| { | |
| "best_metric": 0.5015928149223328, | |
| "best_model_checkpoint": "model_output/e2e_opentable_5_way__approximate__0-shot__seed-42__lstm/checkpoint-1550", | |
| "epoch": 2.6586620926243567, | |
| "global_step": 1550, | |
| "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": 43.0786, | |
| "step": 20 | |
| }, | |
| { | |
| "epoch": 0.07, | |
| "learning_rate": 4.956672443674177e-05, | |
| "loss": 38.7276, | |
| "step": 40 | |
| }, | |
| { | |
| "epoch": 0.09, | |
| "eval_distillation_accuracy_counterfactual": 0.413801949717804, | |
| "eval_distillation_accuracy_factual": 0.7095946639302206, | |
| "eval_distillation_f1_counterfactual": 0.3810686639321449, | |
| "eval_distillation_f1_factual": 0.6957955686304648, | |
| "eval_groundtruth_accuracy_counterfactual": 0.3301693175987686, | |
| "eval_groundtruth_f1_counterfactual": 0.3009236688640019, | |
| "eval_groundtruth_f1_factual": 0.4899999707848249, | |
| "eval_icace_cosine": 0.7151280045509338, | |
| "eval_icace_l2": 0.7787489891052246, | |
| "eval_icace_normdiff": 0.5948067307472229, | |
| "eval_loss": 20.85917091369629, | |
| "eval_runtime": 4.2958, | |
| "eval_samples_per_second": 907.39, | |
| "eval_steps_per_second": 7.216, | |
| "step": 50 | |
| }, | |
| { | |
| "epoch": 0.1, | |
| "learning_rate": 4.935008665511265e-05, | |
| "loss": 35.9292, | |
| "step": 60 | |
| }, | |
| { | |
| "epoch": 0.14, | |
| "learning_rate": 4.913344887348354e-05, | |
| "loss": 32.7092, | |
| "step": 80 | |
| }, | |
| { | |
| "epoch": 0.17, | |
| "learning_rate": 4.891681109185442e-05, | |
| "loss": 28.9099, | |
| "step": 100 | |
| }, | |
| { | |
| "epoch": 0.17, | |
| "eval_distillation_accuracy_counterfactual": 0.44971780400205236, | |
| "eval_distillation_accuracy_factual": 0.7447408927655208, | |
| "eval_distillation_f1_counterfactual": 0.4113364356485835, | |
| "eval_distillation_f1_factual": 0.7313179359518789, | |
| "eval_groundtruth_accuracy_counterfactual": 0.3763468445356593, | |
| "eval_groundtruth_f1_counterfactual": 0.340596840887459, | |
| "eval_groundtruth_f1_factual": 0.5410055670992968, | |
| "eval_icace_cosine": 0.5891422629356384, | |
| "eval_icace_l2": 0.708361029624939, | |
| "eval_icace_normdiff": 0.5622901916503906, | |
| "eval_loss": 15.46045207977295, | |
| "eval_runtime": 4.856, | |
| "eval_samples_per_second": 802.712, | |
| "eval_steps_per_second": 6.384, | |
| "step": 100 | |
| }, | |
| { | |
| "epoch": 0.21, | |
| "learning_rate": 4.8700173310225307e-05, | |
| "loss": 23.7929, | |
| "step": 120 | |
| }, | |
| { | |
| "epoch": 0.24, | |
| "learning_rate": 4.848353552859619e-05, | |
| "loss": 19.9476, | |
| "step": 140 | |
| }, | |
| { | |
| "epoch": 0.26, | |
| "eval_distillation_accuracy_counterfactual": 0.4807593637762955, | |
| "eval_distillation_accuracy_factual": 0.6631605951770139, | |
| "eval_distillation_f1_counterfactual": 0.4198039457002024, | |
| "eval_distillation_f1_factual": 0.6325010855821678, | |
| "eval_groundtruth_accuracy_counterfactual": 0.439712673165726, | |
| "eval_groundtruth_f1_counterfactual": 0.38279540235888865, | |
| "eval_groundtruth_f1_factual": 0.49479750987904253, | |
| "eval_icace_cosine": 0.5389060378074646, | |
| "eval_icace_l2": 0.6715835332870483, | |
| "eval_icace_normdiff": 0.43847718834877014, | |
| "eval_loss": 12.59195613861084, | |
| "eval_runtime": 4.5282, | |
| "eval_samples_per_second": 860.827, | |
| "eval_steps_per_second": 6.846, | |
| "step": 150 | |
| }, | |
| { | |
| "epoch": 0.27, | |
| "learning_rate": 4.826689774696707e-05, | |
| "loss": 18.2834, | |
| "step": 160 | |
| }, | |
| { | |
| "epoch": 0.31, | |
| "learning_rate": 4.8050259965337955e-05, | |
| "loss": 16.7884, | |
| "step": 180 | |
| }, | |
| { | |
| "epoch": 0.34, | |
| "learning_rate": 4.7833622183708845e-05, | |
| "loss": 16.1986, | |
| "step": 200 | |
| }, | |
| { | |
| "epoch": 0.34, | |
| "eval_distillation_accuracy_counterfactual": 0.5100051308363264, | |
| "eval_distillation_accuracy_factual": 0.7208825038481272, | |
| "eval_distillation_f1_counterfactual": 0.44399257347283927, | |
| "eval_distillation_f1_factual": 0.6985538094507235, | |
| "eval_groundtruth_accuracy_counterfactual": 0.469728065674705, | |
| "eval_groundtruth_f1_counterfactual": 0.4066620855800582, | |
| "eval_groundtruth_f1_factual": 0.5453203881049435, | |
| "eval_icace_cosine": 0.5247439742088318, | |
| "eval_icace_l2": 0.6397933959960938, | |
| "eval_icace_normdiff": 0.42328375577926636, | |
| "eval_loss": 11.142946243286133, | |
| "eval_runtime": 4.7099, | |
| "eval_samples_per_second": 827.623, | |
| "eval_steps_per_second": 6.582, | |
| "step": 200 | |
| }, | |
| { | |
| "epoch": 0.38, | |
| "learning_rate": 4.761698440207972e-05, | |
| "loss": 16.3317, | |
| "step": 220 | |
| }, | |
| { | |
| "epoch": 0.41, | |
| "learning_rate": 4.740034662045061e-05, | |
| "loss": 15.6449, | |
| "step": 240 | |
| }, | |
| { | |
| "epoch": 0.43, | |
| "eval_distillation_accuracy_counterfactual": 0.5177013853258081, | |
| "eval_distillation_accuracy_factual": 0.7614161108260646, | |
| "eval_distillation_f1_counterfactual": 0.43564292507442814, | |
| "eval_distillation_f1_factual": 0.7349550078954596, | |
| "eval_groundtruth_accuracy_counterfactual": 0.48255515649050795, | |
| "eval_groundtruth_f1_counterfactual": 0.40504065429982916, | |
| "eval_groundtruth_f1_factual": 0.5535077645302229, | |
| "eval_icace_cosine": 0.516343355178833, | |
| "eval_icace_l2": 0.6303517818450928, | |
| "eval_icace_normdiff": 0.4126652777194977, | |
| "eval_loss": 10.60521411895752, | |
| "eval_runtime": 4.617, | |
| "eval_samples_per_second": 844.277, | |
| "eval_steps_per_second": 6.714, | |
| "step": 250 | |
| }, | |
| { | |
| "epoch": 0.45, | |
| "learning_rate": 4.7183708838821494e-05, | |
| "loss": 15.2981, | |
| "step": 260 | |
| }, | |
| { | |
| "epoch": 0.48, | |
| "learning_rate": 4.6967071057192376e-05, | |
| "loss": 15.1658, | |
| "step": 280 | |
| }, | |
| { | |
| "epoch": 0.51, | |
| "learning_rate": 4.675043327556326e-05, | |
| "loss": 15.0967, | |
| "step": 300 | |
| }, | |
| { | |
| "epoch": 0.51, | |
| "eval_distillation_accuracy_counterfactual": 0.5151359671626475, | |
| "eval_distillation_accuracy_factual": 0.7678296562339662, | |
| "eval_distillation_f1_counterfactual": 0.4566628678247276, | |
| "eval_distillation_f1_factual": 0.7434229464297083, | |
| "eval_groundtruth_accuracy_counterfactual": 0.4856336582863007, | |
| "eval_groundtruth_f1_counterfactual": 0.43108046151644946, | |
| "eval_groundtruth_f1_factual": 0.5711126744289745, | |
| "eval_icace_cosine": 0.5121855735778809, | |
| "eval_icace_l2": 0.636752188205719, | |
| "eval_icace_normdiff": 0.41414302587509155, | |
| "eval_loss": 10.544304847717285, | |
| "eval_runtime": 4.8047, | |
| "eval_samples_per_second": 811.294, | |
| "eval_steps_per_second": 6.452, | |
| "step": 300 | |
| }, | |
| { | |
| "epoch": 0.55, | |
| "learning_rate": 4.653379549393415e-05, | |
| "loss": 14.9139, | |
| "step": 320 | |
| }, | |
| { | |
| "epoch": 0.58, | |
| "learning_rate": 4.6317157712305025e-05, | |
| "loss": 14.7833, | |
| "step": 340 | |
| }, | |
| { | |
| "epoch": 0.6, | |
| "eval_distillation_accuracy_counterfactual": 0.525910723447922, | |
| "eval_distillation_accuracy_factual": 0.7939969214982042, | |
| "eval_distillation_f1_counterfactual": 0.4562979855469525, | |
| "eval_distillation_f1_factual": 0.7748803987234086, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5010261672652643, | |
| "eval_groundtruth_f1_counterfactual": 0.43277911965474775, | |
| "eval_groundtruth_f1_factual": 0.5673297128547324, | |
| "eval_icace_cosine": 0.5058079957962036, | |
| "eval_icace_l2": 0.6291938424110413, | |
| "eval_icace_normdiff": 0.41149842739105225, | |
| "eval_loss": 10.096619606018066, | |
| "eval_runtime": 4.807, | |
| "eval_samples_per_second": 810.901, | |
| "eval_steps_per_second": 6.449, | |
| "step": 350 | |
| }, | |
| { | |
| "epoch": 0.62, | |
| "learning_rate": 4.6100519930675915e-05, | |
| "loss": 14.3751, | |
| "step": 360 | |
| }, | |
| { | |
| "epoch": 0.65, | |
| "learning_rate": 4.58838821490468e-05, | |
| "loss": 14.1618, | |
| "step": 380 | |
| }, | |
| { | |
| "epoch": 0.69, | |
| "learning_rate": 4.566724436741768e-05, | |
| "loss": 14.1163, | |
| "step": 400 | |
| }, | |
| { | |
| "epoch": 0.69, | |
| "eval_distillation_accuracy_counterfactual": 0.5110312981015905, | |
| "eval_distillation_accuracy_factual": 0.8142637249871729, | |
| "eval_distillation_f1_counterfactual": 0.42866472439070946, | |
| "eval_distillation_f1_factual": 0.7995337396770525, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5002565418163161, | |
| "eval_groundtruth_f1_counterfactual": 0.4196447531444125, | |
| "eval_groundtruth_f1_factual": 0.570939084196488, | |
| "eval_icace_cosine": 0.5057737827301025, | |
| "eval_icace_l2": 0.6323451399803162, | |
| "eval_icace_normdiff": 0.4111088812351227, | |
| "eval_loss": 9.999621391296387, | |
| "eval_runtime": 4.6656, | |
| "eval_samples_per_second": 835.478, | |
| "eval_steps_per_second": 6.644, | |
| "step": 400 | |
| }, | |
| { | |
| "epoch": 0.72, | |
| "learning_rate": 4.5450606585788563e-05, | |
| "loss": 14.1471, | |
| "step": 420 | |
| }, | |
| { | |
| "epoch": 0.75, | |
| "learning_rate": 4.5233968804159446e-05, | |
| "loss": 14.0263, | |
| "step": 440 | |
| }, | |
| { | |
| "epoch": 0.77, | |
| "eval_distillation_accuracy_counterfactual": 0.5094920472036942, | |
| "eval_distillation_accuracy_factual": 0.8224730631092868, | |
| "eval_distillation_f1_counterfactual": 0.4472075955335491, | |
| "eval_distillation_f1_factual": 0.8071684604194267, | |
| "eval_groundtruth_accuracy_counterfactual": 0.49640841457157514, | |
| "eval_groundtruth_f1_counterfactual": 0.43563795723268167, | |
| "eval_groundtruth_f1_factual": 0.5741319757982798, | |
| "eval_icace_cosine": 0.5053402185440063, | |
| "eval_icace_l2": 0.6384401917457581, | |
| "eval_icace_normdiff": 0.41610294580459595, | |
| "eval_loss": 9.992120742797852, | |
| "eval_runtime": 4.6252, | |
| "eval_samples_per_second": 842.767, | |
| "eval_steps_per_second": 6.702, | |
| "step": 450 | |
| }, | |
| { | |
| "epoch": 0.79, | |
| "learning_rate": 4.501733102253033e-05, | |
| "loss": 13.777, | |
| "step": 460 | |
| }, | |
| { | |
| "epoch": 0.82, | |
| "learning_rate": 4.480069324090121e-05, | |
| "loss": 13.7366, | |
| "step": 480 | |
| }, | |
| { | |
| "epoch": 0.86, | |
| "learning_rate": 4.45840554592721e-05, | |
| "loss": 13.645, | |
| "step": 500 | |
| }, | |
| { | |
| "epoch": 0.86, | |
| "eval_distillation_accuracy_counterfactual": 0.5187275525910724, | |
| "eval_distillation_accuracy_factual": 0.8206772703950744, | |
| "eval_distillation_f1_counterfactual": 0.44479298662465616, | |
| "eval_distillation_f1_factual": 0.8084170023738808, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5107747562852745, | |
| "eval_groundtruth_f1_counterfactual": 0.43819547536098524, | |
| "eval_groundtruth_f1_factual": 0.5804064040286399, | |
| "eval_icace_cosine": 0.5102166533470154, | |
| "eval_icace_l2": 0.6428956985473633, | |
| "eval_icace_normdiff": 0.4216759502887726, | |
| "eval_loss": 10.014728546142578, | |
| "eval_runtime": 4.6422, | |
| "eval_samples_per_second": 839.683, | |
| "eval_steps_per_second": 6.678, | |
| "step": 500 | |
| }, | |
| { | |
| "epoch": 0.89, | |
| "learning_rate": 4.436741767764298e-05, | |
| "loss": 13.6366, | |
| "step": 520 | |
| }, | |
| { | |
| "epoch": 0.93, | |
| "learning_rate": 4.415077989601387e-05, | |
| "loss": 13.5142, | |
| "step": 540 | |
| }, | |
| { | |
| "epoch": 0.94, | |
| "eval_distillation_accuracy_counterfactual": 0.521292970754233, | |
| "eval_distillation_accuracy_factual": 0.8222165212929707, | |
| "eval_distillation_f1_counterfactual": 0.44615115849333264, | |
| "eval_distillation_f1_factual": 0.8098912953108461, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5138532580810672, | |
| "eval_groundtruth_f1_counterfactual": 0.4412909861048945, | |
| "eval_groundtruth_f1_factual": 0.5759446881541332, | |
| "eval_icace_cosine": 0.5050559043884277, | |
| "eval_icace_l2": 0.6343655586242676, | |
| "eval_icace_normdiff": 0.4120153784751892, | |
| "eval_loss": 9.914525032043457, | |
| "eval_runtime": 4.5769, | |
| "eval_samples_per_second": 851.659, | |
| "eval_steps_per_second": 6.773, | |
| "step": 550 | |
| }, | |
| { | |
| "epoch": 0.96, | |
| "learning_rate": 4.393414211438475e-05, | |
| "loss": 13.516, | |
| "step": 560 | |
| }, | |
| { | |
| "epoch": 0.99, | |
| "learning_rate": 4.371750433275563e-05, | |
| "loss": 13.2886, | |
| "step": 580 | |
| }, | |
| { | |
| "epoch": 1.03, | |
| "learning_rate": 4.3500866551126516e-05, | |
| "loss": 13.0863, | |
| "step": 600 | |
| }, | |
| { | |
| "epoch": 1.03, | |
| "eval_distillation_accuracy_counterfactual": 0.5225756798358132, | |
| "eval_distillation_accuracy_factual": 0.8322216521292971, | |
| "eval_distillation_f1_counterfactual": 0.44257027446874675, | |
| "eval_distillation_f1_factual": 0.8222712281483833, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5159055926115957, | |
| "eval_groundtruth_f1_counterfactual": 0.43868456910889764, | |
| "eval_groundtruth_f1_factual": 0.5752420543125346, | |
| "eval_icace_cosine": 0.5034614205360413, | |
| "eval_icace_l2": 0.6326570510864258, | |
| "eval_icace_normdiff": 0.4119584262371063, | |
| "eval_loss": 9.747785568237305, | |
| "eval_runtime": 4.5826, | |
| "eval_samples_per_second": 850.614, | |
| "eval_steps_per_second": 6.765, | |
| "step": 600 | |
| }, | |
| { | |
| "epoch": 1.06, | |
| "learning_rate": 4.3284228769497406e-05, | |
| "loss": 12.9339, | |
| "step": 620 | |
| }, | |
| { | |
| "epoch": 1.1, | |
| "learning_rate": 4.306759098786828e-05, | |
| "loss": 13.0796, | |
| "step": 640 | |
| }, | |
| { | |
| "epoch": 1.11, | |
| "eval_distillation_accuracy_counterfactual": 0.5110312981015905, | |
| "eval_distillation_accuracy_factual": 0.8273473576192919, | |
| "eval_distillation_f1_counterfactual": 0.43220510255327216, | |
| "eval_distillation_f1_factual": 0.8195939186393728, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5094920472036942, | |
| "eval_groundtruth_f1_counterfactual": 0.43438648520235124, | |
| "eval_groundtruth_f1_factual": 0.5765313785081062, | |
| "eval_icace_cosine": 0.5092682242393494, | |
| "eval_icace_l2": 0.6488710045814514, | |
| "eval_icace_normdiff": 0.42473679780960083, | |
| "eval_loss": 10.03084659576416, | |
| "eval_runtime": 4.5757, | |
| "eval_samples_per_second": 851.883, | |
| "eval_steps_per_second": 6.775, | |
| "step": 650 | |
| }, | |
| { | |
| "epoch": 1.13, | |
| "learning_rate": 4.285095320623917e-05, | |
| "loss": 13.3328, | |
| "step": 660 | |
| }, | |
| { | |
| "epoch": 1.17, | |
| "learning_rate": 4.2634315424610055e-05, | |
| "loss": 13.2103, | |
| "step": 680 | |
| }, | |
| { | |
| "epoch": 1.2, | |
| "learning_rate": 4.241767764298094e-05, | |
| "loss": 12.8642, | |
| "step": 700 | |
| }, | |
| { | |
| "epoch": 1.2, | |
| "eval_distillation_accuracy_counterfactual": 0.5202668034889687, | |
| "eval_distillation_accuracy_factual": 0.8227296049256029, | |
| "eval_distillation_f1_counterfactual": 0.44726878993383945, | |
| "eval_distillation_f1_factual": 0.8144807445603238, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5146228835300154, | |
| "eval_groundtruth_f1_counterfactual": 0.44463257796889355, | |
| "eval_groundtruth_f1_factual": 0.5825450226028441, | |
| "eval_icace_cosine": 0.5063404440879822, | |
| "eval_icace_l2": 0.6389003396034241, | |
| "eval_icace_normdiff": 0.41857677698135376, | |
| "eval_loss": 9.888192176818848, | |
| "eval_runtime": 4.6226, | |
| "eval_samples_per_second": 843.25, | |
| "eval_steps_per_second": 6.706, | |
| "step": 700 | |
| }, | |
| { | |
| "epoch": 1.23, | |
| "learning_rate": 4.220103986135182e-05, | |
| "loss": 12.5476, | |
| "step": 720 | |
| }, | |
| { | |
| "epoch": 1.27, | |
| "learning_rate": 4.198440207972271e-05, | |
| "loss": 13.1858, | |
| "step": 740 | |
| }, | |
| { | |
| "epoch": 1.29, | |
| "eval_distillation_accuracy_counterfactual": 0.5243714725500257, | |
| "eval_distillation_accuracy_factual": 0.8327347357619292, | |
| "eval_distillation_f1_counterfactual": 0.4460710855321577, | |
| "eval_distillation_f1_factual": 0.8222063551634736, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5182144689584403, | |
| "eval_groundtruth_f1_counterfactual": 0.4434489043121113, | |
| "eval_groundtruth_f1_factual": 0.5879251850701077, | |
| "eval_icace_cosine": 0.5027849674224854, | |
| "eval_icace_l2": 0.6359062790870667, | |
| "eval_icace_normdiff": 0.41075506806373596, | |
| "eval_loss": 9.917060852050781, | |
| "eval_runtime": 4.7653, | |
| "eval_samples_per_second": 817.997, | |
| "eval_steps_per_second": 6.505, | |
| "step": 750 | |
| }, | |
| { | |
| "epoch": 1.3, | |
| "learning_rate": 4.1767764298093586e-05, | |
| "loss": 12.7591, | |
| "step": 760 | |
| }, | |
| { | |
| "epoch": 1.34, | |
| "learning_rate": 4.1551126516464476e-05, | |
| "loss": 12.8519, | |
| "step": 780 | |
| }, | |
| { | |
| "epoch": 1.37, | |
| "learning_rate": 4.133448873483536e-05, | |
| "loss": 12.7851, | |
| "step": 800 | |
| }, | |
| { | |
| "epoch": 1.37, | |
| "eval_distillation_accuracy_counterfactual": 0.5118009235505387, | |
| "eval_distillation_accuracy_factual": 0.8137506413545408, | |
| "eval_distillation_f1_counterfactual": 0.43914802457549407, | |
| "eval_distillation_f1_factual": 0.8049341794937319, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5079527963057978, | |
| "eval_groundtruth_f1_counterfactual": 0.4407143465132375, | |
| "eval_groundtruth_f1_factual": 0.5835148341351625, | |
| "eval_icace_cosine": 0.5061963796615601, | |
| "eval_icace_l2": 0.6484118103981018, | |
| "eval_icace_normdiff": 0.42115285992622375, | |
| "eval_loss": 10.143532752990723, | |
| "eval_runtime": 4.6865, | |
| "eval_samples_per_second": 831.75, | |
| "eval_steps_per_second": 6.615, | |
| "step": 800 | |
| }, | |
| { | |
| "epoch": 1.41, | |
| "learning_rate": 4.111785095320624e-05, | |
| "loss": 12.8646, | |
| "step": 820 | |
| }, | |
| { | |
| "epoch": 1.44, | |
| "learning_rate": 4.0901213171577124e-05, | |
| "loss": 12.558, | |
| "step": 840 | |
| }, | |
| { | |
| "epoch": 1.46, | |
| "eval_distillation_accuracy_counterfactual": 0.5307850179579271, | |
| "eval_distillation_accuracy_factual": 0.831965110312981, | |
| "eval_distillation_f1_counterfactual": 0.4484519611985586, | |
| "eval_distillation_f1_factual": 0.822168247394948, | |
| "eval_groundtruth_accuracy_counterfactual": 0.517444843509492, | |
| "eval_groundtruth_f1_counterfactual": 0.43872302047817147, | |
| "eval_groundtruth_f1_factual": 0.5800644441761371, | |
| "eval_icace_cosine": 0.5030389428138733, | |
| "eval_icace_l2": 0.632723867893219, | |
| "eval_icace_normdiff": 0.4054679870605469, | |
| "eval_loss": 9.84372615814209, | |
| "eval_runtime": 5.0033, | |
| "eval_samples_per_second": 779.083, | |
| "eval_steps_per_second": 6.196, | |
| "step": 850 | |
| }, | |
| { | |
| "epoch": 1.48, | |
| "learning_rate": 4.068457538994801e-05, | |
| "loss": 12.828, | |
| "step": 860 | |
| }, | |
| { | |
| "epoch": 1.51, | |
| "learning_rate": 4.04679376083189e-05, | |
| "loss": 12.3635, | |
| "step": 880 | |
| }, | |
| { | |
| "epoch": 1.54, | |
| "learning_rate": 4.025129982668977e-05, | |
| "loss": 12.6661, | |
| "step": 900 | |
| }, | |
| { | |
| "epoch": 1.54, | |
| "eval_distillation_accuracy_counterfactual": 0.5253976398152899, | |
| "eval_distillation_accuracy_factual": 0.8365828630066701, | |
| "eval_distillation_f1_counterfactual": 0.450312280094982, | |
| "eval_distillation_f1_factual": 0.8262625380550788, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5189840944073884, | |
| "eval_groundtruth_f1_counterfactual": 0.44867654499260745, | |
| "eval_groundtruth_f1_factual": 0.5750955137521894, | |
| "eval_icace_cosine": 0.5052414536476135, | |
| "eval_icace_l2": 0.6440022587776184, | |
| "eval_icace_normdiff": 0.41912850737571716, | |
| "eval_loss": 9.941903114318848, | |
| "eval_runtime": 4.5789, | |
| "eval_samples_per_second": 851.288, | |
| "eval_steps_per_second": 6.77, | |
| "step": 900 | |
| }, | |
| { | |
| "epoch": 1.58, | |
| "learning_rate": 4.003466204506066e-05, | |
| "loss": 12.9309, | |
| "step": 920 | |
| }, | |
| { | |
| "epoch": 1.61, | |
| "learning_rate": 3.9818024263431546e-05, | |
| "loss": 12.463, | |
| "step": 940 | |
| }, | |
| { | |
| "epoch": 1.63, | |
| "eval_distillation_accuracy_counterfactual": 0.513340174448435, | |
| "eval_distillation_accuracy_factual": 0.8099025141097999, | |
| "eval_distillation_f1_counterfactual": 0.4451801537186667, | |
| "eval_distillation_f1_factual": 0.8018083383301378, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5135967162647511, | |
| "eval_groundtruth_f1_counterfactual": 0.45122773824298, | |
| "eval_groundtruth_f1_factual": 0.582459738995094, | |
| "eval_icace_cosine": 0.5057380199432373, | |
| "eval_icace_l2": 0.6424580216407776, | |
| "eval_icace_normdiff": 0.4136655926704407, | |
| "eval_loss": 10.070512771606445, | |
| "eval_runtime": 4.5868, | |
| "eval_samples_per_second": 849.823, | |
| "eval_steps_per_second": 6.758, | |
| "step": 950 | |
| }, | |
| { | |
| "epoch": 1.65, | |
| "learning_rate": 3.960138648180243e-05, | |
| "loss": 12.6212, | |
| "step": 960 | |
| }, | |
| { | |
| "epoch": 1.68, | |
| "learning_rate": 3.938474870017331e-05, | |
| "loss": 12.6221, | |
| "step": 980 | |
| }, | |
| { | |
| "epoch": 1.72, | |
| "learning_rate": 3.91681109185442e-05, | |
| "loss": 12.5131, | |
| "step": 1000 | |
| }, | |
| { | |
| "epoch": 1.72, | |
| "eval_distillation_accuracy_counterfactual": 0.5192406362237044, | |
| "eval_distillation_accuracy_factual": 0.8127244740892765, | |
| "eval_distillation_f1_counterfactual": 0.4342093180083893, | |
| "eval_distillation_f1_factual": 0.8020837833036161, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5156490507952797, | |
| "eval_groundtruth_f1_counterfactual": 0.4371545167935388, | |
| "eval_groundtruth_f1_factual": 0.5865676169660375, | |
| "eval_icace_cosine": 0.5021520256996155, | |
| "eval_icace_l2": 0.6322320699691772, | |
| "eval_icace_normdiff": 0.40523087978363037, | |
| "eval_loss": 9.895723342895508, | |
| "eval_runtime": 4.6904, | |
| "eval_samples_per_second": 831.068, | |
| "eval_steps_per_second": 6.609, | |
| "step": 1000 | |
| }, | |
| { | |
| "epoch": 1.75, | |
| "learning_rate": 3.895147313691508e-05, | |
| "loss": 12.5027, | |
| "step": 1020 | |
| }, | |
| { | |
| "epoch": 1.78, | |
| "learning_rate": 3.873483535528597e-05, | |
| "loss": 12.7083, | |
| "step": 1040 | |
| }, | |
| { | |
| "epoch": 1.8, | |
| "eval_distillation_accuracy_counterfactual": 0.5192406362237044, | |
| "eval_distillation_accuracy_factual": 0.8114417650076963, | |
| "eval_distillation_f1_counterfactual": 0.4263386972710933, | |
| "eval_distillation_f1_factual": 0.7994083878081278, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5192406362237044, | |
| "eval_groundtruth_f1_counterfactual": 0.4329553891077026, | |
| "eval_groundtruth_f1_factual": 0.5912280932648739, | |
| "eval_icace_cosine": 0.5038776397705078, | |
| "eval_icace_l2": 0.6317997574806213, | |
| "eval_icace_normdiff": 0.39978867769241333, | |
| "eval_loss": 10.077560424804688, | |
| "eval_runtime": 5.0771, | |
| "eval_samples_per_second": 767.767, | |
| "eval_steps_per_second": 6.106, | |
| "step": 1050 | |
| }, | |
| { | |
| "epoch": 1.82, | |
| "learning_rate": 3.851819757365685e-05, | |
| "loss": 12.3779, | |
| "step": 1060 | |
| }, | |
| { | |
| "epoch": 1.85, | |
| "learning_rate": 3.830155979202773e-05, | |
| "loss": 12.1297, | |
| "step": 1080 | |
| }, | |
| { | |
| "epoch": 1.89, | |
| "learning_rate": 3.8084922010398616e-05, | |
| "loss": 12.2778, | |
| "step": 1100 | |
| }, | |
| { | |
| "epoch": 1.89, | |
| "eval_distillation_accuracy_counterfactual": 0.5076962544894817, | |
| "eval_distillation_accuracy_factual": 0.8093894304771678, | |
| "eval_distillation_f1_counterfactual": 0.42915135801638715, | |
| "eval_distillation_f1_factual": 0.7988387328754103, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5074397126731657, | |
| "eval_groundtruth_f1_counterfactual": 0.43467229220494497, | |
| "eval_groundtruth_f1_factual": 0.5863614417021632, | |
| "eval_icace_cosine": 0.5077196955680847, | |
| "eval_icace_l2": 0.6488126516342163, | |
| "eval_icace_normdiff": 0.4152773916721344, | |
| "eval_loss": 10.29930591583252, | |
| "eval_runtime": 4.7917, | |
| "eval_samples_per_second": 813.496, | |
| "eval_steps_per_second": 6.47, | |
| "step": 1100 | |
| }, | |
| { | |
| "epoch": 1.92, | |
| "learning_rate": 3.78682842287695e-05, | |
| "loss": 12.391, | |
| "step": 1120 | |
| }, | |
| { | |
| "epoch": 1.96, | |
| "learning_rate": 3.765164644714038e-05, | |
| "loss": 12.1049, | |
| "step": 1140 | |
| }, | |
| { | |
| "epoch": 1.97, | |
| "eval_distillation_accuracy_counterfactual": 0.5277065161621344, | |
| "eval_distillation_accuracy_factual": 0.8114417650076963, | |
| "eval_distillation_f1_counterfactual": 0.4379402413535344, | |
| "eval_distillation_f1_factual": 0.8009596923401426, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5184710107747563, | |
| "eval_groundtruth_f1_counterfactual": 0.4351158243705179, | |
| "eval_groundtruth_f1_factual": 0.5809362022910343, | |
| "eval_icace_cosine": 0.5041907429695129, | |
| "eval_icace_l2": 0.6319468021392822, | |
| "eval_icace_normdiff": 0.4021805226802826, | |
| "eval_loss": 10.031487464904785, | |
| "eval_runtime": 4.6011, | |
| "eval_samples_per_second": 847.186, | |
| "eval_steps_per_second": 6.737, | |
| "step": 1150 | |
| }, | |
| { | |
| "epoch": 1.99, | |
| "learning_rate": 3.7435008665511264e-05, | |
| "loss": 12.3362, | |
| "step": 1160 | |
| }, | |
| { | |
| "epoch": 2.02, | |
| "learning_rate": 3.7218370883882154e-05, | |
| "loss": 12.3522, | |
| "step": 1180 | |
| }, | |
| { | |
| "epoch": 2.06, | |
| "learning_rate": 3.700173310225303e-05, | |
| "loss": 12.2852, | |
| "step": 1200 | |
| }, | |
| { | |
| "epoch": 2.06, | |
| "eval_distillation_accuracy_counterfactual": 0.5089789635710621, | |
| "eval_distillation_accuracy_factual": 0.8093894304771678, | |
| "eval_distillation_f1_counterfactual": 0.4327873004366438, | |
| "eval_distillation_f1_factual": 0.7978534612810504, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5094920472036942, | |
| "eval_groundtruth_f1_counterfactual": 0.44147489438032766, | |
| "eval_groundtruth_f1_factual": 0.5875097166633956, | |
| "eval_icace_cosine": 0.5080219507217407, | |
| "eval_icace_l2": 0.6420192122459412, | |
| "eval_icace_normdiff": 0.40897807478904724, | |
| "eval_loss": 10.132452011108398, | |
| "eval_runtime": 4.7301, | |
| "eval_samples_per_second": 824.088, | |
| "eval_steps_per_second": 6.554, | |
| "step": 1200 | |
| }, | |
| { | |
| "epoch": 2.09, | |
| "learning_rate": 3.678509532062392e-05, | |
| "loss": 12.2919, | |
| "step": 1220 | |
| }, | |
| { | |
| "epoch": 2.13, | |
| "learning_rate": 3.65684575389948e-05, | |
| "loss": 12.2076, | |
| "step": 1240 | |
| }, | |
| { | |
| "epoch": 2.14, | |
| "eval_distillation_accuracy_counterfactual": 0.5082093381221139, | |
| "eval_distillation_accuracy_factual": 0.8052847614161108, | |
| "eval_distillation_f1_counterfactual": 0.421327115393484, | |
| "eval_distillation_f1_factual": 0.7938972956790581, | |
| "eval_groundtruth_accuracy_counterfactual": 0.508722421754746, | |
| "eval_groundtruth_f1_counterfactual": 0.4317767637232814, | |
| "eval_groundtruth_f1_factual": 0.5884182312452589, | |
| "eval_icace_cosine": 0.5102863907814026, | |
| "eval_icace_l2": 0.6493218541145325, | |
| "eval_icace_normdiff": 0.41481640934944153, | |
| "eval_loss": 10.397825241088867, | |
| "eval_runtime": 4.5642, | |
| "eval_samples_per_second": 854.035, | |
| "eval_steps_per_second": 6.792, | |
| "step": 1250 | |
| }, | |
| { | |
| "epoch": 2.16, | |
| "learning_rate": 3.6351819757365686e-05, | |
| "loss": 12.1793, | |
| "step": 1260 | |
| }, | |
| { | |
| "epoch": 2.2, | |
| "learning_rate": 3.613518197573657e-05, | |
| "loss": 11.9955, | |
| "step": 1280 | |
| }, | |
| { | |
| "epoch": 2.23, | |
| "learning_rate": 3.591854419410746e-05, | |
| "loss": 12.0414, | |
| "step": 1300 | |
| }, | |
| { | |
| "epoch": 2.23, | |
| "eval_distillation_accuracy_counterfactual": 0.5159055926115957, | |
| "eval_distillation_accuracy_factual": 0.8088763468445357, | |
| "eval_distillation_f1_counterfactual": 0.43928335663414925, | |
| "eval_distillation_f1_factual": 0.798007092912107, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5192406362237044, | |
| "eval_groundtruth_f1_counterfactual": 0.45071084200597555, | |
| "eval_groundtruth_f1_factual": 0.5881302910281572, | |
| "eval_icace_cosine": 0.5018934607505798, | |
| "eval_icace_l2": 0.6314676403999329, | |
| "eval_icace_normdiff": 0.4030630588531494, | |
| "eval_loss": 10.04751968383789, | |
| "eval_runtime": 4.5717, | |
| "eval_samples_per_second": 852.63, | |
| "eval_steps_per_second": 6.781, | |
| "step": 1300 | |
| }, | |
| { | |
| "epoch": 2.26, | |
| "learning_rate": 3.5701906412478334e-05, | |
| "loss": 12.0019, | |
| "step": 1320 | |
| }, | |
| { | |
| "epoch": 2.3, | |
| "learning_rate": 3.5485268630849224e-05, | |
| "loss": 11.8994, | |
| "step": 1340 | |
| }, | |
| { | |
| "epoch": 2.32, | |
| "eval_distillation_accuracy_counterfactual": 0.5120574653668548, | |
| "eval_distillation_accuracy_factual": 0.8001539250897897, | |
| "eval_distillation_f1_counterfactual": 0.43648655334155206, | |
| "eval_distillation_f1_factual": 0.791237363619484, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5035915854284249, | |
| "eval_groundtruth_f1_counterfactual": 0.4365541899982916, | |
| "eval_groundtruth_f1_factual": 0.5892108143561166, | |
| "eval_icace_cosine": 0.5103155970573425, | |
| "eval_icace_l2": 0.648665726184845, | |
| "eval_icace_normdiff": 0.4152573347091675, | |
| "eval_loss": 10.379837989807129, | |
| "eval_runtime": 4.5419, | |
| "eval_samples_per_second": 858.233, | |
| "eval_steps_per_second": 6.825, | |
| "step": 1350 | |
| }, | |
| { | |
| "epoch": 2.33, | |
| "learning_rate": 3.526863084922011e-05, | |
| "loss": 12.1835, | |
| "step": 1360 | |
| }, | |
| { | |
| "epoch": 2.37, | |
| "learning_rate": 3.505199306759099e-05, | |
| "loss": 12.1258, | |
| "step": 1380 | |
| }, | |
| { | |
| "epoch": 2.4, | |
| "learning_rate": 3.483535528596187e-05, | |
| "loss": 11.8758, | |
| "step": 1400 | |
| }, | |
| { | |
| "epoch": 2.4, | |
| "eval_distillation_accuracy_counterfactual": 0.5092355053873782, | |
| "eval_distillation_accuracy_factual": 0.8055413032324269, | |
| "eval_distillation_f1_counterfactual": 0.43572512517349055, | |
| "eval_distillation_f1_factual": 0.7959903986995227, | |
| "eval_groundtruth_accuracy_counterfactual": 0.512827090815803, | |
| "eval_groundtruth_f1_counterfactual": 0.44567177435185246, | |
| "eval_groundtruth_f1_factual": 0.5944081922689601, | |
| "eval_icace_cosine": 0.504929780960083, | |
| "eval_icace_l2": 0.6384508013725281, | |
| "eval_icace_normdiff": 0.404530793428421, | |
| "eval_loss": 10.173859596252441, | |
| "eval_runtime": 4.614, | |
| "eval_samples_per_second": 844.817, | |
| "eval_steps_per_second": 6.719, | |
| "step": 1400 | |
| }, | |
| { | |
| "epoch": 2.44, | |
| "learning_rate": 3.461871750433276e-05, | |
| "loss": 12.0309, | |
| "step": 1420 | |
| }, | |
| { | |
| "epoch": 2.47, | |
| "learning_rate": 3.440207972270364e-05, | |
| "loss": 11.738, | |
| "step": 1440 | |
| }, | |
| { | |
| "epoch": 2.49, | |
| "eval_distillation_accuracy_counterfactual": 0.5151359671626475, | |
| "eval_distillation_accuracy_factual": 0.8065674704976912, | |
| "eval_distillation_f1_counterfactual": 0.42557736492521486, | |
| "eval_distillation_f1_factual": 0.7962593487138783, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5207798871216008, | |
| "eval_groundtruth_f1_counterfactual": 0.4400972677897436, | |
| "eval_groundtruth_f1_factual": 0.5898699262629978, | |
| "eval_icace_cosine": 0.504587709903717, | |
| "eval_icace_l2": 0.6404756903648376, | |
| "eval_icace_normdiff": 0.4062090516090393, | |
| "eval_loss": 10.268754005432129, | |
| "eval_runtime": 4.5043, | |
| "eval_samples_per_second": 865.395, | |
| "eval_steps_per_second": 6.882, | |
| "step": 1450 | |
| }, | |
| { | |
| "epoch": 2.5, | |
| "learning_rate": 3.418544194107453e-05, | |
| "loss": 11.847, | |
| "step": 1460 | |
| }, | |
| { | |
| "epoch": 2.54, | |
| "learning_rate": 3.396880415944541e-05, | |
| "loss": 11.7855, | |
| "step": 1480 | |
| }, | |
| { | |
| "epoch": 2.57, | |
| "learning_rate": 3.3752166377816294e-05, | |
| "loss": 12.2395, | |
| "step": 1500 | |
| }, | |
| { | |
| "epoch": 2.57, | |
| "eval_distillation_accuracy_counterfactual": 0.5076962544894817, | |
| "eval_distillation_accuracy_factual": 0.793227296049256, | |
| "eval_distillation_f1_counterfactual": 0.42559353759463275, | |
| "eval_distillation_f1_factual": 0.7832657828168742, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5105182144689584, | |
| "eval_groundtruth_f1_counterfactual": 0.4380208778877491, | |
| "eval_groundtruth_f1_factual": 0.5829268254157943, | |
| "eval_icace_cosine": 0.508986234664917, | |
| "eval_icace_l2": 0.6547452211380005, | |
| "eval_icace_normdiff": 0.4193916618824005, | |
| "eval_loss": 10.556840896606445, | |
| "eval_runtime": 4.5426, | |
| "eval_samples_per_second": 858.107, | |
| "eval_steps_per_second": 6.824, | |
| "step": 1500 | |
| }, | |
| { | |
| "epoch": 2.61, | |
| "learning_rate": 3.353552859618718e-05, | |
| "loss": 12.2207, | |
| "step": 1520 | |
| }, | |
| { | |
| "epoch": 2.64, | |
| "learning_rate": 3.331889081455806e-05, | |
| "loss": 12.2085, | |
| "step": 1540 | |
| }, | |
| { | |
| "epoch": 2.66, | |
| "eval_distillation_accuracy_counterfactual": 0.5177013853258081, | |
| "eval_distillation_accuracy_factual": 0.7978450487429452, | |
| "eval_distillation_f1_counterfactual": 0.4343412617259876, | |
| "eval_distillation_f1_factual": 0.7885736470487629, | |
| "eval_groundtruth_accuracy_counterfactual": 0.5241149307337096, | |
| "eval_groundtruth_f1_counterfactual": 0.44795498771470293, | |
| "eval_groundtruth_f1_factual": 0.5848176602353289, | |
| "eval_icace_cosine": 0.5015928149223328, | |
| "eval_icace_l2": 0.6313996911048889, | |
| "eval_icace_normdiff": 0.400736927986145, | |
| "eval_loss": 10.176055908203125, | |
| "eval_runtime": 6.0094, | |
| "eval_samples_per_second": 648.649, | |
| "eval_steps_per_second": 5.159, | |
| "step": 1550 | |
| } | |
| ], | |
| "max_steps": 4616, | |
| "num_train_epochs": 8, | |
| "total_flos": 234153553766400.0, | |
| "trial_name": null, | |
| "trial_params": null | |
| } | |