps_per_second": 3.297, "step": 2800 }, { "epoch": 0.43, "learning_rate": 4.7826738609112713e-05, "loss": 0.2134, "step": 2900 }, { "epoch": 0.43, "eval_F1": 81.6516, "eval_Precision": 78.0488, "eval_Recall": 85.6031, "eval_accuracy": 77.9426, "eval_accuracy_negative": 70.2821, "eval_accuracy_positive": 85.6031, "eval_gen_len": 3.0, "eval_loss": 0.2863008975982666, "eval_runtime": 78.0388, "eval_samples_per_second": 52.692, "eval_steps_per_second": 3.293, "step": 2900 }, { "epoch": 0.45, "learning_rate": 4.775179856115108e-05, "loss": 0.2349, "step": 3000 }, { "epoch": 0.45, "eval_F1": 81.0536, "eval_Precision": 80.5476, "eval_Recall": 81.5661, "eval_accuracy": 77.8696, "eval_accuracy_negative": 74.1732, "eval_accuracy_positive": 81.5661, "eval_gen_len": 3.0, "eval_loss": 0.23327529430389404, "eval_runtime": 77.998, "eval_samples_per_second": 52.719, "eval_steps_per_second": 3.295, "step": 3000 }, { "epoch": 0.46, "learning_rate": 4.767685851318945e-05, "loss": 0.2031, "step": 3100 }, { "epoch": 0.46, "eval_F1": 80.2638, "eval_Precision": 83.8812, "eval_Recall": 76.9455, "eval_accuracy": 78.429, "eval_accuracy_negative": 79.9125, "eval_accuracy_positive": 76.9455, "eval_gen_len": 3.0, "eval_loss": 0.2780900001525879, "eval_runtime": 78.0382, "eval_samples_per_second": 52.692, "eval_steps_per_second": 3.293, "step": 3100 }, { "epoch": 0.48, "learning_rate": 4.7601918465227817e-05, "loss": 0.2704, "step": 3200 }, { "epoch": 0.48, "eval_F1": 80.8479, "eval_Precision": 82.958, "eval_Recall": 78.8424, "eval_accuracy": 79.8881, "eval_accuracy_negative": 80.9339, "eval_accuracy_positive": 78.8424, "eval_gen_len": 3.0, "eval_loss": 0.22213193774223328, "eval_runtime": 77.9705, "eval_samples_per_second": 52.738, "eval_steps_per_second": 3.296, "step": 3200 }, { "epoch": 0.49, "learning_rate": 4.752697841726619e-05, "loss": 0.2413, "step": 3300 }, { "epoch": 0.49, "eval_F1": 82.2274, "eval_Precision": 78.6158, "eval_Recall": 86.1868, "eval_accuracy": 79.3045, "eval_accuracy_negative": 72.4222, "eval_accuracy_positive": 86.1868, "eval_gen_len": 3.0, "eval_loss": 0.21565163135528564, "eval_runtime": 77.9735, "eval_samples_per_second": 52.736, "eval_steps_per_second": 3.296, "step": 3300 }, { "epoch": 0.51, "learning_rate": 4.745203836930456e-05, "loss": 0.2549, "step": 3400 }, { "epoch": 0.51, "eval_F1": 80.991, "eval_Precision": 70.8045, "eval_Recall": 94.6012, "eval_accuracy": 75.5107, "eval_accuracy_negative": 56.4202, "eval_accuracy_positive": 94.6012, "eval_gen_len": 3.0, "eval_loss": 0.3653569221496582, "eval_runtime": 77.9428, "eval_samples_per_second": 52.757, "eval_steps_per_second": 3.297, "step": 3400 }, { "epoch": 0.52, "learning_rate": 4.7377098321342926e-05, "loss": 0.2327, "step": 3500 }, { "epoch": 0.52, "eval_F1": 80.6348, "eval_Precision": 82.2458, "eval_Recall": 79.0856, "eval_accuracy": 78.7938, "eval_accuracy_negative": 78.5019, "eval_accuracy_positive": 79.0856, "eval_gen_len": 3.0, "eval_loss": 0.22318707406520844, "eval_runtime": 77.9908, "eval_samples_per_second": 52.724, "eval_steps_per_second": 3.295, "step": 3500 }, { "epoch": 0.54, "learning_rate": 4.7302158273381294e-05, "loss": 0.234, "step": 3600 }, { "epoch": 0.54, "eval_F1": 82.8591, "eval_Precision": 76.8176, "eval_Recall": 89.9319, "eval_accuracy": 77.5535, "eval_accuracy_negative": 65.1751, "eval_accuracy_positive": 89.9319, "eval_gen_len": 3.0, "eval_loss": 0.28604593873023987, "eval_runtime": 77.9808, "eval_samples_per_second": 52.731, "eval_steps_per_second": 3.296, "step": 3600 }, { "epoch": 0.55, "learning_rate": 4.722721822541967e-05, "loss": 0.258, "step": 3700 }, { "epoch": 0.55, "eval_F1": 74.4311, "eval_Precision": 82.7874, "eval_Recall": 67.607, "eval_accuracy": 67.9718, "eval_accuracy_negative": 68.3366, "eval_accuracy_positive": 67.607, "eval_gen_len": 3.0, "eval_loss": 0.2335677146911621, "eval_runtime": 78.0681, "eval_samples_per_second": 52.672, "eval_steps_per_second": 3.292, "step": 3700 }, { "epoch": 0.57, "learning_rate": 4.7152278177458036e-05, "loss": 0.2367, "step": 3800 }, { "epoch": 0.57, "eval_F1": 69.2561, "eval_Precision": 89.7754, "eval_Recall": 56.3716, "eval_accuracy": 67.4368, "eval_accuracy_negative": 78.5019, "eval_accuracy_positive": 56.3716, "eval_gen_len": 3.0, "eval_loss": 0.2100234031677246, "eval_runtime": 77.9683, "eval_samples_per_second": 52.739, "eval_steps_per_second": 3.296, "step": 3800 }, { "epoch": 0.58, "learning_rate": 4.7077338129496404e-05, "loss": 0.2157, "step": 3900 }, { "epoch": 0.58, "eval_F1": 79.2374, "eval_Precision": 85.5612, "eval_Recall": 73.784, "eval_accuracy": 76.8482, "eval_accuracy_negative": 79.9125, "eval_accuracy_positive": 73.784, "eval_gen_len": 3.0, "eval_loss": 0.23944681882858276, "eval_runtime": 77.9314, "eval_samples_per_second": 52.764, "eval_steps_per_second": 3.298, "step": 3900 }, { "epoch": 0.6, "learning_rate": 4.700239808153478e-05, "loss": 0.2431, "step": 4000 }, { "epoch": 0.6, "eval_F1": 72.7115, "eval_Precision": 89.4812, "eval_Recall": 61.2354, "eval_accuracy": 74.0272, "eval_accuracy_negative": 86.8191, "eval_accuracy_positive": 61.2354, "eval_gen_len": 3.0, "eval_loss": 0.2376098483800888, "eval_runtime": 78.0857, "eval_samples_per_second": 52.66, "eval_steps_per_second": 3.291, "step": 4000 }, { "epoch": 0.61, "learning_rate": 4.6927458033573146e-05, "loss": 0.2442, "step": 4100 }, { "epoch": 0.61, "eval_F1": 82.3446, "eval_Precision": 79.7901, "eval_Recall": 85.0681, "eval_accuracy": 78.4047, "eval_accuracy_negative": 71.7412, "eval_accuracy_positive": 85.0681, "eval_gen_len": 3.0, "eval_loss": 0.2564048767089844, "eval_runtime": 78.0736, "eval_samples_per_second": 52.668, "eval_steps_per_second": 3.292, "step": 4100 }, { "epoch": 0.63, "learning_rate": 4.685251798561151e-05, "loss": 0.2125, "step": 4200 }, { "epoch": 0.63, "eval_F1": 81.7357, "eval_Precision": 82.4024, "eval_Recall": 81.0798, "eval_accuracy": 77.4562, "eval_accuracy_negative": 73.8327, "eval_accuracy_positive": 81.0798, "eval_gen_len": 3.0, "eval_loss": 0.26422980427742004, "eval_runtime": 77.9729, "eval_samples_per_second": 52.736, "eval_steps_per_second": 3.296, "step": 4200 }, { "epoch": 0.64, "learning_rate": 4.677757793764988e-05, "loss": 0.2213, "step": 4300 }, { "epoch": 0.64, "eval_F1": 61.6757, "eval_Precision": 89.3813, "eval_Recall": 47.0817, "eval_accuracy": 66.5856, "eval_accuracy_negative": 86.0895, "eval_accuracy_positive": 47.0817, "eval_gen_len": 3.0, "eval_loss": 0.2372213900089264, "eval_runtime": 78.0051, "eval_samples_per_second": 52.714, "eval_steps_per_second": 3.295, "step": 4300 }, { "epoch": 0.66, "learning_rate": 4.670263788968825e-05, "loss": 0.2277, "step": 4400 }, { "epoch": 0.66, "eval_F1": 65.461, "eval_Precision": 83.8269, "eval_Recall": 53.6965, "eval_accuracy": 63.2539, "eval_accuracy_negative": 72.8113, "eval_accuracy_positive": 53.6965, "eval_gen_len": 3.0, "eval_loss": 0.21777845919132233, "eval_runtime": 78.04, "eval_samples_per_second": 52.691, "eval_steps_per_second": 3.293, "step": 4400 }, { "epoch": 0.67, "learning_rate": 4.662769784172662e-05, "loss": 0.1893, "step": 4500 }, { "epoch": 0.67, "eval_F1": 61.2192, "eval_Precision": 91.8288, "eval_Recall": 45.9144, "eval_accuracy": 62.1839, "eval_accuracy_negative": 78.4533, "eval_accuracy_positive": 45.9144, "eval_gen_len": 3.0, "eval_loss": 0.25438836216926575, "eval_runtime": 77.9949, "eval_samples_per_second": 52.721, "eval_steps_per_second": 3.295, "step": 4500 }, { "epoch": 0.69, "learning_rate": 4.655275779376499e-05, "loss": 0.2577, "step": 4600 }, { "epoch": 0.69, "eval_F1": 48.5335, "eval_Precision": 86.0149, "eval_Recall": 33.8035, "eval_accuracy": 56.9309, "eval_accuracy_negative": 80.0584, "eval_accuracy_positive": 33.8035, "eval_gen_len": 3.0, "eval_loss": 0.20975889265537262, "eval_runtime": 77.9509, "eval_samples_per_second": 52.751, "eval_steps_per_second": 3.297, "step": 4600 }, { "epoch": 0.7, "learning_rate": 4.647781774580336e-05, "loss": 0.2255, "step": 4700 }, { "epoch": 0.7, "eval_F1": 48.856, "eval_Precision": 88.4076, "eval_Recall": 33.7549, "eval_accuracy": 59.3872, "eval_accuracy_negative": 85.0195, "eval_accuracy_positive": 33.7549, "eval_gen_len": 3.0, "eval_loss": 0.2087894082069397, "eval_runtime": 78.0665, "eval_samples_per_second": 52.673, "eval_steps_per_second": 3.292, "step": 4700 }, { "epoch": 0.72, "learning_rate": 4.640287769784173e-05, "loss": 0.2052, "step": 4800 }, { "epoch": 0.72, "eval_F1": 62.0057, "eval_Precision": 88.6878, "eval_Recall": 47.6654, "eval_accuracy": 64.6887, "eval_accuracy_negative": 81.7121, "eval_accuracy_positive": 47.6654, "eval_gen_len": 3.0, "eval_loss": 0.2096925675868988, "eval_runtime": 78.1278, "eval_samples_per_second": 52.632, "eval_steps_per_second": 3.289, "step": 4800 }, { "epoch": 0.73, "learning_rate": 4.63279376498801e-05, "loss": 0.1803, "step": 4900 }, { "epoch": 0.73, "eval_F1": 76.7251, "eval_Precision": 87.782, "eval_Recall": 68.142, "eval_accuracy": 75.8025, "eval_accuracy_negative": 83.463, "eval_accuracy_positive": 68.142, "eval_gen_len": 3.0, "eval_loss": 0.23566679656505585, "eval_runtime": 78.1099, "eval_samples_per_second": 52.644, "eval_steps_per_second": 3.29, "step": 4900 }, { "epoch": 0.75, "learning_rate": 4.625299760191847e-05, "loss": 0.2405, "step": 5000 }, { "epoch": 0.75, "eval_F1": 79.0246, "eval_Precision": 85.7224, "eval_Recall": 73.2977, "eval_accuracy": 75.8998, "eval_accuracy_negative": 78.5019, "eval_accuracy_positive": 73.2977, "eval_gen_len": 3.0, "eval_loss": 0.22571609914302826, "eval_runtime": 77.9758, "eval_samples_per_second": 52.734, "eval_steps_per_second": 3.296, "step": 5000 }, { "epoch": 0.76, "learning_rate": 4.617805755395684e-05, "loss": 0.195, "step": 5100 }, { "epoch": 0.76, "eval_F1": 38.7195, "eval_Precision": 89.4366, "eval_Recall": 24.7082, "eval_accuracy": 55.0097, "eval_accuracy_negative": 85.3113, "eval_accuracy_positive": 24.7082, "eval_gen_len": 3.0, "eval_loss": 0.25449925661087036, "eval_runtime": 78.1451, "eval_samples_per_second": 52.62, "eval_steps_per_second": 3.289, "step": 5100 }, { "epoch": 0.78, "learning_rate": 4.610311750599521e-05, "loss": 0.2136, "step": 5200 }, { "epoch": 0.78, "eval_F1": 72.9004, "eval_Precision": 89.638, "eval_Recall": 61.43, "eval_accuracy": 70.1605, "eval_accuracy_negative": 78.8911, "eval_accuracy_positive": 61.43, "eval_gen_len": 3.0, "eval_loss": 0.23430722951889038, "eval_runtime": 78.0577, "eval_samples_per_second": 52.679, "eval_steps_per_second": 3.292, "step": 5200 }, { "epoch": 0.79, "learning_rate": 4.602817745803358e-05, "loss": 0.205, "step": 5300 }, { "epoch": 0.79, "eval_F1": 61.5385, "eval_Precision": 88.8073, "eval_Recall": 47.0817, "eval_accuracy": 63.6916, "eval_accuracy_negative": 80.3016, "eval_accuracy_positive": 47.0817, "eval_gen_len": 3.0, "eval_loss": 0.20516321063041687, "eval_runtime": 78.06, "eval_samples_per_second": 52.677, "eval_steps_per_second": 3.292, "step": 5300 }, { "epoch": 0.81, "learning_rate": 4.595323741007194e-05, "loss": 0.227, "step": 5400 }, { "epoch": 0.81, "eval_F1": 71.5575, "eval_Precision": 89.1226, "eval_Recall": 59.7763, "eval_accuracy": 70.1119, "eval_accuracy_negative": 80.4475, "eval_accuracy_positive": 59.7763, "eval_gen_len": 3.0, "eval_loss": 0.20273716747760773, "eval_runtime": 78.0817, "eval_samples_per_second": 52.663, "eval_steps_per_second": 3.291, "step": 5400 }, { "epoch": 0.82, "learning_rate": 4.5878297362110315e-05, "loss": 0.221, "step": 5500 }, { "epoch": 0.82, "eval_F1": 70.1068, "eval_Precision": 89.8176, "eval_Recall": 57.4903, "eval_accuracy": 68.3123, "eval_accuracy_negative": 79.1342, "eval_accuracy_positive": 57.4903, "eval_gen_len": 3.0, "eval_loss": 0.20614123344421387, "eval_runtime": 78.0595, "eval_samples_per_second": 52.678, "eval_steps_per_second": 3.292, "step": 5500 }, { "epoch": 0.84, "learning_rate": 4.580335731414868e-05, "loss": 0.2148, "step": 5600 }, { "epoch": 0.84, "eval_F1": 78.0293, "eval_Precision": 86.2272, "eval_Recall": 71.2549, "eval_accuracy": 71.6683, "eval_accuracy_negative": 72.0817, "eval_accuracy_positive": 71.2549, "eval_gen_len": 3.0, "eval_loss": 0.23381440341472626, "eval_runtime": 78.0578, "eval_samples_per_second": 52.679, "eval_steps_per_second": 3.292, "step": 5600 }, { "epoch": 0.85, "learning_rate": 4.572841726618705e-05, "loss": 0.203, "step": 5700 }, { "epoch": 0.85, "eval_F1": 58.205, "eval_Precision": 93.4409, "eval_Recall": 42.2665, "eval_accuracy": 60.4572, "eval_accuracy_negative": 78.6479, "eval_accuracy_positive": 42.2665, "eval_gen_len": 3.0, "eval_loss": 0.2337663322687149, "eval_runtime": 78.2027, "eval_samples_per_second": 52.581, "eval_steps_per_second": 3.286, "step": 5700 }, { "epoch": 0.87, "learning_rate": 4.5653477218225424e-05, "loss": 0.1912, "step": 5800 }, { "epoch": 0.87, "eval_F1": 82.295, "eval_Precision": 83.3416, "eval_Recall": 81.2743, "eval_accuracy": 78.6722, "eval_accuracy_negative": 76.07, "eval_accuracy_positive": 81.2743, "eval_gen_len": 3.0, "eval_loss": 0.23967531323432922, "eval_runtime": 78.2026, "eval_samples_per_second": 52.581, "eval_steps_per_second": 3.286, "step": 5800 }, { "epoch": 0.88, "learning_rate": 4.557853717026379e-05, "loss": 0.2292, "step": 5900 }, { "epoch": 0.88, "eval_F1": 77.3544, "eval_Precision": 87.8245, "eval_Recall": 69.1148, "eval_accuracy": 71.8142, "eval_accuracy_negative": 74.5136, "eval_accuracy_positive": 69.1148, "eval_gen_len": 3.0, "eval_loss": 0.21783529222011566, "eval_runtime": 78.0505, "eval_samples_per_second": 52.684, "eval_steps_per_second": 3.293, "step": 5900 }, { "epoch": 0.9, "learning_rate": 4.550359712230216e-05, "loss": 0.2366, "step": 6000 }, { "epoch": 0.9, "eval_F1": 73.1382, "eval_Precision": 90.4659, "eval_Recall": 61.3813, "eval_accuracy": 69.3337, "eval_accuracy_negative": 77.286, "eval_accuracy_positive": 61.3813, "eval_gen_len": 3.0, "eval_loss": 0.21324363350868225, "eval_runtime": 78.0489, "eval_samples_per_second": 52.685, "eval_steps_per_second": 3.293, "step": 6000 }, { "epoch": 0.91, "learning_rate": 4.542865707434053e-05, "loss": 0.2172, "step": 6100 }, { "epoch": 0.91, "eval_F1": 82.6628, "eval_Precision": 81.4151, "eval_Recall": 83.9494, "eval_accuracy": 80.0584, "eval_accuracy_negative": 76.1673, "eval_accuracy_positive": 83.9494, "eval_gen_len": 3.0, "eval_loss": 0.21681031584739685, "eval_runtime": 77.9889, "eval_samples_per_second": 52.725, "eval_steps_per_second": 3.295, "step": 6100 }, { "epoch": 0.93, "learning_rate": 4.53537170263789e-05, "loss": 0.2176, "step": 6200 }, { "epoch": 0.93, "eval_F1": 81.7284, "eval_Precision": 82.999, "eval_Recall": 80.4961, "eval_accuracy": 79.2802, "eval_accuracy_negative": 78.0642, "eval_accuracy_positive": 80.4961, "eval_gen_len": 3.0, "eval_loss": 0.19668923318386078, "eval_runtime": 78.1339, "eval_samples_per_second": 52.628, "eval_steps_per_second": 3.289, "step": 6200 }, { "epoch": 0.94, "learning_rate": 4.527877697841727e-05, "loss": 0.1703, "step": 6300 }, { "epoch": 0.94, "eval_F1": 82.7941, "eval_Precision": 83.4486, "eval_Recall": 82.1498, "eval_accuracy": 79.9125, "eval_accuracy_negative": 77.6751, "eval_accuracy_positive": 82.1498, "eval_gen_len": 3.0, "eval_loss": 0.23434630036354065, "eval_runtime": 77.9722, "eval_samples_per_second": 52.737, "eval_steps_per_second": 3.296, "step": 6300 }, { "epoch": 0.96, "learning_rate": 4.520383693045564e-05, "loss": 0.2234, "step": 6400 }, { "epoch": 0.96, "eval_F1": 83.6217, "eval_Precision": 81.5534, "eval_Recall": 85.7977, "eval_accuracy": 80.2529, "eval_accuracy_negative": 74.7082, "eval_accuracy_positive": 85.7977, "eval_gen_len": 3.0, "eval_loss": 0.22015611827373505, "eval_runtime": 78.0553, "eval_samples_per_second": 52.681, "eval_steps_per_second": 3.293, "step": 6400 }, { "epoch": 0.97, "learning_rate": 4.512889688249401e-05, "loss": 0.2047, "step": 6500 }, { "epoch": 0.97, "eval_F1": 83.6764, "eval_Precision": 82.3751, "eval_Recall": 85.0195, "eval_accuracy": 79.1586, "eval_accuracy_negative": 73.2977, "eval_accuracy_positive": 85.0195, "eval_gen_len": 3.0, "eval_loss": 0.20958371460437775, "eval_runtime": 77.9877, "eval_samples_per_second": 52.726, "eval_steps_per_second": 3.295, "step": 6500 } ], "logging_steps": 100, "max_steps": 66720, "num_train_epochs": 10, "save_steps": 100, "total_flos": 1.83775713030144e+16, "trial_name": null, "trial_params": null }