| { | |
| "best_metric": 0.45797210931777954, | |
| "best_model_checkpoint": "model_output/e2e_opentable_5_way__approximate__0-shot__seed-77__bert-base-uncased/checkpoint-1000", | |
| "epoch": 1.7152658662092626, | |
| "global_step": 1000, | |
| "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": 15.0971, | |
| "step": 20 | |
| }, | |
| { | |
| "epoch": 0.07, | |
| "learning_rate": 4.956672443674177e-05, | |
| "loss": 10.8654, | |
| "step": 40 | |
| }, | |
| { | |
| "epoch": 0.09, | |
| "eval_distillation_accuracy_counterfactual": 0.512827090815803, | |
| "eval_distillation_accuracy_factual": 0.8973832734735762, | |
| "eval_distillation_f1_counterfactual": 0.38283775595303926, | |
| "eval_distillation_f1_factual": 0.8849844553565298, | |
| "eval_groundtruth_accuracy_counterfactual": 0.47998973832734737, | |
| "eval_groundtruth_f1_counterfactual": 0.35799290350733415, | |
| "eval_groundtruth_f1_factual": 0.6974663941237289, | |
| "eval_icace_cosine": 0.4845265746116638, | |
| "eval_icace_l2": 0.6881101131439209, | |
| "eval_icace_normdiff": 0.4512614905834198, | |
| "eval_loss": 7.50442361831665, | |
| "eval_runtime": 22.2098, | |
| "eval_samples_per_second": 175.508, | |
| "eval_steps_per_second": 2.747, | |
| "step": 50 | |
| }, | |
| { | |
| "epoch": 0.1, | |
| "learning_rate": 4.935008665511265e-05, | |
| "loss": 10.0023, | |
| "step": 60 | |
| }, | |
| { | |
| "epoch": 0.14, | |
| "learning_rate": 4.913344887348354e-05, | |
| "loss": 9.8393, | |
| "step": 80 | |
| }, | |
| { | |
| "epoch": 0.17, | |
| "learning_rate": 4.891681109185442e-05, | |
| "loss": 9.2084, | |
| "step": 100 | |
| }, | |
| { | |
| "epoch": 0.17, | |
| "eval_distillation_accuracy_counterfactual": 0.534376603386352, | |
| "eval_distillation_accuracy_factual": 0.879168804515136, | |
| "eval_distillation_f1_counterfactual": 0.47020198824723397, | |
| "eval_distillation_f1_factual": 0.8636644497351712, | |
| "eval_groundtruth_accuracy_counterfactual": 0.521806054386865, | |
| "eval_groundtruth_f1_counterfactual": 0.46415698862260973, | |
| "eval_groundtruth_f1_factual": 0.6787169266269484, | |
| "eval_icace_cosine": 0.474692165851593, | |
| "eval_icace_l2": 0.6402959823608398, | |
| "eval_icace_normdiff": 0.41231343150138855, | |
| "eval_loss": 7.25449275970459, | |
| "eval_runtime": 22.2189, | |
| "eval_samples_per_second": 175.436, | |
| "eval_steps_per_second": 2.745, | |
| "step": 100 | |
| }, | |
| { | |
| "epoch": 0.21, | |
| "learning_rate": 4.8700173310225307e-05, | |
| "loss": 9.2547, | |
| "step": 120 | |
| }, | |
| { | |
| "epoch": 0.24, | |
| "learning_rate": 4.848353552859619e-05, | |
| "loss": 9.2275, | |
| "step": 140 | |
| }, | |
| { | |
| "epoch": 0.26, | |
| "eval_distillation_accuracy_counterfactual": 0.5028219599794767, | |
| "eval_distillation_accuracy_factual": 0.8245253976398152, | |
| "eval_distillation_f1_counterfactual": 0.41359201224181497, | |
| "eval_distillation_f1_factual": 0.8013559665708911, | |
| "eval_groundtruth_accuracy_counterfactual": 0.49204720369420213, | |
| "eval_groundtruth_f1_counterfactual": 0.4109951641517333, | |
| "eval_groundtruth_f1_factual": 0.6495764505483296, | |
| "eval_icace_cosine": 0.46765828132629395, | |
| "eval_icace_l2": 0.6432615518569946, | |
| "eval_icace_normdiff": 0.38947969675064087, | |
| "eval_loss": 7.776533126831055, | |
| "eval_runtime": 22.1488, | |
| "eval_samples_per_second": 175.991, | |
| "eval_steps_per_second": 2.754, | |
| "step": 150 | |
| }, | |
| { | |
| "epoch": 0.27, | |
| "learning_rate": 4.826689774696707e-05, | |
| "loss": 8.9152, | |
| "step": 160 | |
| }, | |
| { | |
| "epoch": 0.31, | |
| "learning_rate": 4.8050259965337955e-05, | |
| "loss": 8.7736, | |
| "step": 180 | |
| }, | |
| { | |
| "epoch": 0.34, | |
| "learning_rate": 4.7833622183708845e-05, | |
| "loss": 8.8387, | |
| "step": 200 | |
| }, | |
| { | |
| "epoch": 0.34, | |
| "eval_distillation_accuracy_counterfactual": 0.5177013853258081, | |
| "eval_distillation_accuracy_factual": 0.8614674191893279, | |
| "eval_distillation_f1_counterfactual": 0.41268689402882053, | |
| "eval_distillation_f1_factual": 0.8521590725130646, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5030785017957927, | |
| "eval_groundtruth_f1_counterfactual": 0.40850979925255315, | |
| "eval_groundtruth_f1_factual": 0.6669518573115495, | |
| "eval_icace_cosine": 0.4691610634326935, | |
| "eval_icace_l2": 0.6340324282646179, | |
| "eval_icace_normdiff": 0.39094486832618713, | |
| "eval_loss": 7.458261013031006, | |
| "eval_runtime": 22.3381, | |
| "eval_samples_per_second": 174.5, | |
| "eval_steps_per_second": 2.731, | |
| "step": 200 | |
| }, | |
| { | |
| "epoch": 0.38, | |
| "learning_rate": 4.761698440207972e-05, | |
| "loss": 8.772, | |
| "step": 220 | |
| }, | |
| { | |
| "epoch": 0.41, | |
| "learning_rate": 4.740034662045061e-05, | |
| "loss": 8.7734, | |
| "step": 240 | |
| }, | |
| { | |
| "epoch": 0.43, | |
| "eval_distillation_accuracy_counterfactual": 0.4671626475115444, | |
| "eval_distillation_accuracy_factual": 0.8671113391482812, | |
| "eval_distillation_f1_counterfactual": 0.38582609157213904, | |
| "eval_distillation_f1_factual": 0.8563293543368673, | |
| "eval_groundtruth_accuracy_counterfactual": 0.46151872755259105, | |
| "eval_groundtruth_f1_counterfactual": 0.39182214143822086, | |
| "eval_groundtruth_f1_factual": 0.6883791067687154, | |
| "eval_icace_cosine": 0.4946799874305725, | |
| "eval_icace_l2": 0.6939188241958618, | |
| "eval_icace_normdiff": 0.43225187063217163, | |
| "eval_loss": 8.251187324523926, | |
| "eval_runtime": 22.1682, | |
| "eval_samples_per_second": 175.838, | |
| "eval_steps_per_second": 2.752, | |
| "step": 250 | |
| }, | |
| { | |
| "epoch": 0.45, | |
| "learning_rate": 4.7183708838821494e-05, | |
| "loss": 8.5326, | |
| "step": 260 | |
| }, | |
| { | |
| "epoch": 0.48, | |
| "learning_rate": 4.6967071057192376e-05, | |
| "loss": 8.593, | |
| "step": 280 | |
| }, | |
| { | |
| "epoch": 0.51, | |
| "learning_rate": 4.675043327556326e-05, | |
| "loss": 8.5207, | |
| "step": 300 | |
| }, | |
| { | |
| "epoch": 0.51, | |
| "eval_distillation_accuracy_counterfactual": 0.5187275525910724, | |
| "eval_distillation_accuracy_factual": 0.8471010774756286, | |
| "eval_distillation_f1_counterfactual": 0.4809461537601309, | |
| "eval_distillation_f1_factual": 0.8415144400606737, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5023088763468445, | |
| "eval_groundtruth_f1_counterfactual": 0.46717757401427457, | |
| "eval_groundtruth_f1_factual": 0.6817982025512722, | |
| "eval_icace_cosine": 0.48118895292282104, | |
| "eval_icace_l2": 0.6521334052085876, | |
| "eval_icace_normdiff": 0.4162002503871918, | |
| "eval_loss": 7.6342902183532715, | |
| "eval_runtime": 22.1084, | |
| "eval_samples_per_second": 176.313, | |
| "eval_steps_per_second": 2.759, | |
| "step": 300 | |
| }, | |
| { | |
| "epoch": 0.55, | |
| "learning_rate": 4.653379549393415e-05, | |
| "loss": 8.3762, | |
| "step": 320 | |
| }, | |
| { | |
| "epoch": 0.58, | |
| "learning_rate": 4.6317157712305025e-05, | |
| "loss": 8.3251, | |
| "step": 340 | |
| }, | |
| { | |
| "epoch": 0.6, | |
| "eval_distillation_accuracy_counterfactual": 0.5295023088763469, | |
| "eval_distillation_accuracy_factual": 0.8568496664956388, | |
| "eval_distillation_f1_counterfactual": 0.4641469168961435, | |
| "eval_distillation_f1_factual": 0.8494977579970042, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5166752180605438, | |
| "eval_groundtruth_f1_counterfactual": 0.4567649352125045, | |
| "eval_groundtruth_f1_factual": 0.6694324318001126, | |
| "eval_icace_cosine": 0.46943578124046326, | |
| "eval_icace_l2": 0.6349437236785889, | |
| "eval_icace_normdiff": 0.39818716049194336, | |
| "eval_loss": 7.341701507568359, | |
| "eval_runtime": 22.1905, | |
| "eval_samples_per_second": 175.66, | |
| "eval_steps_per_second": 2.749, | |
| "step": 350 | |
| }, | |
| { | |
| "epoch": 0.62, | |
| "learning_rate": 4.6100519930675915e-05, | |
| "loss": 8.2535, | |
| "step": 360 | |
| }, | |
| { | |
| "epoch": 0.65, | |
| "learning_rate": 4.58838821490468e-05, | |
| "loss": 8.3502, | |
| "step": 380 | |
| }, | |
| { | |
| "epoch": 0.69, | |
| "learning_rate": 4.566724436741768e-05, | |
| "loss": 8.1628, | |
| "step": 400 | |
| }, | |
| { | |
| "epoch": 0.69, | |
| "eval_distillation_accuracy_counterfactual": 0.5397639815289892, | |
| "eval_distillation_accuracy_factual": 0.8355566957414059, | |
| "eval_distillation_f1_counterfactual": 0.46033001340715457, | |
| "eval_distillation_f1_factual": 0.8292453735319636, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5223191380194971, | |
| "eval_groundtruth_f1_counterfactual": 0.44829969536055286, | |
| "eval_groundtruth_f1_factual": 0.6718756022768833, | |
| "eval_icace_cosine": 0.460631787776947, | |
| "eval_icace_l2": 0.6126735210418701, | |
| "eval_icace_normdiff": 0.3810984492301941, | |
| "eval_loss": 7.215612411499023, | |
| "eval_runtime": 22.2088, | |
| "eval_samples_per_second": 175.516, | |
| "eval_steps_per_second": 2.747, | |
| "step": 400 | |
| }, | |
| { | |
| "epoch": 0.72, | |
| "learning_rate": 4.5450606585788563e-05, | |
| "loss": 8.1842, | |
| "step": 420 | |
| }, | |
| { | |
| "epoch": 0.75, | |
| "learning_rate": 4.5233968804159446e-05, | |
| "loss": 8.2209, | |
| "step": 440 | |
| }, | |
| { | |
| "epoch": 0.77, | |
| "eval_distillation_accuracy_counterfactual": 0.5256541816316059, | |
| "eval_distillation_accuracy_factual": 0.8045151359671626, | |
| "eval_distillation_f1_counterfactual": 0.41138103040788077, | |
| "eval_distillation_f1_factual": 0.798165576043234, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5248845561826577, | |
| "eval_groundtruth_f1_counterfactual": 0.4184296633356827, | |
| "eval_groundtruth_f1_factual": 0.66434863984353, | |
| "eval_icace_cosine": 0.47930216789245605, | |
| "eval_icace_l2": 0.6362362504005432, | |
| "eval_icace_normdiff": 0.4014618396759033, | |
| "eval_loss": 7.654078483581543, | |
| "eval_runtime": 22.2042, | |
| "eval_samples_per_second": 175.552, | |
| "eval_steps_per_second": 2.747, | |
| "step": 450 | |
| }, | |
| { | |
| "epoch": 0.79, | |
| "learning_rate": 4.501733102253033e-05, | |
| "loss": 8.2117, | |
| "step": 460 | |
| }, | |
| { | |
| "epoch": 0.82, | |
| "learning_rate": 4.480069324090121e-05, | |
| "loss": 8.246, | |
| "step": 480 | |
| }, | |
| { | |
| "epoch": 0.86, | |
| "learning_rate": 4.45840554592721e-05, | |
| "loss": 8.1357, | |
| "step": 500 | |
| }, | |
| { | |
| "epoch": 0.86, | |
| "eval_distillation_accuracy_counterfactual": 0.5497691123653156, | |
| "eval_distillation_accuracy_factual": 0.8237557721908672, | |
| "eval_distillation_f1_counterfactual": 0.45037064661810805, | |
| "eval_distillation_f1_factual": 0.8056668211328342, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5402770651616213, | |
| "eval_groundtruth_f1_counterfactual": 0.44977490137963505, | |
| "eval_groundtruth_f1_factual": 0.6736698728564816, | |
| "eval_icace_cosine": 0.4668065309524536, | |
| "eval_icace_l2": 0.6126391887664795, | |
| "eval_icace_normdiff": 0.39072343707084656, | |
| "eval_loss": 7.208556175231934, | |
| "eval_runtime": 22.0856, | |
| "eval_samples_per_second": 176.495, | |
| "eval_steps_per_second": 2.762, | |
| "step": 500 | |
| }, | |
| { | |
| "epoch": 0.89, | |
| "learning_rate": 4.436741767764298e-05, | |
| "loss": 8.0972, | |
| "step": 520 | |
| }, | |
| { | |
| "epoch": 0.93, | |
| "learning_rate": 4.415077989601387e-05, | |
| "loss": 7.9764, | |
| "step": 540 | |
| }, | |
| { | |
| "epoch": 0.94, | |
| "eval_distillation_accuracy_counterfactual": 0.5413032324268856, | |
| "eval_distillation_accuracy_factual": 0.8263211903540277, | |
| "eval_distillation_f1_counterfactual": 0.4579261774469936, | |
| "eval_distillation_f1_factual": 0.8169241589118235, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5338635197537198, | |
| "eval_groundtruth_f1_counterfactual": 0.45648415392539654, | |
| "eval_groundtruth_f1_factual": 0.6666874640710526, | |
| "eval_icace_cosine": 0.4673744738101959, | |
| "eval_icace_l2": 0.6225017309188843, | |
| "eval_icace_normdiff": 0.3859357535839081, | |
| "eval_loss": 7.50295877456665, | |
| "eval_runtime": 22.0967, | |
| "eval_samples_per_second": 176.406, | |
| "eval_steps_per_second": 2.761, | |
| "step": 550 | |
| }, | |
| { | |
| "epoch": 0.96, | |
| "learning_rate": 4.393414211438475e-05, | |
| "loss": 8.0277, | |
| "step": 560 | |
| }, | |
| { | |
| "epoch": 0.99, | |
| "learning_rate": 4.371750433275563e-05, | |
| "loss": 7.9231, | |
| "step": 580 | |
| }, | |
| { | |
| "epoch": 1.03, | |
| "learning_rate": 4.3500866551126516e-05, | |
| "loss": 8.1605, | |
| "step": 600 | |
| }, | |
| { | |
| "epoch": 1.03, | |
| "eval_distillation_accuracy_counterfactual": 0.538737814263725, | |
| "eval_distillation_accuracy_factual": 0.8424833247819394, | |
| "eval_distillation_f1_counterfactual": 0.4424763530876773, | |
| "eval_distillation_f1_factual": 0.8305147218419746, | |
| "eval_groundtruth_accuracy_counterfactual": 0.530528476141611, | |
| "eval_groundtruth_f1_counterfactual": 0.44459191212630245, | |
| "eval_groundtruth_f1_factual": 0.675893177241911, | |
| "eval_icace_cosine": 0.47036415338516235, | |
| "eval_icace_l2": 0.6303676962852478, | |
| "eval_icace_normdiff": 0.40152838826179504, | |
| "eval_loss": 7.446086406707764, | |
| "eval_runtime": 22.193, | |
| "eval_samples_per_second": 175.641, | |
| "eval_steps_per_second": 2.749, | |
| "step": 600 | |
| }, | |
| { | |
| "epoch": 1.06, | |
| "learning_rate": 4.3284228769497406e-05, | |
| "loss": 8.0369, | |
| "step": 620 | |
| }, | |
| { | |
| "epoch": 1.1, | |
| "learning_rate": 4.306759098786828e-05, | |
| "loss": 7.8522, | |
| "step": 640 | |
| }, | |
| { | |
| "epoch": 1.11, | |
| "eval_distillation_accuracy_counterfactual": 0.5325808106721396, | |
| "eval_distillation_accuracy_factual": 0.7945100051308364, | |
| "eval_distillation_f1_counterfactual": 0.45243712925320734, | |
| "eval_distillation_f1_factual": 0.7760743993737376, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5248845561826577, | |
| "eval_groundtruth_f1_counterfactual": 0.4539872598778271, | |
| "eval_groundtruth_f1_factual": 0.6485178712630223, | |
| "eval_icace_cosine": 0.4656234085559845, | |
| "eval_icace_l2": 0.6243467330932617, | |
| "eval_icace_normdiff": 0.38494807481765747, | |
| "eval_loss": 7.678111553192139, | |
| "eval_runtime": 22.2461, | |
| "eval_samples_per_second": 175.222, | |
| "eval_steps_per_second": 2.742, | |
| "step": 650 | |
| }, | |
| { | |
| "epoch": 1.13, | |
| "learning_rate": 4.285095320623917e-05, | |
| "loss": 7.7876, | |
| "step": 660 | |
| }, | |
| { | |
| "epoch": 1.17, | |
| "learning_rate": 4.2634315424610055e-05, | |
| "loss": 7.8585, | |
| "step": 680 | |
| }, | |
| { | |
| "epoch": 1.2, | |
| "learning_rate": 4.241767764298094e-05, | |
| "loss": 7.9963, | |
| "step": 700 | |
| }, | |
| { | |
| "epoch": 1.2, | |
| "eval_distillation_accuracy_counterfactual": 0.5220625962031811, | |
| "eval_distillation_accuracy_factual": 0.8532580810672139, | |
| "eval_distillation_f1_counterfactual": 0.421394437601727, | |
| "eval_distillation_f1_factual": 0.843810107668179, | |
| "eval_groundtruth_accuracy_counterfactual": 0.513083632632119, | |
| "eval_groundtruth_f1_counterfactual": 0.41873912487772846, | |
| "eval_groundtruth_f1_factual": 0.6890304825574103, | |
| "eval_icace_cosine": 0.47455719113349915, | |
| "eval_icace_l2": 0.6358718276023865, | |
| "eval_icace_normdiff": 0.3861025869846344, | |
| "eval_loss": 7.682065010070801, | |
| "eval_runtime": 22.4604, | |
| "eval_samples_per_second": 173.55, | |
| "eval_steps_per_second": 2.716, | |
| "step": 700 | |
| }, | |
| { | |
| "epoch": 1.23, | |
| "learning_rate": 4.220103986135182e-05, | |
| "loss": 7.9938, | |
| "step": 720 | |
| }, | |
| { | |
| "epoch": 1.27, | |
| "learning_rate": 4.198440207972271e-05, | |
| "loss": 7.7139, | |
| "step": 740 | |
| }, | |
| { | |
| "epoch": 1.29, | |
| "eval_distillation_accuracy_counterfactual": 0.5405336069779374, | |
| "eval_distillation_accuracy_factual": 0.8509492047203694, | |
| "eval_distillation_f1_counterfactual": 0.47182248378312613, | |
| "eval_distillation_f1_factual": 0.8355752615379481, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5269368907131863, | |
| "eval_groundtruth_f1_counterfactual": 0.4620646268597837, | |
| "eval_groundtruth_f1_factual": 0.6925217992694058, | |
| "eval_icace_cosine": 0.45927247405052185, | |
| "eval_icace_l2": 0.6215088367462158, | |
| "eval_icace_normdiff": 0.39000242948532104, | |
| "eval_loss": 7.237956523895264, | |
| "eval_runtime": 22.1166, | |
| "eval_samples_per_second": 176.248, | |
| "eval_steps_per_second": 2.758, | |
| "step": 750 | |
| }, | |
| { | |
| "epoch": 1.3, | |
| "learning_rate": 4.1767764298093586e-05, | |
| "loss": 7.8577, | |
| "step": 760 | |
| }, | |
| { | |
| "epoch": 1.34, | |
| "learning_rate": 4.1551126516464476e-05, | |
| "loss": 7.8288, | |
| "step": 780 | |
| }, | |
| { | |
| "epoch": 1.37, | |
| "learning_rate": 4.133448873483536e-05, | |
| "loss": 7.6519, | |
| "step": 800 | |
| }, | |
| { | |
| "epoch": 1.37, | |
| "eval_distillation_accuracy_counterfactual": 0.5279630579784504, | |
| "eval_distillation_accuracy_factual": 0.805797845048743, | |
| "eval_distillation_f1_counterfactual": 0.44398440406113127, | |
| "eval_distillation_f1_factual": 0.7889017139693454, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5207798871216008, | |
| "eval_groundtruth_f1_counterfactual": 0.4413428310825916, | |
| "eval_groundtruth_f1_factual": 0.6594914555824491, | |
| "eval_icace_cosine": 0.47146424651145935, | |
| "eval_icace_l2": 0.6373152732849121, | |
| "eval_icace_normdiff": 0.40072518587112427, | |
| "eval_loss": 7.812716960906982, | |
| "eval_runtime": 22.1689, | |
| "eval_samples_per_second": 175.832, | |
| "eval_steps_per_second": 2.752, | |
| "step": 800 | |
| }, | |
| { | |
| "epoch": 1.41, | |
| "learning_rate": 4.111785095320624e-05, | |
| "loss": 7.8938, | |
| "step": 820 | |
| }, | |
| { | |
| "epoch": 1.44, | |
| "learning_rate": 4.0901213171577124e-05, | |
| "loss": 7.7746, | |
| "step": 840 | |
| }, | |
| { | |
| "epoch": 1.46, | |
| "eval_distillation_accuracy_counterfactual": 0.5243714725500257, | |
| "eval_distillation_accuracy_factual": 0.836326321190354, | |
| "eval_distillation_f1_counterfactual": 0.4465427270486524, | |
| "eval_distillation_f1_factual": 0.82730940599887, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5179579271421242, | |
| "eval_groundtruth_f1_counterfactual": 0.4470807155891924, | |
| "eval_groundtruth_f1_factual": 0.6724916960045889, | |
| "eval_icace_cosine": 0.4680478870868683, | |
| "eval_icace_l2": 0.6270192861557007, | |
| "eval_icace_normdiff": 0.3876318037509918, | |
| "eval_loss": 7.512824535369873, | |
| "eval_runtime": 22.2529, | |
| "eval_samples_per_second": 175.168, | |
| "eval_steps_per_second": 2.741, | |
| "step": 850 | |
| }, | |
| { | |
| "epoch": 1.48, | |
| "learning_rate": 4.068457538994801e-05, | |
| "loss": 7.6544, | |
| "step": 860 | |
| }, | |
| { | |
| "epoch": 1.51, | |
| "learning_rate": 4.04679376083189e-05, | |
| "loss": 7.7708, | |
| "step": 880 | |
| }, | |
| { | |
| "epoch": 1.54, | |
| "learning_rate": 4.025129982668977e-05, | |
| "loss": 7.6759, | |
| "step": 900 | |
| }, | |
| { | |
| "epoch": 1.54, | |
| "eval_distillation_accuracy_counterfactual": 0.5194971780400205, | |
| "eval_distillation_accuracy_factual": 0.814520266803489, | |
| "eval_distillation_f1_counterfactual": 0.45570094642434356, | |
| "eval_distillation_f1_factual": 0.805867060477024, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5007696254489482, | |
| "eval_groundtruth_f1_counterfactual": 0.43858879322504424, | |
| "eval_groundtruth_f1_factual": 0.657928782779729, | |
| "eval_icace_cosine": 0.4692392945289612, | |
| "eval_icace_l2": 0.6410536766052246, | |
| "eval_icace_normdiff": 0.4014453887939453, | |
| "eval_loss": 7.68869161605835, | |
| "eval_runtime": 22.1915, | |
| "eval_samples_per_second": 175.653, | |
| "eval_steps_per_second": 2.749, | |
| "step": 900 | |
| }, | |
| { | |
| "epoch": 1.58, | |
| "learning_rate": 4.003466204506066e-05, | |
| "loss": 7.8545, | |
| "step": 920 | |
| }, | |
| { | |
| "epoch": 1.61, | |
| "learning_rate": 3.9818024263431546e-05, | |
| "loss": 7.5503, | |
| "step": 940 | |
| }, | |
| { | |
| "epoch": 1.63, | |
| "eval_distillation_accuracy_counterfactual": 0.5402770651616213, | |
| "eval_distillation_accuracy_factual": 0.8211903540277066, | |
| "eval_distillation_f1_counterfactual": 0.453286607279591, | |
| "eval_distillation_f1_factual": 0.8132738069526344, | |
| "eval_groundtruth_accuracy_counterfactual": 0.530528476141611, | |
| "eval_groundtruth_f1_counterfactual": 0.4485755018961964, | |
| "eval_groundtruth_f1_factual": 0.6565776752764576, | |
| "eval_icace_cosine": 0.4694054126739502, | |
| "eval_icace_l2": 0.625873863697052, | |
| "eval_icace_normdiff": 0.39317741990089417, | |
| "eval_loss": 7.5347418785095215, | |
| "eval_runtime": 22.1059, | |
| "eval_samples_per_second": 176.333, | |
| "eval_steps_per_second": 2.759, | |
| "step": 950 | |
| }, | |
| { | |
| "epoch": 1.65, | |
| "learning_rate": 3.960138648180243e-05, | |
| "loss": 7.7698, | |
| "step": 960 | |
| }, | |
| { | |
| "epoch": 1.68, | |
| "learning_rate": 3.938474870017331e-05, | |
| "loss": 7.7985, | |
| "step": 980 | |
| }, | |
| { | |
| "epoch": 1.72, | |
| "learning_rate": 3.91681109185442e-05, | |
| "loss": 7.686, | |
| "step": 1000 | |
| }, | |
| { | |
| "epoch": 1.72, | |
| "eval_distillation_accuracy_counterfactual": 0.5420728578758338, | |
| "eval_distillation_accuracy_factual": 0.8396613648024628, | |
| "eval_distillation_f1_counterfactual": 0.456055619550645, | |
| "eval_distillation_f1_factual": 0.827631226039748, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5292457670600308, | |
| "eval_groundtruth_f1_counterfactual": 0.45013798447650544, | |
| "eval_groundtruth_f1_factual": 0.6763146243844917, | |
| "eval_icace_cosine": 0.45797210931777954, | |
| "eval_icace_l2": 0.611527144908905, | |
| "eval_icace_normdiff": 0.37649333477020264, | |
| "eval_loss": 7.2710371017456055, | |
| "eval_runtime": 22.2409, | |
| "eval_samples_per_second": 175.262, | |
| "eval_steps_per_second": 2.743, | |
| "step": 1000 | |
| } | |
| ], | |
| "max_steps": 4616, | |
| "num_train_epochs": 8, | |
| "total_flos": 8474526459302400.0, | |
| "trial_name": null, | |
| "trial_params": null | |
| } | |