| { | |
| "best_metric": 0.44814541935920715, | |
| "best_model_checkpoint": "model_output/e2e_opentable_5_way__approximate__0-shot__seed-42__roberta-base/checkpoint-1450", | |
| "epoch": 2.4871355060034306, | |
| "global_step": 1450, | |
| "is_hyper_param_search": false, | |
| "is_local_process_zero": true, | |
| "is_world_process_zero": true, | |
| "log_history": [ | |
| { | |
| "epoch": 0.03, | |
| "learning_rate": 4.9783362218370885e-05, | |
| "loss": 15.3848, | |
| "step": 20 | |
| }, | |
| { | |
| "epoch": 0.07, | |
| "learning_rate": 4.956672443674177e-05, | |
| "loss": 12.8162, | |
| "step": 40 | |
| }, | |
| { | |
| "epoch": 0.09, | |
| "eval_distillation_accuracy_counterfactual": 0.4474089276552078, | |
| "eval_distillation_accuracy_factual": 0.7721908671113391, | |
| "eval_distillation_f1_counterfactual": 0.34384085542828363, | |
| "eval_distillation_f1_factual": 0.7449065740174884, | |
| "eval_groundtruth_accuracy_counterfactual": 0.43560800410466904, | |
| "eval_groundtruth_f1_counterfactual": 0.3322658600412939, | |
| "eval_groundtruth_f1_factual": 0.6497934851385894, | |
| "eval_icace_cosine": 0.4707597494125366, | |
| "eval_icace_l2": 0.7290830612182617, | |
| "eval_icace_normdiff": 0.5084162950515747, | |
| "eval_loss": 9.263871192932129, | |
| "eval_runtime": 12.7695, | |
| "eval_samples_per_second": 305.258, | |
| "eval_steps_per_second": 2.428, | |
| "step": 50 | |
| }, | |
| { | |
| "epoch": 0.1, | |
| "learning_rate": 4.935008665511265e-05, | |
| "loss": 11.524, | |
| "step": 60 | |
| }, | |
| { | |
| "epoch": 0.14, | |
| "learning_rate": 4.913344887348354e-05, | |
| "loss": 10.7187, | |
| "step": 80 | |
| }, | |
| { | |
| "epoch": 0.17, | |
| "learning_rate": 4.891681109185442e-05, | |
| "loss": 10.3266, | |
| "step": 100 | |
| }, | |
| { | |
| "epoch": 0.17, | |
| "eval_distillation_accuracy_counterfactual": 0.47485890200102615, | |
| "eval_distillation_accuracy_factual": 0.8653155464340687, | |
| "eval_distillation_f1_counterfactual": 0.352733121767057, | |
| "eval_distillation_f1_factual": 0.8531998683010464, | |
| "eval_groundtruth_accuracy_counterfactual": 0.47691123653155465, | |
| "eval_groundtruth_f1_counterfactual": 0.3544569665529119, | |
| "eval_groundtruth_f1_factual": 0.7055142150493701, | |
| "eval_icace_cosine": 0.46635791659355164, | |
| "eval_icace_l2": 0.7056604623794556, | |
| "eval_icace_normdiff": 0.4767073690891266, | |
| "eval_loss": 8.300187110900879, | |
| "eval_runtime": 13.1062, | |
| "eval_samples_per_second": 297.416, | |
| "eval_steps_per_second": 2.365, | |
| "step": 100 | |
| }, | |
| { | |
| "epoch": 0.21, | |
| "learning_rate": 4.8700173310225307e-05, | |
| "loss": 9.8902, | |
| "step": 120 | |
| }, | |
| { | |
| "epoch": 0.24, | |
| "learning_rate": 4.848353552859619e-05, | |
| "loss": 9.9859, | |
| "step": 140 | |
| }, | |
| { | |
| "epoch": 0.26, | |
| "eval_distillation_accuracy_counterfactual": 0.46587993842996406, | |
| "eval_distillation_accuracy_factual": 0.8286300667008722, | |
| "eval_distillation_f1_counterfactual": 0.3884433215928239, | |
| "eval_distillation_f1_factual": 0.8139544299165415, | |
| "eval_groundtruth_accuracy_counterfactual": 0.46151872755259105, | |
| "eval_groundtruth_f1_counterfactual": 0.3891775674468404, | |
| "eval_groundtruth_f1_factual": 0.6795376489414218, | |
| "eval_icace_cosine": 0.4685736894607544, | |
| "eval_icace_l2": 0.6997411251068115, | |
| "eval_icace_normdiff": 0.4409981966018677, | |
| "eval_loss": 8.554189682006836, | |
| "eval_runtime": 12.8948, | |
| "eval_samples_per_second": 302.293, | |
| "eval_steps_per_second": 2.404, | |
| "step": 150 | |
| }, | |
| { | |
| "epoch": 0.27, | |
| "learning_rate": 4.826689774696707e-05, | |
| "loss": 9.9401, | |
| "step": 160 | |
| }, | |
| { | |
| "epoch": 0.31, | |
| "learning_rate": 4.8050259965337955e-05, | |
| "loss": 9.8133, | |
| "step": 180 | |
| }, | |
| { | |
| "epoch": 0.34, | |
| "learning_rate": 4.7833622183708845e-05, | |
| "loss": 9.5608, | |
| "step": 200 | |
| }, | |
| { | |
| "epoch": 0.34, | |
| "eval_distillation_accuracy_counterfactual": 0.49589533093894306, | |
| "eval_distillation_accuracy_factual": 0.8476141611082606, | |
| "eval_distillation_f1_counterfactual": 0.41223146480687944, | |
| "eval_distillation_f1_factual": 0.839115183452182, | |
| "eval_groundtruth_accuracy_counterfactual": 0.4979476654694715, | |
| "eval_groundtruth_f1_counterfactual": 0.4149936071266812, | |
| "eval_groundtruth_f1_factual": 0.6763507473779132, | |
| "eval_icace_cosine": 0.461904913187027, | |
| "eval_icace_l2": 0.6693803668022156, | |
| "eval_icace_normdiff": 0.4146590232849121, | |
| "eval_loss": 8.107243537902832, | |
| "eval_runtime": 13.0158, | |
| "eval_samples_per_second": 299.481, | |
| "eval_steps_per_second": 2.382, | |
| "step": 200 | |
| }, | |
| { | |
| "epoch": 0.38, | |
| "learning_rate": 4.761698440207972e-05, | |
| "loss": 9.6041, | |
| "step": 220 | |
| }, | |
| { | |
| "epoch": 0.41, | |
| "learning_rate": 4.740034662045061e-05, | |
| "loss": 9.4542, | |
| "step": 240 | |
| }, | |
| { | |
| "epoch": 0.43, | |
| "eval_distillation_accuracy_counterfactual": 0.5123140071831709, | |
| "eval_distillation_accuracy_factual": 0.8365828630066701, | |
| "eval_distillation_f1_counterfactual": 0.41742508528245137, | |
| "eval_distillation_f1_factual": 0.8170598509732152, | |
| "eval_groundtruth_accuracy_counterfactual": 0.4923037455105182, | |
| "eval_groundtruth_f1_counterfactual": 0.39565449482548276, | |
| "eval_groundtruth_f1_factual": 0.695773274631551, | |
| "eval_icace_cosine": 0.4568592607975006, | |
| "eval_icace_l2": 0.6699839234352112, | |
| "eval_icace_normdiff": 0.4179259240627289, | |
| "eval_loss": 8.091322898864746, | |
| "eval_runtime": 13.0177, | |
| "eval_samples_per_second": 299.438, | |
| "eval_steps_per_second": 2.381, | |
| "step": 250 | |
| }, | |
| { | |
| "epoch": 0.45, | |
| "learning_rate": 4.7183708838821494e-05, | |
| "loss": 9.4606, | |
| "step": 260 | |
| }, | |
| { | |
| "epoch": 0.48, | |
| "learning_rate": 4.6967071057192376e-05, | |
| "loss": 9.4165, | |
| "step": 280 | |
| }, | |
| { | |
| "epoch": 0.51, | |
| "learning_rate": 4.675043327556326e-05, | |
| "loss": 9.3996, | |
| "step": 300 | |
| }, | |
| { | |
| "epoch": 0.51, | |
| "eval_distillation_accuracy_counterfactual": 0.5161621344279117, | |
| "eval_distillation_accuracy_factual": 0.8445356593124679, | |
| "eval_distillation_f1_counterfactual": 0.43140398163976057, | |
| "eval_distillation_f1_factual": 0.8243538167177713, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5071831708568497, | |
| "eval_groundtruth_f1_counterfactual": 0.4211696295155843, | |
| "eval_groundtruth_f1_factual": 0.6804350127556201, | |
| "eval_icace_cosine": 0.4562739133834839, | |
| "eval_icace_l2": 0.671053409576416, | |
| "eval_icace_normdiff": 0.42435094714164734, | |
| "eval_loss": 7.912776470184326, | |
| "eval_runtime": 13.0458, | |
| "eval_samples_per_second": 298.794, | |
| "eval_steps_per_second": 2.376, | |
| "step": 300 | |
| }, | |
| { | |
| "epoch": 0.55, | |
| "learning_rate": 4.653379549393415e-05, | |
| "loss": 9.2297, | |
| "step": 320 | |
| }, | |
| { | |
| "epoch": 0.58, | |
| "learning_rate": 4.6317157712305025e-05, | |
| "loss": 9.2203, | |
| "step": 340 | |
| }, | |
| { | |
| "epoch": 0.6, | |
| "eval_distillation_accuracy_counterfactual": 0.5179579271421242, | |
| "eval_distillation_accuracy_factual": 0.8155464340687532, | |
| "eval_distillation_f1_counterfactual": 0.4682566845228989, | |
| "eval_distillation_f1_factual": 0.8006032026406558, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5020523345305284, | |
| "eval_groundtruth_f1_counterfactual": 0.45492108472946924, | |
| "eval_groundtruth_f1_factual": 0.666677263616981, | |
| "eval_icace_cosine": 0.455741286277771, | |
| "eval_icace_l2": 0.6516066789627075, | |
| "eval_icace_normdiff": 0.4051017761230469, | |
| "eval_loss": 8.105178833007812, | |
| "eval_runtime": 12.9117, | |
| "eval_samples_per_second": 301.897, | |
| "eval_steps_per_second": 2.401, | |
| "step": 350 | |
| }, | |
| { | |
| "epoch": 0.62, | |
| "learning_rate": 4.6100519930675915e-05, | |
| "loss": 9.0931, | |
| "step": 360 | |
| }, | |
| { | |
| "epoch": 0.65, | |
| "learning_rate": 4.58838821490468e-05, | |
| "loss": 8.9297, | |
| "step": 380 | |
| }, | |
| { | |
| "epoch": 0.69, | |
| "learning_rate": 4.566724436741768e-05, | |
| "loss": 8.8054, | |
| "step": 400 | |
| }, | |
| { | |
| "epoch": 0.69, | |
| "eval_distillation_accuracy_counterfactual": 0.5053873781426372, | |
| "eval_distillation_accuracy_factual": 0.8222165212929707, | |
| "eval_distillation_f1_counterfactual": 0.38822036619209827, | |
| "eval_distillation_f1_factual": 0.8117129824518028, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5010261672652643, | |
| "eval_groundtruth_f1_counterfactual": 0.38820947576310105, | |
| "eval_groundtruth_f1_factual": 0.6914330659012916, | |
| "eval_icace_cosine": 0.45915645360946655, | |
| "eval_icace_l2": 0.6688275933265686, | |
| "eval_icace_normdiff": 0.4207516312599182, | |
| "eval_loss": 8.215511322021484, | |
| "eval_runtime": 13.0382, | |
| "eval_samples_per_second": 298.967, | |
| "eval_steps_per_second": 2.378, | |
| "step": 400 | |
| }, | |
| { | |
| "epoch": 0.72, | |
| "learning_rate": 4.5450606585788563e-05, | |
| "loss": 9.0175, | |
| "step": 420 | |
| }, | |
| { | |
| "epoch": 0.75, | |
| "learning_rate": 4.5233968804159446e-05, | |
| "loss": 9.0479, | |
| "step": 440 | |
| }, | |
| { | |
| "epoch": 0.77, | |
| "eval_distillation_accuracy_counterfactual": 0.5161621344279117, | |
| "eval_distillation_accuracy_factual": 0.8322216521292971, | |
| "eval_distillation_f1_counterfactual": 0.43163093824456045, | |
| "eval_distillation_f1_factual": 0.8179434971552079, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5164186762442278, | |
| "eval_groundtruth_f1_counterfactual": 0.43368043593910377, | |
| "eval_groundtruth_f1_factual": 0.7010946530829851, | |
| "eval_icace_cosine": 0.46967530250549316, | |
| "eval_icace_l2": 0.6567780375480652, | |
| "eval_icace_normdiff": 0.38229528069496155, | |
| "eval_loss": 8.273370742797852, | |
| "eval_runtime": 12.9013, | |
| "eval_samples_per_second": 302.139, | |
| "eval_steps_per_second": 2.403, | |
| "step": 450 | |
| }, | |
| { | |
| "epoch": 0.79, | |
| "learning_rate": 4.501733102253033e-05, | |
| "loss": 8.6738, | |
| "step": 460 | |
| }, | |
| { | |
| "epoch": 0.82, | |
| "learning_rate": 4.480069324090121e-05, | |
| "loss": 8.8553, | |
| "step": 480 | |
| }, | |
| { | |
| "epoch": 0.86, | |
| "learning_rate": 4.45840554592721e-05, | |
| "loss": 8.7079, | |
| "step": 500 | |
| }, | |
| { | |
| "epoch": 0.86, | |
| "eval_distillation_accuracy_counterfactual": 0.5023088763468445, | |
| "eval_distillation_accuracy_factual": 0.8260646485377117, | |
| "eval_distillation_f1_counterfactual": 0.39076709834402334, | |
| "eval_distillation_f1_factual": 0.8131187804629831, | |
| "eval_groundtruth_accuracy_counterfactual": 0.4974345818368394, | |
| "eval_groundtruth_f1_counterfactual": 0.38463954865815575, | |
| "eval_groundtruth_f1_factual": 0.7000450508607393, | |
| "eval_icace_cosine": 0.474086195230484, | |
| "eval_icace_l2": 0.6828737258911133, | |
| "eval_icace_normdiff": 0.40336331725120544, | |
| "eval_loss": 8.616800308227539, | |
| "eval_runtime": 13.0775, | |
| "eval_samples_per_second": 298.07, | |
| "eval_steps_per_second": 2.37, | |
| "step": 500 | |
| }, | |
| { | |
| "epoch": 0.89, | |
| "learning_rate": 4.436741767764298e-05, | |
| "loss": 8.795, | |
| "step": 520 | |
| }, | |
| { | |
| "epoch": 0.93, | |
| "learning_rate": 4.415077989601387e-05, | |
| "loss": 8.8781, | |
| "step": 540 | |
| }, | |
| { | |
| "epoch": 0.94, | |
| "eval_distillation_accuracy_counterfactual": 0.5330938943047717, | |
| "eval_distillation_accuracy_factual": 0.8201641867624423, | |
| "eval_distillation_f1_counterfactual": 0.4628830838776178, | |
| "eval_distillation_f1_factual": 0.8086604463285127, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5389943560800411, | |
| "eval_groundtruth_f1_counterfactual": 0.47066366125954884, | |
| "eval_groundtruth_f1_factual": 0.7039950616687438, | |
| "eval_icace_cosine": 0.4582160413265228, | |
| "eval_icace_l2": 0.6311160326004028, | |
| "eval_icace_normdiff": 0.3835700452327728, | |
| "eval_loss": 7.966909408569336, | |
| "eval_runtime": 12.9904, | |
| "eval_samples_per_second": 300.068, | |
| "eval_steps_per_second": 2.386, | |
| "step": 550 | |
| }, | |
| { | |
| "epoch": 0.96, | |
| "learning_rate": 4.393414211438475e-05, | |
| "loss": 8.8926, | |
| "step": 560 | |
| }, | |
| { | |
| "epoch": 0.99, | |
| "learning_rate": 4.371750433275563e-05, | |
| "loss": 8.704, | |
| "step": 580 | |
| }, | |
| { | |
| "epoch": 1.03, | |
| "learning_rate": 4.3500866551126516e-05, | |
| "loss": 8.5031, | |
| "step": 600 | |
| }, | |
| { | |
| "epoch": 1.03, | |
| "eval_distillation_accuracy_counterfactual": 0.513083632632119, | |
| "eval_distillation_accuracy_factual": 0.822986146741919, | |
| "eval_distillation_f1_counterfactual": 0.44551358939731, | |
| "eval_distillation_f1_factual": 0.8106574040803418, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5100051308363264, | |
| "eval_groundtruth_f1_counterfactual": 0.44647837744933117, | |
| "eval_groundtruth_f1_factual": 0.7133952359814255, | |
| "eval_icace_cosine": 0.46173933148384094, | |
| "eval_icace_l2": 0.6598149538040161, | |
| "eval_icace_normdiff": 0.4006192982196808, | |
| "eval_loss": 8.42484188079834, | |
| "eval_runtime": 12.9066, | |
| "eval_samples_per_second": 302.016, | |
| "eval_steps_per_second": 2.402, | |
| "step": 600 | |
| }, | |
| { | |
| "epoch": 1.06, | |
| "learning_rate": 4.3284228769497406e-05, | |
| "loss": 8.3235, | |
| "step": 620 | |
| }, | |
| { | |
| "epoch": 1.1, | |
| "learning_rate": 4.306759098786828e-05, | |
| "loss": 8.3831, | |
| "step": 640 | |
| }, | |
| { | |
| "epoch": 1.11, | |
| "eval_distillation_accuracy_counterfactual": 0.5289892252437147, | |
| "eval_distillation_accuracy_factual": 0.8260646485377117, | |
| "eval_distillation_f1_counterfactual": 0.47012361221009824, | |
| "eval_distillation_f1_factual": 0.8064982655163991, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5302719343252951, | |
| "eval_groundtruth_f1_counterfactual": 0.47179239129259665, | |
| "eval_groundtruth_f1_factual": 0.6996350113744447, | |
| "eval_icace_cosine": 0.4589817523956299, | |
| "eval_icace_l2": 0.6445746421813965, | |
| "eval_icace_normdiff": 0.39702358841896057, | |
| "eval_loss": 8.185053825378418, | |
| "eval_runtime": 12.9149, | |
| "eval_samples_per_second": 301.822, | |
| "eval_steps_per_second": 2.4, | |
| "step": 650 | |
| }, | |
| { | |
| "epoch": 1.13, | |
| "learning_rate": 4.285095320623917e-05, | |
| "loss": 8.5789, | |
| "step": 660 | |
| }, | |
| { | |
| "epoch": 1.17, | |
| "learning_rate": 4.2634315424610055e-05, | |
| "loss": 8.4277, | |
| "step": 680 | |
| }, | |
| { | |
| "epoch": 1.2, | |
| "learning_rate": 4.241767764298094e-05, | |
| "loss": 8.4591, | |
| "step": 700 | |
| }, | |
| { | |
| "epoch": 1.2, | |
| "eval_distillation_accuracy_counterfactual": 0.5071831708568497, | |
| "eval_distillation_accuracy_factual": 0.8291431503335044, | |
| "eval_distillation_f1_counterfactual": 0.4229278158258259, | |
| "eval_distillation_f1_factual": 0.8158920555245694, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5094920472036942, | |
| "eval_groundtruth_f1_counterfactual": 0.42468748967332876, | |
| "eval_groundtruth_f1_factual": 0.7058832006651468, | |
| "eval_icace_cosine": 0.46126317977905273, | |
| "eval_icace_l2": 0.6510809659957886, | |
| "eval_icace_normdiff": 0.39125320315361023, | |
| "eval_loss": 8.368897438049316, | |
| "eval_runtime": 12.9121, | |
| "eval_samples_per_second": 301.886, | |
| "eval_steps_per_second": 2.401, | |
| "step": 700 | |
| }, | |
| { | |
| "epoch": 1.23, | |
| "learning_rate": 4.220103986135182e-05, | |
| "loss": 8.4671, | |
| "step": 720 | |
| }, | |
| { | |
| "epoch": 1.27, | |
| "learning_rate": 4.198440207972271e-05, | |
| "loss": 8.4019, | |
| "step": 740 | |
| }, | |
| { | |
| "epoch": 1.29, | |
| "eval_distillation_accuracy_counterfactual": 0.49281682914315034, | |
| "eval_distillation_accuracy_factual": 0.827860441251924, | |
| "eval_distillation_f1_counterfactual": 0.41244951531506935, | |
| "eval_distillation_f1_factual": 0.8170673706163509, | |
| "eval_groundtruth_accuracy_counterfactual": 0.4856336582863007, | |
| "eval_groundtruth_f1_counterfactual": 0.4077764907535967, | |
| "eval_groundtruth_f1_factual": 0.7051511278749412, | |
| "eval_icace_cosine": 0.4673001766204834, | |
| "eval_icace_l2": 0.6646483540534973, | |
| "eval_icace_normdiff": 0.3972298800945282, | |
| "eval_loss": 8.459917068481445, | |
| "eval_runtime": 12.9198, | |
| "eval_samples_per_second": 301.708, | |
| "eval_steps_per_second": 2.399, | |
| "step": 750 | |
| }, | |
| { | |
| "epoch": 1.3, | |
| "learning_rate": 4.1767764298093586e-05, | |
| "loss": 8.5387, | |
| "step": 760 | |
| }, | |
| { | |
| "epoch": 1.34, | |
| "learning_rate": 4.1551126516464476e-05, | |
| "loss": 8.4688, | |
| "step": 780 | |
| }, | |
| { | |
| "epoch": 1.37, | |
| "learning_rate": 4.133448873483536e-05, | |
| "loss": 8.5516, | |
| "step": 800 | |
| }, | |
| { | |
| "epoch": 1.37, | |
| "eval_distillation_accuracy_counterfactual": 0.5184710107747563, | |
| "eval_distillation_accuracy_factual": 0.822986146741919, | |
| "eval_distillation_f1_counterfactual": 0.4457687437030219, | |
| "eval_distillation_f1_factual": 0.8022682230012104, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5159055926115957, | |
| "eval_groundtruth_f1_counterfactual": 0.4464137100072511, | |
| "eval_groundtruth_f1_factual": 0.6935766782924615, | |
| "eval_icace_cosine": 0.45302730798721313, | |
| "eval_icace_l2": 0.6743680238723755, | |
| "eval_icace_normdiff": 0.42486244440078735, | |
| "eval_loss": 8.359912872314453, | |
| "eval_runtime": 12.9096, | |
| "eval_samples_per_second": 301.945, | |
| "eval_steps_per_second": 2.401, | |
| "step": 800 | |
| }, | |
| { | |
| "epoch": 1.41, | |
| "learning_rate": 4.111785095320624e-05, | |
| "loss": 8.4687, | |
| "step": 820 | |
| }, | |
| { | |
| "epoch": 1.44, | |
| "learning_rate": 4.0901213171577124e-05, | |
| "loss": 8.3884, | |
| "step": 840 | |
| }, | |
| { | |
| "epoch": 1.46, | |
| "eval_distillation_accuracy_counterfactual": 0.530015392508979, | |
| "eval_distillation_accuracy_factual": 0.8209338122113905, | |
| "eval_distillation_f1_counterfactual": 0.44736802933677255, | |
| "eval_distillation_f1_factual": 0.8048461947798791, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5323242688558235, | |
| "eval_groundtruth_f1_counterfactual": 0.45097140064821445, | |
| "eval_groundtruth_f1_factual": 0.7159403464102823, | |
| "eval_icace_cosine": 0.4506547152996063, | |
| "eval_icace_l2": 0.6407148241996765, | |
| "eval_icace_normdiff": 0.395082026720047, | |
| "eval_loss": 8.082590103149414, | |
| "eval_runtime": 12.9483, | |
| "eval_samples_per_second": 301.044, | |
| "eval_steps_per_second": 2.394, | |
| "step": 850 | |
| }, | |
| { | |
| "epoch": 1.48, | |
| "learning_rate": 4.068457538994801e-05, | |
| "loss": 8.4295, | |
| "step": 860 | |
| }, | |
| { | |
| "epoch": 1.51, | |
| "learning_rate": 4.04679376083189e-05, | |
| "loss": 8.3828, | |
| "step": 880 | |
| }, | |
| { | |
| "epoch": 1.54, | |
| "learning_rate": 4.025129982668977e-05, | |
| "loss": 8.4004, | |
| "step": 900 | |
| }, | |
| { | |
| "epoch": 1.54, | |
| "eval_distillation_accuracy_counterfactual": 0.5220625962031811, | |
| "eval_distillation_accuracy_factual": 0.8083632632119035, | |
| "eval_distillation_f1_counterfactual": 0.44838757222370995, | |
| "eval_distillation_f1_factual": 0.7919448574347471, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5243714725500257, | |
| "eval_groundtruth_f1_counterfactual": 0.44872773811316585, | |
| "eval_groundtruth_f1_factual": 0.7046094122055937, | |
| "eval_icace_cosine": 0.4570869207382202, | |
| "eval_icace_l2": 0.6522126197814941, | |
| "eval_icace_normdiff": 0.406441867351532, | |
| "eval_loss": 8.266593933105469, | |
| "eval_runtime": 12.9504, | |
| "eval_samples_per_second": 300.995, | |
| "eval_steps_per_second": 2.394, | |
| "step": 900 | |
| }, | |
| { | |
| "epoch": 1.58, | |
| "learning_rate": 4.003466204506066e-05, | |
| "loss": 8.335, | |
| "step": 920 | |
| }, | |
| { | |
| "epoch": 1.61, | |
| "learning_rate": 3.9818024263431546e-05, | |
| "loss": 8.3741, | |
| "step": 940 | |
| }, | |
| { | |
| "epoch": 1.63, | |
| "eval_distillation_accuracy_counterfactual": 0.5312981015905592, | |
| "eval_distillation_accuracy_factual": 0.8055413032324269, | |
| "eval_distillation_f1_counterfactual": 0.4617410563594155, | |
| "eval_distillation_f1_factual": 0.7931325749560401, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5194971780400205, | |
| "eval_groundtruth_f1_counterfactual": 0.4513515695315669, | |
| "eval_groundtruth_f1_factual": 0.7019030914005773, | |
| "eval_icace_cosine": 0.45170876383781433, | |
| "eval_icace_l2": 0.6569298505783081, | |
| "eval_icace_normdiff": 0.41902977228164673, | |
| "eval_loss": 8.250597953796387, | |
| "eval_runtime": 12.89, | |
| "eval_samples_per_second": 302.405, | |
| "eval_steps_per_second": 2.405, | |
| "step": 950 | |
| }, | |
| { | |
| "epoch": 1.65, | |
| "learning_rate": 3.960138648180243e-05, | |
| "loss": 8.4944, | |
| "step": 960 | |
| }, | |
| { | |
| "epoch": 1.68, | |
| "learning_rate": 3.938474870017331e-05, | |
| "loss": 8.4011, | |
| "step": 980 | |
| }, | |
| { | |
| "epoch": 1.72, | |
| "learning_rate": 3.91681109185442e-05, | |
| "loss": 8.4287, | |
| "step": 1000 | |
| }, | |
| { | |
| "epoch": 1.72, | |
| "eval_distillation_accuracy_counterfactual": 0.5371985633658286, | |
| "eval_distillation_accuracy_factual": 0.7909184197024115, | |
| "eval_distillation_f1_counterfactual": 0.46983763588777594, | |
| "eval_distillation_f1_factual": 0.7799708268449532, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5318111852231914, | |
| "eval_groundtruth_f1_counterfactual": 0.4659493577740304, | |
| "eval_groundtruth_f1_factual": 0.681861236036038, | |
| "eval_icace_cosine": 0.4572613537311554, | |
| "eval_icace_l2": 0.6384456753730774, | |
| "eval_icace_normdiff": 0.3894091844558716, | |
| "eval_loss": 8.310347557067871, | |
| "eval_runtime": 12.9611, | |
| "eval_samples_per_second": 300.747, | |
| "eval_steps_per_second": 2.392, | |
| "step": 1000 | |
| }, | |
| { | |
| "epoch": 1.75, | |
| "learning_rate": 3.895147313691508e-05, | |
| "loss": 8.2598, | |
| "step": 1020 | |
| }, | |
| { | |
| "epoch": 1.78, | |
| "learning_rate": 3.873483535528597e-05, | |
| "loss": 8.2625, | |
| "step": 1040 | |
| }, | |
| { | |
| "epoch": 1.8, | |
| "eval_distillation_accuracy_counterfactual": 0.5397639815289892, | |
| "eval_distillation_accuracy_factual": 0.7875833760903027, | |
| "eval_distillation_f1_counterfactual": 0.4799542696533782, | |
| "eval_distillation_f1_factual": 0.7726528869168954, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5369420215495125, | |
| "eval_groundtruth_f1_counterfactual": 0.4771898703044782, | |
| "eval_groundtruth_f1_factual": 0.7023731499922957, | |
| "eval_icace_cosine": 0.45770639181137085, | |
| "eval_icace_l2": 0.6276518702507019, | |
| "eval_icace_normdiff": 0.3755152225494385, | |
| "eval_loss": 8.230777740478516, | |
| "eval_runtime": 12.9942, | |
| "eval_samples_per_second": 299.981, | |
| "eval_steps_per_second": 2.386, | |
| "step": 1050 | |
| }, | |
| { | |
| "epoch": 1.82, | |
| "learning_rate": 3.851819757365685e-05, | |
| "loss": 8.0979, | |
| "step": 1060 | |
| }, | |
| { | |
| "epoch": 1.85, | |
| "learning_rate": 3.830155979202773e-05, | |
| "loss": 8.2971, | |
| "step": 1080 | |
| }, | |
| { | |
| "epoch": 1.89, | |
| "learning_rate": 3.8084922010398616e-05, | |
| "loss": 8.343, | |
| "step": 1100 | |
| }, | |
| { | |
| "epoch": 1.89, | |
| "eval_distillation_accuracy_counterfactual": 0.5415597742432017, | |
| "eval_distillation_accuracy_factual": 0.806054386865059, | |
| "eval_distillation_f1_counterfactual": 0.4594314006577007, | |
| "eval_distillation_f1_factual": 0.7921003157959313, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5400205233453053, | |
| "eval_groundtruth_f1_counterfactual": 0.45903598954343383, | |
| "eval_groundtruth_f1_factual": 0.6999320297801038, | |
| "eval_icace_cosine": 0.45384570956230164, | |
| "eval_icace_l2": 0.6284541487693787, | |
| "eval_icace_normdiff": 0.3843493163585663, | |
| "eval_loss": 8.147943496704102, | |
| "eval_runtime": 13.1714, | |
| "eval_samples_per_second": 295.944, | |
| "eval_steps_per_second": 2.354, | |
| "step": 1100 | |
| }, | |
| { | |
| "epoch": 1.92, | |
| "learning_rate": 3.78682842287695e-05, | |
| "loss": 8.4675, | |
| "step": 1120 | |
| }, | |
| { | |
| "epoch": 1.96, | |
| "learning_rate": 3.765164644714038e-05, | |
| "loss": 8.3651, | |
| "step": 1140 | |
| }, | |
| { | |
| "epoch": 1.97, | |
| "eval_distillation_accuracy_counterfactual": 0.5330938943047717, | |
| "eval_distillation_accuracy_factual": 0.8273473576192919, | |
| "eval_distillation_f1_counterfactual": 0.4438315486422179, | |
| "eval_distillation_f1_factual": 0.8114092211560717, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5282195997947665, | |
| "eval_groundtruth_f1_counterfactual": 0.43912524130185177, | |
| "eval_groundtruth_f1_factual": 0.696958037051927, | |
| "eval_icace_cosine": 0.45857784152030945, | |
| "eval_icace_l2": 0.6330906748771667, | |
| "eval_icace_normdiff": 0.3742936849594116, | |
| "eval_loss": 8.261628150939941, | |
| "eval_runtime": 12.8921, | |
| "eval_samples_per_second": 302.355, | |
| "eval_steps_per_second": 2.405, | |
| "step": 1150 | |
| }, | |
| { | |
| "epoch": 1.99, | |
| "learning_rate": 3.7435008665511264e-05, | |
| "loss": 8.3106, | |
| "step": 1160 | |
| }, | |
| { | |
| "epoch": 2.02, | |
| "learning_rate": 3.7218370883882154e-05, | |
| "loss": 8.1733, | |
| "step": 1180 | |
| }, | |
| { | |
| "epoch": 2.06, | |
| "learning_rate": 3.700173310225303e-05, | |
| "loss": 8.1566, | |
| "step": 1200 | |
| }, | |
| { | |
| "epoch": 2.06, | |
| "eval_distillation_accuracy_counterfactual": 0.521806054386865, | |
| "eval_distillation_accuracy_factual": 0.8083632632119035, | |
| "eval_distillation_f1_counterfactual": 0.4568626686367353, | |
| "eval_distillation_f1_factual": 0.793454263258986, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5210364289379169, | |
| "eval_groundtruth_f1_counterfactual": 0.4589985732848542, | |
| "eval_groundtruth_f1_factual": 0.6909411446988759, | |
| "eval_icace_cosine": 0.4588005244731903, | |
| "eval_icace_l2": 0.6523511409759521, | |
| "eval_icace_normdiff": 0.40103352069854736, | |
| "eval_loss": 8.481898307800293, | |
| "eval_runtime": 12.9624, | |
| "eval_samples_per_second": 300.717, | |
| "eval_steps_per_second": 2.392, | |
| "step": 1200 | |
| }, | |
| { | |
| "epoch": 2.09, | |
| "learning_rate": 3.678509532062392e-05, | |
| "loss": 8.0229, | |
| "step": 1220 | |
| }, | |
| { | |
| "epoch": 2.13, | |
| "learning_rate": 3.65684575389948e-05, | |
| "loss": 8.3726, | |
| "step": 1240 | |
| }, | |
| { | |
| "epoch": 2.14, | |
| "eval_distillation_accuracy_counterfactual": 0.5246280143663418, | |
| "eval_distillation_accuracy_factual": 0.8040020523345305, | |
| "eval_distillation_f1_counterfactual": 0.4638517088409774, | |
| "eval_distillation_f1_factual": 0.7892649860700691, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5166752180605438, | |
| "eval_groundtruth_f1_counterfactual": 0.4574432564220393, | |
| "eval_groundtruth_f1_factual": 0.7129214655450473, | |
| "eval_icace_cosine": 0.45950308442115784, | |
| "eval_icace_l2": 0.6577529311180115, | |
| "eval_icace_normdiff": 0.4075782299041748, | |
| "eval_loss": 8.51949691772461, | |
| "eval_runtime": 12.9172, | |
| "eval_samples_per_second": 301.769, | |
| "eval_steps_per_second": 2.4, | |
| "step": 1250 | |
| }, | |
| { | |
| "epoch": 2.16, | |
| "learning_rate": 3.6351819757365686e-05, | |
| "loss": 8.2033, | |
| "step": 1260 | |
| }, | |
| { | |
| "epoch": 2.2, | |
| "learning_rate": 3.613518197573657e-05, | |
| "loss": 8.2507, | |
| "step": 1280 | |
| }, | |
| { | |
| "epoch": 2.23, | |
| "learning_rate": 3.591854419410746e-05, | |
| "loss": 8.0753, | |
| "step": 1300 | |
| }, | |
| { | |
| "epoch": 2.23, | |
| "eval_distillation_accuracy_counterfactual": 0.5397639815289892, | |
| "eval_distillation_accuracy_factual": 0.8165726013340174, | |
| "eval_distillation_f1_counterfactual": 0.47682727590749313, | |
| "eval_distillation_f1_factual": 0.7967308018931318, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5325808106721396, | |
| "eval_groundtruth_f1_counterfactual": 0.47018455613210597, | |
| "eval_groundtruth_f1_factual": 0.7043239542576792, | |
| "eval_icace_cosine": 0.4495893120765686, | |
| "eval_icace_l2": 0.6213250160217285, | |
| "eval_icace_normdiff": 0.37207138538360596, | |
| "eval_loss": 8.11995792388916, | |
| "eval_runtime": 13.9585, | |
| "eval_samples_per_second": 279.256, | |
| "eval_steps_per_second": 2.221, | |
| "step": 1300 | |
| }, | |
| { | |
| "epoch": 2.26, | |
| "learning_rate": 3.5701906412478334e-05, | |
| "loss": 8.2577, | |
| "step": 1320 | |
| }, | |
| { | |
| "epoch": 2.3, | |
| "learning_rate": 3.5485268630849224e-05, | |
| "loss": 8.3416, | |
| "step": 1340 | |
| }, | |
| { | |
| "epoch": 2.32, | |
| "eval_distillation_accuracy_counterfactual": 0.5246280143663418, | |
| "eval_distillation_accuracy_factual": 0.8183683940482298, | |
| "eval_distillation_f1_counterfactual": 0.4426092418935421, | |
| "eval_distillation_f1_factual": 0.8016386570803451, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5066700872242176, | |
| "eval_groundtruth_f1_counterfactual": 0.43023988421850207, | |
| "eval_groundtruth_f1_factual": 0.7132188999952972, | |
| "eval_icace_cosine": 0.45355769991874695, | |
| "eval_icace_l2": 0.6458914279937744, | |
| "eval_icace_normdiff": 0.3992498517036438, | |
| "eval_loss": 8.449216842651367, | |
| "eval_runtime": 12.8687, | |
| "eval_samples_per_second": 302.906, | |
| "eval_steps_per_second": 2.409, | |
| "step": 1350 | |
| }, | |
| { | |
| "epoch": 2.33, | |
| "learning_rate": 3.526863084922011e-05, | |
| "loss": 8.2811, | |
| "step": 1360 | |
| }, | |
| { | |
| "epoch": 2.37, | |
| "learning_rate": 3.505199306759099e-05, | |
| "loss": 8.192, | |
| "step": 1380 | |
| }, | |
| { | |
| "epoch": 2.4, | |
| "learning_rate": 3.483535528596187e-05, | |
| "loss": 8.1159, | |
| "step": 1400 | |
| }, | |
| { | |
| "epoch": 2.4, | |
| "eval_distillation_accuracy_counterfactual": 0.5348896870189841, | |
| "eval_distillation_accuracy_factual": 0.8070805541303232, | |
| "eval_distillation_f1_counterfactual": 0.4442360778869393, | |
| "eval_distillation_f1_factual": 0.793089455272879, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5389943560800411, | |
| "eval_groundtruth_f1_counterfactual": 0.45311551419493207, | |
| "eval_groundtruth_f1_factual": 0.713127845360182, | |
| "eval_icace_cosine": 0.45396122336387634, | |
| "eval_icace_l2": 0.6456016302108765, | |
| "eval_icace_normdiff": 0.397720605134964, | |
| "eval_loss": 8.26937198638916, | |
| "eval_runtime": 13.0046, | |
| "eval_samples_per_second": 299.74, | |
| "eval_steps_per_second": 2.384, | |
| "step": 1400 | |
| }, | |
| { | |
| "epoch": 2.44, | |
| "learning_rate": 3.461871750433276e-05, | |
| "loss": 8.1856, | |
| "step": 1420 | |
| }, | |
| { | |
| "epoch": 2.47, | |
| "learning_rate": 3.440207972270364e-05, | |
| "loss": 8.2343, | |
| "step": 1440 | |
| }, | |
| { | |
| "epoch": 2.49, | |
| "eval_distillation_accuracy_counterfactual": 0.5105182144689584, | |
| "eval_distillation_accuracy_factual": 0.8150333504361211, | |
| "eval_distillation_f1_counterfactual": 0.46519275113167036, | |
| "eval_distillation_f1_factual": 0.7986983661724563, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5051308363263212, | |
| "eval_groundtruth_f1_counterfactual": 0.4604440377605913, | |
| "eval_groundtruth_f1_factual": 0.6976730796070983, | |
| "eval_icace_cosine": 0.44814541935920715, | |
| "eval_icace_l2": 0.6517916917800903, | |
| "eval_icace_normdiff": 0.4058307409286499, | |
| "eval_loss": 8.407746315002441, | |
| "eval_runtime": 12.9229, | |
| "eval_samples_per_second": 301.635, | |
| "eval_steps_per_second": 2.399, | |
| "step": 1450 | |
| } | |
| ], | |
| "max_steps": 4616, | |
| "num_train_epochs": 8, | |
| "total_flos": 1.2202103856768e+16, | |
| "trial_name": null, | |
| "trial_params": null | |
| } | |