| { | |
| "best_metric": null, | |
| "best_model_checkpoint": null, | |
| "epoch": 46.99733333333333, | |
| "eval_steps": 500, | |
| "global_step": 8812, | |
| "is_hyper_param_search": false, | |
| "is_local_process_zero": true, | |
| "is_world_process_zero": true, | |
| "log_history": [ | |
| { | |
| "epoch": 0.53, | |
| "learning_rate": 5e-05, | |
| "loss": 1.3988, | |
| "step": 100 | |
| }, | |
| { | |
| "epoch": 1.0, | |
| "eval_accuracy": 0.6660784313725491, | |
| "eval_loss": 1.2107256650924683, | |
| "eval_runtime": 7.728, | |
| "eval_samples_per_second": 64.7, | |
| "eval_steps_per_second": 8.152, | |
| "step": 187 | |
| }, | |
| { | |
| "epoch": 1.0, | |
| "eval_exact_match": 19.4, | |
| "eval_f1": 27.911959489312448, | |
| "eval_qa_bleu": 2.842200543476061, | |
| "eval_qa_exact_match": 0.152, | |
| "eval_recite_bleu": 12.149078895446083, | |
| "eval_recite_exact_match": 0.0, | |
| "step": 187 | |
| }, | |
| { | |
| "epoch": 1.07, | |
| "learning_rate": 5e-05, | |
| "loss": 1.238, | |
| "step": 200 | |
| }, | |
| { | |
| "epoch": 1.6, | |
| "learning_rate": 5e-05, | |
| "loss": 1.1977, | |
| "step": 300 | |
| }, | |
| { | |
| "epoch": 2.0, | |
| "eval_accuracy": 0.6676078431372549, | |
| "eval_loss": 1.1987448930740356, | |
| "eval_runtime": 7.4628, | |
| "eval_samples_per_second": 66.999, | |
| "eval_steps_per_second": 8.442, | |
| "step": 375 | |
| }, | |
| { | |
| "epoch": 2.0, | |
| "eval_exact_match": 21.4, | |
| "eval_f1": 27.979969946667268, | |
| "eval_qa_bleu": 3.4273838915653356, | |
| "eval_qa_exact_match": 0.158, | |
| "eval_recite_bleu": 11.474877690710539, | |
| "eval_recite_exact_match": 0.0, | |
| "step": 375 | |
| }, | |
| { | |
| "epoch": 2.13, | |
| "learning_rate": 5e-05, | |
| "loss": 1.1628, | |
| "step": 400 | |
| }, | |
| { | |
| "epoch": 2.67, | |
| "learning_rate": 5e-05, | |
| "loss": 1.1458, | |
| "step": 500 | |
| }, | |
| { | |
| "epoch": 3.0, | |
| "eval_accuracy": 0.6675764705882353, | |
| "eval_loss": 1.1964411735534668, | |
| "eval_runtime": 6.6677, | |
| "eval_samples_per_second": 74.988, | |
| "eval_steps_per_second": 9.448, | |
| "step": 562 | |
| }, | |
| { | |
| "epoch": 3.0, | |
| "eval_exact_match": 23.4, | |
| "eval_f1": 31.525761083856903, | |
| "eval_qa_bleu": 2.0606924929592334, | |
| "eval_qa_exact_match": 0.17, | |
| "eval_recite_bleu": 12.027903845545522, | |
| "eval_recite_exact_match": 0.0, | |
| "step": 562 | |
| }, | |
| { | |
| "epoch": 3.2, | |
| "learning_rate": 5e-05, | |
| "loss": 1.1179, | |
| "step": 600 | |
| }, | |
| { | |
| "epoch": 3.73, | |
| "learning_rate": 5e-05, | |
| "loss": 1.0707, | |
| "step": 700 | |
| }, | |
| { | |
| "epoch": 4.0, | |
| "eval_accuracy": 0.6665960784313726, | |
| "eval_loss": 1.208784580230713, | |
| "eval_runtime": 6.6546, | |
| "eval_samples_per_second": 75.136, | |
| "eval_steps_per_second": 9.467, | |
| "step": 750 | |
| }, | |
| { | |
| "epoch": 4.0, | |
| "eval_exact_match": 23.4, | |
| "eval_f1": 33.18893657873498, | |
| "eval_qa_bleu": 3.8183981024520506, | |
| "eval_qa_exact_match": 0.176, | |
| "eval_recite_bleu": 14.071857337313705, | |
| "eval_recite_exact_match": 0.0, | |
| "step": 750 | |
| }, | |
| { | |
| "epoch": 4.27, | |
| "learning_rate": 5e-05, | |
| "loss": 1.0179, | |
| "step": 800 | |
| }, | |
| { | |
| "epoch": 4.8, | |
| "learning_rate": 5e-05, | |
| "loss": 1.0066, | |
| "step": 900 | |
| }, | |
| { | |
| "epoch": 5.0, | |
| "eval_accuracy": 0.6651764705882353, | |
| "eval_loss": 1.2342489957809448, | |
| "eval_runtime": 6.699, | |
| "eval_samples_per_second": 74.638, | |
| "eval_steps_per_second": 9.404, | |
| "step": 937 | |
| }, | |
| { | |
| "epoch": 5.0, | |
| "eval_exact_match": 24.0, | |
| "eval_f1": 32.88692081962977, | |
| "eval_qa_bleu": 3.6747877275980905, | |
| "eval_qa_exact_match": 0.174, | |
| "eval_recite_bleu": 14.36158051041567, | |
| "eval_recite_exact_match": 0.006, | |
| "step": 937 | |
| }, | |
| { | |
| "epoch": 5.33, | |
| "learning_rate": 5e-05, | |
| "loss": 0.9595, | |
| "step": 1000 | |
| }, | |
| { | |
| "epoch": 5.87, | |
| "learning_rate": 5e-05, | |
| "loss": 0.9353, | |
| "step": 1100 | |
| }, | |
| { | |
| "epoch": 6.0, | |
| "eval_accuracy": 0.6628392156862745, | |
| "eval_loss": 1.2645816802978516, | |
| "eval_runtime": 7.4539, | |
| "eval_samples_per_second": 67.079, | |
| "eval_steps_per_second": 8.452, | |
| "step": 1125 | |
| }, | |
| { | |
| "epoch": 6.0, | |
| "eval_exact_match": 23.4, | |
| "eval_f1": 33.50380952380953, | |
| "eval_qa_bleu": 13.968612067397348, | |
| "eval_qa_exact_match": 0.168, | |
| "eval_recite_bleu": 15.026821124315997, | |
| "eval_recite_exact_match": 0.012, | |
| "step": 1125 | |
| }, | |
| { | |
| "epoch": 6.4, | |
| "learning_rate": 5e-05, | |
| "loss": 0.8774, | |
| "step": 1200 | |
| }, | |
| { | |
| "epoch": 6.93, | |
| "learning_rate": 5e-05, | |
| "loss": 0.8629, | |
| "step": 1300 | |
| }, | |
| { | |
| "epoch": 7.0, | |
| "eval_accuracy": 0.6617333333333333, | |
| "eval_loss": 1.2990769147872925, | |
| "eval_runtime": 6.6963, | |
| "eval_samples_per_second": 74.668, | |
| "eval_steps_per_second": 9.408, | |
| "step": 1312 | |
| }, | |
| { | |
| "epoch": 7.0, | |
| "eval_exact_match": 22.0, | |
| "eval_f1": 31.8904761904762, | |
| "eval_qa_bleu": 15.632357578213835, | |
| "eval_qa_exact_match": 0.166, | |
| "eval_recite_bleu": 14.318212267740977, | |
| "eval_recite_exact_match": 0.016, | |
| "step": 1312 | |
| }, | |
| { | |
| "epoch": 7.47, | |
| "learning_rate": 5e-05, | |
| "loss": 0.8, | |
| "step": 1400 | |
| }, | |
| { | |
| "epoch": 8.0, | |
| "learning_rate": 5e-05, | |
| "loss": 0.7933, | |
| "step": 1500 | |
| }, | |
| { | |
| "epoch": 8.0, | |
| "eval_accuracy": 0.659607843137255, | |
| "eval_loss": 1.3466852903366089, | |
| "eval_runtime": 6.6563, | |
| "eval_samples_per_second": 75.117, | |
| "eval_steps_per_second": 9.465, | |
| "step": 1500 | |
| }, | |
| { | |
| "epoch": 8.0, | |
| "eval_exact_match": 22.6, | |
| "eval_f1": 31.16493506493508, | |
| "eval_qa_bleu": 21.45561748665761, | |
| "eval_qa_exact_match": 0.17, | |
| "eval_recite_bleu": 14.23612921374107, | |
| "eval_recite_exact_match": 0.012, | |
| "step": 1500 | |
| }, | |
| { | |
| "epoch": 8.53, | |
| "learning_rate": 5e-05, | |
| "loss": 0.7185, | |
| "step": 1600 | |
| }, | |
| { | |
| "epoch": 9.0, | |
| "eval_accuracy": 0.6574980392156863, | |
| "eval_loss": 1.391683578491211, | |
| "eval_runtime": 7.7056, | |
| "eval_samples_per_second": 64.888, | |
| "eval_steps_per_second": 8.176, | |
| "step": 1687 | |
| }, | |
| { | |
| "epoch": 9.0, | |
| "eval_exact_match": 21.8, | |
| "eval_f1": 31.150317460317464, | |
| "eval_qa_bleu": 18.411626756160196, | |
| "eval_qa_exact_match": 0.17, | |
| "eval_recite_bleu": 15.144735177405446, | |
| "eval_recite_exact_match": 0.014, | |
| "step": 1687 | |
| }, | |
| { | |
| "epoch": 9.07, | |
| "learning_rate": 5e-05, | |
| "loss": 0.7176, | |
| "step": 1700 | |
| }, | |
| { | |
| "epoch": 9.6, | |
| "learning_rate": 5e-05, | |
| "loss": 0.6489, | |
| "step": 1800 | |
| }, | |
| { | |
| "epoch": 10.0, | |
| "eval_accuracy": 0.6549647058823529, | |
| "eval_loss": 1.4609253406524658, | |
| "eval_runtime": 6.6599, | |
| "eval_samples_per_second": 75.076, | |
| "eval_steps_per_second": 9.46, | |
| "step": 1875 | |
| }, | |
| { | |
| "epoch": 10.0, | |
| "eval_exact_match": 24.8, | |
| "eval_f1": 32.80031746031747, | |
| "eval_qa_bleu": 19.203751437503612, | |
| "eval_qa_exact_match": 0.204, | |
| "eval_recite_bleu": 14.497298759377461, | |
| "eval_recite_exact_match": 0.016, | |
| "step": 1875 | |
| }, | |
| { | |
| "epoch": 10.13, | |
| "learning_rate": 5e-05, | |
| "loss": 0.6426, | |
| "step": 1900 | |
| }, | |
| { | |
| "epoch": 10.67, | |
| "learning_rate": 5e-05, | |
| "loss": 0.586, | |
| "step": 2000 | |
| }, | |
| { | |
| "epoch": 11.0, | |
| "eval_accuracy": 0.6531372549019608, | |
| "eval_loss": 1.5205447673797607, | |
| "eval_runtime": 6.6744, | |
| "eval_samples_per_second": 74.913, | |
| "eval_steps_per_second": 9.439, | |
| "step": 2062 | |
| }, | |
| { | |
| "epoch": 11.0, | |
| "eval_exact_match": 24.4, | |
| "eval_f1": 32.14761904761905, | |
| "eval_qa_bleu": 17.66314588893528, | |
| "eval_qa_exact_match": 0.198, | |
| "eval_recite_bleu": 14.285546168231928, | |
| "eval_recite_exact_match": 0.016, | |
| "step": 2062 | |
| }, | |
| { | |
| "epoch": 11.2, | |
| "learning_rate": 5e-05, | |
| "loss": 0.5625, | |
| "step": 2100 | |
| }, | |
| { | |
| "epoch": 11.73, | |
| "learning_rate": 5e-05, | |
| "loss": 0.5267, | |
| "step": 2200 | |
| }, | |
| { | |
| "epoch": 12.0, | |
| "eval_accuracy": 0.6518117647058823, | |
| "eval_loss": 1.5615981817245483, | |
| "eval_runtime": 7.0623, | |
| "eval_samples_per_second": 70.799, | |
| "eval_steps_per_second": 8.921, | |
| "step": 2250 | |
| }, | |
| { | |
| "epoch": 12.0, | |
| "eval_exact_match": 22.0, | |
| "eval_f1": 29.764505494505517, | |
| "eval_qa_bleu": 14.595863052742681, | |
| "eval_qa_exact_match": 0.174, | |
| "eval_recite_bleu": 15.269859538584459, | |
| "eval_recite_exact_match": 0.02, | |
| "step": 2250 | |
| }, | |
| { | |
| "epoch": 12.27, | |
| "learning_rate": 5e-05, | |
| "loss": 0.5074, | |
| "step": 2300 | |
| }, | |
| { | |
| "epoch": 12.8, | |
| "learning_rate": 5e-05, | |
| "loss": 0.4702, | |
| "step": 2400 | |
| }, | |
| { | |
| "epoch": 13.0, | |
| "eval_accuracy": 0.6498588235294117, | |
| "eval_loss": 1.6518086194992065, | |
| "eval_runtime": 7.4952, | |
| "eval_samples_per_second": 66.71, | |
| "eval_steps_per_second": 8.405, | |
| "step": 2437 | |
| }, | |
| { | |
| "epoch": 13.0, | |
| "eval_exact_match": 19.8, | |
| "eval_f1": 29.199523809523825, | |
| "eval_qa_bleu": 15.26272043877601, | |
| "eval_qa_exact_match": 0.156, | |
| "eval_recite_bleu": 13.712487638236569, | |
| "eval_recite_exact_match": 0.012, | |
| "step": 2437 | |
| }, | |
| { | |
| "epoch": 13.33, | |
| "learning_rate": 5e-05, | |
| "loss": 0.4335, | |
| "step": 2500 | |
| }, | |
| { | |
| "epoch": 13.87, | |
| "learning_rate": 5e-05, | |
| "loss": 0.42, | |
| "step": 2600 | |
| }, | |
| { | |
| "epoch": 14.0, | |
| "eval_accuracy": 0.6481411764705882, | |
| "eval_loss": 1.7193351984024048, | |
| "eval_runtime": 7.6621, | |
| "eval_samples_per_second": 65.256, | |
| "eval_steps_per_second": 8.222, | |
| "step": 2625 | |
| }, | |
| { | |
| "epoch": 14.0, | |
| "eval_exact_match": 18.4, | |
| "eval_f1": 27.123376623376636, | |
| "eval_qa_bleu": 13.906554273237205, | |
| "eval_qa_exact_match": 0.15, | |
| "eval_recite_bleu": 14.177162892556794, | |
| "eval_recite_exact_match": 0.016, | |
| "step": 2625 | |
| }, | |
| { | |
| "epoch": 14.4, | |
| "learning_rate": 5e-05, | |
| "loss": 0.3784, | |
| "step": 2700 | |
| }, | |
| { | |
| "epoch": 14.93, | |
| "learning_rate": 5e-05, | |
| "loss": 0.3811, | |
| "step": 2800 | |
| }, | |
| { | |
| "epoch": 15.0, | |
| "eval_accuracy": 0.6458509803921568, | |
| "eval_loss": 1.8015950918197632, | |
| "eval_runtime": 7.7113, | |
| "eval_samples_per_second": 64.84, | |
| "eval_steps_per_second": 8.17, | |
| "step": 2812 | |
| }, | |
| { | |
| "epoch": 15.0, | |
| "eval_exact_match": 19.0, | |
| "eval_f1": 28.891111111111115, | |
| "eval_qa_bleu": 16.83679026235492, | |
| "eval_qa_exact_match": 0.154, | |
| "eval_recite_bleu": 13.04215589870331, | |
| "eval_recite_exact_match": 0.012, | |
| "step": 2812 | |
| }, | |
| { | |
| "epoch": 15.47, | |
| "learning_rate": 5e-05, | |
| "loss": 0.3287, | |
| "step": 2900 | |
| }, | |
| { | |
| "epoch": 16.0, | |
| "learning_rate": 5e-05, | |
| "loss": 0.3373, | |
| "step": 3000 | |
| }, | |
| { | |
| "epoch": 16.0, | |
| "eval_accuracy": 0.6464627450980392, | |
| "eval_loss": 1.8555551767349243, | |
| "eval_runtime": 6.6548, | |
| "eval_samples_per_second": 75.133, | |
| "eval_steps_per_second": 9.467, | |
| "step": 3000 | |
| }, | |
| { | |
| "epoch": 16.0, | |
| "eval_exact_match": 21.0, | |
| "eval_f1": 30.546060606060603, | |
| "eval_qa_bleu": 15.51818073374785, | |
| "eval_qa_exact_match": 0.178, | |
| "eval_recite_bleu": 14.611971996497491, | |
| "eval_recite_exact_match": 0.016, | |
| "step": 3000 | |
| }, | |
| { | |
| "epoch": 16.53, | |
| "learning_rate": 5e-05, | |
| "loss": 0.284, | |
| "step": 3100 | |
| }, | |
| { | |
| "epoch": 17.0, | |
| "eval_accuracy": 0.644313725490196, | |
| "eval_loss": 1.9406696557998657, | |
| "eval_runtime": 6.7553, | |
| "eval_samples_per_second": 74.016, | |
| "eval_steps_per_second": 9.326, | |
| "step": 3187 | |
| }, | |
| { | |
| "epoch": 17.0, | |
| "eval_exact_match": 19.2, | |
| "eval_f1": 26.95666666666667, | |
| "eval_qa_bleu": 13.067462248071651, | |
| "eval_qa_exact_match": 0.148, | |
| "eval_recite_bleu": 13.516002177466241, | |
| "eval_recite_exact_match": 0.016, | |
| "step": 3187 | |
| }, | |
| { | |
| "epoch": 17.07, | |
| "learning_rate": 5e-05, | |
| "loss": 0.2933, | |
| "step": 3200 | |
| }, | |
| { | |
| "epoch": 17.6, | |
| "learning_rate": 5e-05, | |
| "loss": 0.25, | |
| "step": 3300 | |
| }, | |
| { | |
| "epoch": 18.0, | |
| "eval_accuracy": 0.6435764705882353, | |
| "eval_loss": 1.9759635925292969, | |
| "eval_runtime": 7.1655, | |
| "eval_samples_per_second": 69.779, | |
| "eval_steps_per_second": 8.792, | |
| "step": 3375 | |
| }, | |
| { | |
| "epoch": 18.0, | |
| "eval_exact_match": 18.6, | |
| "eval_f1": 27.20079365079366, | |
| "eval_qa_bleu": 12.454425793037345, | |
| "eval_qa_exact_match": 0.158, | |
| "eval_recite_bleu": 13.86644732825742, | |
| "eval_recite_exact_match": 0.014, | |
| "step": 3375 | |
| }, | |
| { | |
| "epoch": 18.13, | |
| "learning_rate": 5e-05, | |
| "loss": 0.2587, | |
| "step": 3400 | |
| }, | |
| { | |
| "epoch": 18.67, | |
| "learning_rate": 5e-05, | |
| "loss": 0.2274, | |
| "step": 3500 | |
| }, | |
| { | |
| "epoch": 19.0, | |
| "eval_accuracy": 0.6424313725490196, | |
| "eval_loss": 2.100283145904541, | |
| "eval_runtime": 7.4607, | |
| "eval_samples_per_second": 67.018, | |
| "eval_steps_per_second": 8.444, | |
| "step": 3562 | |
| }, | |
| { | |
| "epoch": 19.0, | |
| "eval_exact_match": 18.8, | |
| "eval_f1": 26.78857142857144, | |
| "eval_qa_bleu": 6.755652921468526, | |
| "eval_qa_exact_match": 0.16, | |
| "eval_recite_bleu": 13.96083887839662, | |
| "eval_recite_exact_match": 0.018, | |
| "step": 3562 | |
| }, | |
| { | |
| "epoch": 19.2, | |
| "learning_rate": 5e-05, | |
| "loss": 0.2209, | |
| "step": 3600 | |
| }, | |
| { | |
| "epoch": 19.73, | |
| "learning_rate": 5e-05, | |
| "loss": 0.2033, | |
| "step": 3700 | |
| }, | |
| { | |
| "epoch": 20.0, | |
| "eval_accuracy": 0.6424627450980392, | |
| "eval_loss": 2.141883373260498, | |
| "eval_runtime": 6.6558, | |
| "eval_samples_per_second": 75.122, | |
| "eval_steps_per_second": 9.465, | |
| "step": 3750 | |
| }, | |
| { | |
| "epoch": 20.0, | |
| "eval_exact_match": 20.4, | |
| "eval_f1": 28.975054945054936, | |
| "eval_qa_bleu": 14.791681538972547, | |
| "eval_qa_exact_match": 0.168, | |
| "eval_recite_bleu": 14.920846894015328, | |
| "eval_recite_exact_match": 0.016, | |
| "step": 3750 | |
| }, | |
| { | |
| "epoch": 20.27, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1928, | |
| "step": 3800 | |
| }, | |
| { | |
| "epoch": 20.8, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1832, | |
| "step": 3900 | |
| }, | |
| { | |
| "epoch": 21.0, | |
| "eval_accuracy": 0.6405333333333333, | |
| "eval_loss": 2.2004899978637695, | |
| "eval_runtime": 7.0682, | |
| "eval_samples_per_second": 70.739, | |
| "eval_steps_per_second": 8.913, | |
| "step": 3937 | |
| }, | |
| { | |
| "epoch": 21.0, | |
| "eval_exact_match": 19.8, | |
| "eval_f1": 27.03857142857144, | |
| "eval_qa_bleu": 14.16698111965369, | |
| "eval_qa_exact_match": 0.152, | |
| "eval_recite_bleu": 14.34411744472555, | |
| "eval_recite_exact_match": 0.016, | |
| "step": 3937 | |
| }, | |
| { | |
| "epoch": 21.33, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1674, | |
| "step": 4000 | |
| }, | |
| { | |
| "epoch": 21.87, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1678, | |
| "step": 4100 | |
| }, | |
| { | |
| "epoch": 22.0, | |
| "eval_accuracy": 0.6396627450980392, | |
| "eval_loss": 2.2465262413024902, | |
| "eval_runtime": 7.6611, | |
| "eval_samples_per_second": 65.264, | |
| "eval_steps_per_second": 8.223, | |
| "step": 4125 | |
| }, | |
| { | |
| "epoch": 22.0, | |
| "eval_exact_match": 18.8, | |
| "eval_f1": 26.868253968253967, | |
| "eval_qa_bleu": 16.683806863576216, | |
| "eval_qa_exact_match": 0.15, | |
| "eval_recite_bleu": 15.201480663492395, | |
| "eval_recite_exact_match": 0.018, | |
| "step": 4125 | |
| }, | |
| { | |
| "epoch": 22.4, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1516, | |
| "step": 4200 | |
| }, | |
| { | |
| "epoch": 22.93, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1505, | |
| "step": 4300 | |
| }, | |
| { | |
| "epoch": 23.0, | |
| "eval_accuracy": 0.6399607843137255, | |
| "eval_loss": 2.292829990386963, | |
| "eval_runtime": 7.6966, | |
| "eval_samples_per_second": 64.963, | |
| "eval_steps_per_second": 8.185, | |
| "step": 4312 | |
| }, | |
| { | |
| "epoch": 23.0, | |
| "eval_exact_match": 20.4, | |
| "eval_f1": 28.723809523809532, | |
| "eval_qa_bleu": 14.649664836489828, | |
| "eval_qa_exact_match": 0.16, | |
| "eval_recite_bleu": 14.44141809683148, | |
| "eval_recite_exact_match": 0.016, | |
| "step": 4312 | |
| }, | |
| { | |
| "epoch": 23.47, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1345, | |
| "step": 4400 | |
| }, | |
| { | |
| "epoch": 24.0, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1406, | |
| "step": 4500 | |
| }, | |
| { | |
| "epoch": 24.0, | |
| "eval_accuracy": 0.6399686274509804, | |
| "eval_loss": 2.369138240814209, | |
| "eval_runtime": 7.0721, | |
| "eval_samples_per_second": 70.701, | |
| "eval_steps_per_second": 8.908, | |
| "step": 4500 | |
| }, | |
| { | |
| "epoch": 24.0, | |
| "eval_exact_match": 17.8, | |
| "eval_f1": 25.43047619047621, | |
| "eval_qa_bleu": 13.339685450002685, | |
| "eval_qa_exact_match": 0.148, | |
| "eval_recite_bleu": 13.754679861863655, | |
| "eval_recite_exact_match": 0.018, | |
| "step": 4500 | |
| }, | |
| { | |
| "epoch": 24.53, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1214, | |
| "step": 4600 | |
| }, | |
| { | |
| "epoch": 25.0, | |
| "eval_accuracy": 0.6389882352941176, | |
| "eval_loss": 2.41001033782959, | |
| "eval_runtime": 7.2032, | |
| "eval_samples_per_second": 69.414, | |
| "eval_steps_per_second": 8.746, | |
| "step": 4687 | |
| }, | |
| { | |
| "epoch": 25.0, | |
| "eval_exact_match": 18.0, | |
| "eval_f1": 26.35142857142858, | |
| "eval_qa_bleu": 10.51643947107139, | |
| "eval_qa_exact_match": 0.146, | |
| "eval_recite_bleu": 13.73909597941028, | |
| "eval_recite_exact_match": 0.02, | |
| "step": 4687 | |
| }, | |
| { | |
| "epoch": 25.07, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1266, | |
| "step": 4700 | |
| }, | |
| { | |
| "epoch": 25.6, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1128, | |
| "step": 4800 | |
| }, | |
| { | |
| "epoch": 26.0, | |
| "eval_accuracy": 0.6381725490196078, | |
| "eval_loss": 2.484750986099243, | |
| "eval_runtime": 6.6571, | |
| "eval_samples_per_second": 75.107, | |
| "eval_steps_per_second": 9.464, | |
| "step": 4875 | |
| }, | |
| { | |
| "epoch": 26.0, | |
| "eval_exact_match": 19.0, | |
| "eval_f1": 28.15619047619049, | |
| "eval_qa_bleu": 7.883655295314112, | |
| "eval_qa_exact_match": 0.158, | |
| "eval_recite_bleu": 14.563137659888193, | |
| "eval_recite_exact_match": 0.02, | |
| "step": 4875 | |
| }, | |
| { | |
| "epoch": 26.13, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1179, | |
| "step": 4900 | |
| }, | |
| { | |
| "epoch": 26.67, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1076, | |
| "step": 5000 | |
| }, | |
| { | |
| "epoch": 27.0, | |
| "eval_accuracy": 0.638956862745098, | |
| "eval_loss": 2.546912908554077, | |
| "eval_runtime": 6.6822, | |
| "eval_samples_per_second": 74.826, | |
| "eval_steps_per_second": 9.428, | |
| "step": 5062 | |
| }, | |
| { | |
| "epoch": 27.0, | |
| "eval_exact_match": 18.8, | |
| "eval_f1": 25.95916971916973, | |
| "eval_qa_bleu": 6.531854323689431, | |
| "eval_qa_exact_match": 0.162, | |
| "eval_recite_bleu": 13.16436178935274, | |
| "eval_recite_exact_match": 0.022, | |
| "step": 5062 | |
| }, | |
| { | |
| "epoch": 27.2, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1095, | |
| "step": 5100 | |
| }, | |
| { | |
| "epoch": 27.73, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1046, | |
| "step": 5200 | |
| }, | |
| { | |
| "epoch": 28.0, | |
| "eval_accuracy": 0.6391529411764706, | |
| "eval_loss": 2.5205068588256836, | |
| "eval_runtime": 7.4592, | |
| "eval_samples_per_second": 67.031, | |
| "eval_steps_per_second": 8.446, | |
| "step": 5250 | |
| }, | |
| { | |
| "epoch": 28.0, | |
| "eval_exact_match": 17.2, | |
| "eval_f1": 25.731298701298712, | |
| "eval_qa_bleu": 8.206606144769799, | |
| "eval_qa_exact_match": 0.14, | |
| "eval_recite_bleu": 14.236424434530772, | |
| "eval_recite_exact_match": 0.022, | |
| "step": 5250 | |
| }, | |
| { | |
| "epoch": 28.27, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1003, | |
| "step": 5300 | |
| }, | |
| { | |
| "epoch": 28.8, | |
| "learning_rate": 5e-05, | |
| "loss": 0.1007, | |
| "step": 5400 | |
| }, | |
| { | |
| "epoch": 29.0, | |
| "eval_accuracy": 0.6381882352941176, | |
| "eval_loss": 2.575054407119751, | |
| "eval_runtime": 6.7399, | |
| "eval_samples_per_second": 74.185, | |
| "eval_steps_per_second": 9.347, | |
| "step": 5437 | |
| }, | |
| { | |
| "epoch": 29.0, | |
| "eval_exact_match": 16.6, | |
| "eval_f1": 25.44886446886448, | |
| "eval_qa_bleu": 11.416596929945294, | |
| "eval_qa_exact_match": 0.132, | |
| "eval_recite_bleu": 14.936294044523267, | |
| "eval_recite_exact_match": 0.026, | |
| "step": 5437 | |
| }, | |
| { | |
| "epoch": 29.33, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0956, | |
| "step": 5500 | |
| }, | |
| { | |
| "epoch": 29.87, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0967, | |
| "step": 5600 | |
| }, | |
| { | |
| "epoch": 30.0, | |
| "eval_accuracy": 0.6388784313725491, | |
| "eval_loss": 2.598768472671509, | |
| "eval_runtime": 6.6559, | |
| "eval_samples_per_second": 75.122, | |
| "eval_steps_per_second": 9.465, | |
| "step": 5625 | |
| }, | |
| { | |
| "epoch": 30.0, | |
| "eval_exact_match": 18.8, | |
| "eval_f1": 26.82857142857144, | |
| "eval_qa_bleu": 8.539846307112432, | |
| "eval_qa_exact_match": 0.148, | |
| "eval_recite_bleu": 14.818047491107617, | |
| "eval_recite_exact_match": 0.022, | |
| "step": 5625 | |
| }, | |
| { | |
| "epoch": 30.4, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0912, | |
| "step": 5700 | |
| }, | |
| { | |
| "epoch": 30.93, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0947, | |
| "step": 5800 | |
| }, | |
| { | |
| "epoch": 31.0, | |
| "eval_accuracy": 0.6390509803921569, | |
| "eval_loss": 2.6168441772460938, | |
| "eval_runtime": 7.5152, | |
| "eval_samples_per_second": 66.532, | |
| "eval_steps_per_second": 8.383, | |
| "step": 5812 | |
| }, | |
| { | |
| "epoch": 31.0, | |
| "eval_exact_match": 19.8, | |
| "eval_f1": 27.306190476190487, | |
| "eval_qa_bleu": 15.11903516220494, | |
| "eval_qa_exact_match": 0.146, | |
| "eval_recite_bleu": 15.095386960197741, | |
| "eval_recite_exact_match": 0.03, | |
| "step": 5812 | |
| }, | |
| { | |
| "epoch": 31.47, | |
| "learning_rate": 5e-05, | |
| "loss": 0.087, | |
| "step": 5900 | |
| }, | |
| { | |
| "epoch": 32.0, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0914, | |
| "step": 6000 | |
| }, | |
| { | |
| "epoch": 32.0, | |
| "eval_accuracy": 0.6393411764705882, | |
| "eval_loss": 2.666327476501465, | |
| "eval_runtime": 6.6541, | |
| "eval_samples_per_second": 75.142, | |
| "eval_steps_per_second": 9.468, | |
| "step": 6000 | |
| }, | |
| { | |
| "epoch": 32.0, | |
| "eval_exact_match": 18.2, | |
| "eval_f1": 27.904761904761916, | |
| "eval_qa_bleu": 10.102774569302643, | |
| "eval_qa_exact_match": 0.148, | |
| "eval_recite_bleu": 14.86885556528055, | |
| "eval_recite_exact_match": 0.024, | |
| "step": 6000 | |
| }, | |
| { | |
| "epoch": 32.53, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0834, | |
| "step": 6100 | |
| }, | |
| { | |
| "epoch": 33.0, | |
| "eval_accuracy": 0.639521568627451, | |
| "eval_loss": 2.6627161502838135, | |
| "eval_runtime": 6.719, | |
| "eval_samples_per_second": 74.416, | |
| "eval_steps_per_second": 9.376, | |
| "step": 6187 | |
| }, | |
| { | |
| "epoch": 33.0, | |
| "eval_exact_match": 21.0, | |
| "eval_f1": 28.334885844748868, | |
| "eval_qa_bleu": 12.177590782661031, | |
| "eval_qa_exact_match": 0.17, | |
| "eval_recite_bleu": 14.44189612883832, | |
| "eval_recite_exact_match": 0.026, | |
| "step": 6187 | |
| }, | |
| { | |
| "epoch": 33.07, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0878, | |
| "step": 6200 | |
| }, | |
| { | |
| "epoch": 33.6, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0827, | |
| "step": 6300 | |
| }, | |
| { | |
| "epoch": 34.0, | |
| "eval_accuracy": 0.6383450980392157, | |
| "eval_loss": 2.6657156944274902, | |
| "eval_runtime": 6.6564, | |
| "eval_samples_per_second": 75.116, | |
| "eval_steps_per_second": 9.465, | |
| "step": 6375 | |
| }, | |
| { | |
| "epoch": 34.0, | |
| "eval_exact_match": 17.8, | |
| "eval_f1": 25.287619047619064, | |
| "eval_qa_bleu": 13.43249770935417, | |
| "eval_qa_exact_match": 0.144, | |
| "eval_recite_bleu": 14.257830156074794, | |
| "eval_recite_exact_match": 0.026, | |
| "step": 6375 | |
| }, | |
| { | |
| "epoch": 34.13, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0849, | |
| "step": 6400 | |
| }, | |
| { | |
| "epoch": 34.67, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0818, | |
| "step": 6500 | |
| }, | |
| { | |
| "epoch": 35.0, | |
| "eval_accuracy": 0.6389882352941176, | |
| "eval_loss": 2.6941633224487305, | |
| "eval_runtime": 6.6666, | |
| "eval_samples_per_second": 75.001, | |
| "eval_steps_per_second": 9.45, | |
| "step": 6562 | |
| }, | |
| { | |
| "epoch": 35.0, | |
| "eval_exact_match": 20.0, | |
| "eval_f1": 26.61507936507937, | |
| "eval_qa_bleu": 11.98196398422812, | |
| "eval_qa_exact_match": 0.166, | |
| "eval_recite_bleu": 15.297615086859388, | |
| "eval_recite_exact_match": 0.03, | |
| "step": 6562 | |
| }, | |
| { | |
| "epoch": 35.2, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0826, | |
| "step": 6600 | |
| }, | |
| { | |
| "epoch": 35.73, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0817, | |
| "step": 6700 | |
| }, | |
| { | |
| "epoch": 36.0, | |
| "eval_accuracy": 0.6383686274509804, | |
| "eval_loss": 2.7361207008361816, | |
| "eval_runtime": 6.6649, | |
| "eval_samples_per_second": 75.02, | |
| "eval_steps_per_second": 9.453, | |
| "step": 6750 | |
| }, | |
| { | |
| "epoch": 36.0, | |
| "eval_exact_match": 15.6, | |
| "eval_f1": 24.341428571428583, | |
| "eval_qa_bleu": 12.164230009686962, | |
| "eval_qa_exact_match": 0.122, | |
| "eval_recite_bleu": 14.713130786681702, | |
| "eval_recite_exact_match": 0.028, | |
| "step": 6750 | |
| }, | |
| { | |
| "epoch": 36.27, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0806, | |
| "step": 6800 | |
| }, | |
| { | |
| "epoch": 36.8, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0799, | |
| "step": 6900 | |
| }, | |
| { | |
| "epoch": 37.0, | |
| "eval_accuracy": 0.6386823529411765, | |
| "eval_loss": 2.7282583713531494, | |
| "eval_runtime": 7.0995, | |
| "eval_samples_per_second": 70.428, | |
| "eval_steps_per_second": 8.874, | |
| "step": 6937 | |
| }, | |
| { | |
| "epoch": 37.0, | |
| "eval_exact_match": 19.0, | |
| "eval_f1": 26.294444444444455, | |
| "eval_qa_bleu": 12.126807356073423, | |
| "eval_qa_exact_match": 0.158, | |
| "eval_recite_bleu": 14.707435095434732, | |
| "eval_recite_exact_match": 0.028, | |
| "step": 6937 | |
| }, | |
| { | |
| "epoch": 37.33, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0776, | |
| "step": 7000 | |
| }, | |
| { | |
| "epoch": 37.87, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0784, | |
| "step": 7100 | |
| }, | |
| { | |
| "epoch": 38.0, | |
| "eval_accuracy": 0.6379843137254902, | |
| "eval_loss": 2.7539706230163574, | |
| "eval_runtime": 7.4637, | |
| "eval_samples_per_second": 66.991, | |
| "eval_steps_per_second": 8.441, | |
| "step": 7125 | |
| }, | |
| { | |
| "epoch": 38.0, | |
| "eval_exact_match": 17.8, | |
| "eval_f1": 25.752582972582978, | |
| "eval_qa_bleu": 13.042360364860006, | |
| "eval_qa_exact_match": 0.148, | |
| "eval_recite_bleu": 13.422661654351161, | |
| "eval_recite_exact_match": 0.014, | |
| "step": 7125 | |
| }, | |
| { | |
| "epoch": 38.4, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0765, | |
| "step": 7200 | |
| }, | |
| { | |
| "epoch": 38.93, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0788, | |
| "step": 7300 | |
| }, | |
| { | |
| "epoch": 39.0, | |
| "eval_accuracy": 0.6387607843137255, | |
| "eval_loss": 2.7975375652313232, | |
| "eval_runtime": 7.7145, | |
| "eval_samples_per_second": 64.813, | |
| "eval_steps_per_second": 8.166, | |
| "step": 7312 | |
| }, | |
| { | |
| "epoch": 39.0, | |
| "eval_exact_match": 17.8, | |
| "eval_f1": 25.90701298701299, | |
| "eval_qa_bleu": 14.875623156941945, | |
| "eval_qa_exact_match": 0.142, | |
| "eval_recite_bleu": 14.139663492994524, | |
| "eval_recite_exact_match": 0.02, | |
| "step": 7312 | |
| }, | |
| { | |
| "epoch": 39.47, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0751, | |
| "step": 7400 | |
| }, | |
| { | |
| "epoch": 40.0, | |
| "learning_rate": 5e-05, | |
| "loss": 0.079, | |
| "step": 7500 | |
| }, | |
| { | |
| "epoch": 40.0, | |
| "eval_accuracy": 0.6390901960784314, | |
| "eval_loss": 2.7849485874176025, | |
| "eval_runtime": 6.6517, | |
| "eval_samples_per_second": 75.168, | |
| "eval_steps_per_second": 9.471, | |
| "step": 7500 | |
| }, | |
| { | |
| "epoch": 40.0, | |
| "eval_exact_match": 19.6, | |
| "eval_f1": 26.523809523809533, | |
| "eval_qa_bleu": 12.596822978027996, | |
| "eval_qa_exact_match": 0.152, | |
| "eval_recite_bleu": 15.111971741554742, | |
| "eval_recite_exact_match": 0.022, | |
| "step": 7500 | |
| }, | |
| { | |
| "epoch": 40.53, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0734, | |
| "step": 7600 | |
| }, | |
| { | |
| "epoch": 41.0, | |
| "eval_accuracy": 0.6383764705882353, | |
| "eval_loss": 2.8044795989990234, | |
| "eval_runtime": 7.6611, | |
| "eval_samples_per_second": 65.265, | |
| "eval_steps_per_second": 8.223, | |
| "step": 7687 | |
| }, | |
| { | |
| "epoch": 41.0, | |
| "eval_exact_match": 17.2, | |
| "eval_f1": 25.060268620268626, | |
| "eval_qa_bleu": 13.678395766110564, | |
| "eval_qa_exact_match": 0.138, | |
| "eval_recite_bleu": 13.766722868189284, | |
| "eval_recite_exact_match": 0.018, | |
| "step": 7687 | |
| }, | |
| { | |
| "epoch": 41.07, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0773, | |
| "step": 7700 | |
| }, | |
| { | |
| "epoch": 41.6, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0735, | |
| "step": 7800 | |
| }, | |
| { | |
| "epoch": 42.0, | |
| "eval_accuracy": 0.6389254901960785, | |
| "eval_loss": 2.806025981903076, | |
| "eval_runtime": 6.6643, | |
| "eval_samples_per_second": 75.027, | |
| "eval_steps_per_second": 9.453, | |
| "step": 7875 | |
| }, | |
| { | |
| "epoch": 42.0, | |
| "eval_exact_match": 18.4, | |
| "eval_f1": 25.394761904761907, | |
| "eval_qa_bleu": 14.274991754808283, | |
| "eval_qa_exact_match": 0.152, | |
| "eval_recite_bleu": 12.166894147679086, | |
| "eval_recite_exact_match": 0.01, | |
| "step": 7875 | |
| }, | |
| { | |
| "epoch": 42.13, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0751, | |
| "step": 7900 | |
| }, | |
| { | |
| "epoch": 42.67, | |
| "learning_rate": 5e-05, | |
| "loss": 0.073, | |
| "step": 8000 | |
| }, | |
| { | |
| "epoch": 43.0, | |
| "eval_accuracy": 0.6387450980392156, | |
| "eval_loss": 2.8026182651519775, | |
| "eval_runtime": 7.4771, | |
| "eval_samples_per_second": 66.871, | |
| "eval_steps_per_second": 8.426, | |
| "step": 8062 | |
| }, | |
| { | |
| "epoch": 43.0, | |
| "eval_exact_match": 16.0, | |
| "eval_f1": 23.73333333333334, | |
| "eval_qa_bleu": 9.497279831354774, | |
| "eval_qa_exact_match": 0.13, | |
| "eval_recite_bleu": 12.740344920561881, | |
| "eval_recite_exact_match": 0.014, | |
| "step": 8062 | |
| }, | |
| { | |
| "epoch": 43.2, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0746, | |
| "step": 8100 | |
| }, | |
| { | |
| "epoch": 43.73, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0737, | |
| "step": 8200 | |
| }, | |
| { | |
| "epoch": 44.0, | |
| "eval_accuracy": 0.638643137254902, | |
| "eval_loss": 2.8393852710723877, | |
| "eval_runtime": 7.075, | |
| "eval_samples_per_second": 70.671, | |
| "eval_steps_per_second": 8.905, | |
| "step": 8250 | |
| }, | |
| { | |
| "epoch": 44.0, | |
| "eval_exact_match": 14.2, | |
| "eval_f1": 22.128095238095252, | |
| "eval_qa_bleu": 10.28150817786033, | |
| "eval_qa_exact_match": 0.108, | |
| "eval_recite_bleu": 11.743705431581837, | |
| "eval_recite_exact_match": 0.014, | |
| "step": 8250 | |
| }, | |
| { | |
| "epoch": 44.27, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0734, | |
| "step": 8300 | |
| }, | |
| { | |
| "epoch": 44.8, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0736, | |
| "step": 8400 | |
| }, | |
| { | |
| "epoch": 45.0, | |
| "eval_accuracy": 0.6390588235294118, | |
| "eval_loss": 2.811821460723877, | |
| "eval_runtime": 7.4896, | |
| "eval_samples_per_second": 66.759, | |
| "eval_steps_per_second": 8.412, | |
| "step": 8437 | |
| }, | |
| { | |
| "epoch": 45.0, | |
| "eval_exact_match": 16.6, | |
| "eval_f1": 22.322857142857153, | |
| "eval_qa_bleu": 10.231299353118551, | |
| "eval_qa_exact_match": 0.134, | |
| "eval_recite_bleu": 11.37344332610696, | |
| "eval_recite_exact_match": 0.008, | |
| "step": 8437 | |
| }, | |
| { | |
| "epoch": 45.33, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0711, | |
| "step": 8500 | |
| }, | |
| { | |
| "epoch": 45.87, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0724, | |
| "step": 8600 | |
| }, | |
| { | |
| "epoch": 46.0, | |
| "eval_accuracy": 0.6388313725490196, | |
| "eval_loss": 2.849515438079834, | |
| "eval_runtime": 6.6632, | |
| "eval_samples_per_second": 75.039, | |
| "eval_steps_per_second": 9.455, | |
| "step": 8625 | |
| }, | |
| { | |
| "epoch": 46.0, | |
| "eval_exact_match": 16.8, | |
| "eval_f1": 23.55984126984127, | |
| "eval_qa_bleu": 11.575646857443479, | |
| "eval_qa_exact_match": 0.132, | |
| "eval_recite_bleu": 11.803037448505567, | |
| "eval_recite_exact_match": 0.012, | |
| "step": 8625 | |
| }, | |
| { | |
| "epoch": 46.4, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0696, | |
| "step": 8700 | |
| }, | |
| { | |
| "epoch": 46.93, | |
| "learning_rate": 5e-05, | |
| "loss": 0.0721, | |
| "step": 8800 | |
| }, | |
| { | |
| "epoch": 47.0, | |
| "eval_accuracy": 0.638964705882353, | |
| "eval_loss": 2.844177722930908, | |
| "eval_runtime": 7.5389, | |
| "eval_samples_per_second": 66.323, | |
| "eval_steps_per_second": 8.357, | |
| "step": 8812 | |
| }, | |
| { | |
| "epoch": 47.0, | |
| "eval_exact_match": 16.4, | |
| "eval_f1": 22.471111111111114, | |
| "eval_qa_bleu": 7.236042097106983, | |
| "eval_qa_exact_match": 0.124, | |
| "eval_recite_bleu": 12.05434822783003, | |
| "eval_recite_exact_match": 0.012, | |
| "step": 8812 | |
| } | |
| ], | |
| "logging_steps": 100, | |
| "max_steps": 9350, | |
| "num_train_epochs": 50, | |
| "save_steps": 500, | |
| "total_flos": 2.0422792020104315e+18, | |
| "trial_name": null, | |
| "trial_params": null | |
| } | |