trainer_state.json

{
  "best_metric": 0.4927329123020172,
  "best_model_checkpoint": "model_output/e2e_opentable_5_way__approximate__0-shot__seed-77__lstm/checkpoint-900",
  "epoch": 1.5437392795883362,
  "global_step": 900,
  "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": 44.1562,
      "step": 20
    },
    {
      "epoch": 0.07,
      "learning_rate": 4.956672443674177e-05,
      "loss": 40.0981,
      "step": 40
    },
    {
      "epoch": 0.09,
      "eval_distillation_accuracy_counterfactual": 0.40533606977937403,
      "eval_distillation_accuracy_factual": 0.7136993329912775,
      "eval_distillation_f1_counterfactual": 0.37584526793260814,
      "eval_distillation_f1_factual": 0.6903688842294228,
      "eval_groundtruth_accuracy_counterfactual": 0.33119548486403283,
      "eval_groundtruth_f1_counterfactual": 0.30050094417395845,
      "eval_groundtruth_f1_factual": 0.5105855093093573,
      "eval_icace_cosine": 0.7006840109825134,
      "eval_icace_l2": 0.7841421961784363,
      "eval_icace_normdiff": 0.6143233180046082,
      "eval_loss": 21.405780792236328,
      "eval_runtime": 4.5232,
      "eval_samples_per_second": 861.775,
      "eval_steps_per_second": 6.854,
      "step": 50
    },
    {
      "epoch": 0.1,
      "learning_rate": 4.935008665511265e-05,
      "loss": 36.4728,
      "step": 60
    },
    {
      "epoch": 0.14,
      "learning_rate": 4.913344887348354e-05,
      "loss": 33.5128,
      "step": 80
    },
    {
      "epoch": 0.17,
      "learning_rate": 4.891681109185442e-05,
      "loss": 29.1246,
      "step": 100
    },
    {
      "epoch": 0.17,
      "eval_distillation_accuracy_counterfactual": 0.4304771677783479,
      "eval_distillation_accuracy_factual": 0.7060030785017958,
      "eval_distillation_f1_counterfactual": 0.3874499036502149,
      "eval_distillation_f1_factual": 0.672147939569431,
      "eval_groundtruth_accuracy_counterfactual": 0.3599281682914315,
      "eval_groundtruth_f1_counterfactual": 0.317370542081978,
      "eval_groundtruth_f1_factual": 0.5138084441251398,
      "eval_icace_cosine": 0.5940763354301453,
      "eval_icace_l2": 0.7375399470329285,
      "eval_icace_normdiff": 0.5838260650634766,
      "eval_loss": 15.98747730255127,
      "eval_runtime": 4.638,
      "eval_samples_per_second": 840.44,
      "eval_steps_per_second": 6.684,
      "step": 100
    },
    {
      "epoch": 0.21,
      "learning_rate": 4.8700173310225307e-05,
      "loss": 25.2302,
      "step": 120
    },
    {
      "epoch": 0.24,
      "learning_rate": 4.848353552859619e-05,
      "loss": 21.5776,
      "step": 140
    },
    {
      "epoch": 0.26,
      "eval_distillation_accuracy_counterfactual": 0.4620318111852232,
      "eval_distillation_accuracy_factual": 0.6246793227296049,
      "eval_distillation_f1_counterfactual": 0.3998229577563853,
      "eval_distillation_f1_factual": 0.5809545209677998,
      "eval_groundtruth_accuracy_counterfactual": 0.4220112878399179,
      "eval_groundtruth_f1_counterfactual": 0.3560556229162239,
      "eval_groundtruth_f1_factual": 0.4660399182176905,
      "eval_icace_cosine": 0.541000485420227,
      "eval_icace_l2": 0.6904007792472839,
      "eval_icace_normdiff": 0.4991605877876282,
      "eval_loss": 12.918840408325195,
      "eval_runtime": 4.7423,
      "eval_samples_per_second": 821.967,
      "eval_steps_per_second": 6.537,
      "step": 150
    },
    {
      "epoch": 0.27,
      "learning_rate": 4.826689774696707e-05,
      "loss": 19.3343,
      "step": 160
    },
    {
      "epoch": 0.31,
      "learning_rate": 4.8050259965337955e-05,
      "loss": 17.9944,
      "step": 180
    },
    {
      "epoch": 0.34,
      "learning_rate": 4.7833622183708845e-05,
      "loss": 17.4913,
      "step": 200
    },
    {
      "epoch": 0.34,
      "eval_distillation_accuracy_counterfactual": 0.48512057465366853,
      "eval_distillation_accuracy_factual": 0.6765007696254489,
      "eval_distillation_f1_counterfactual": 0.4355374039109995,
      "eval_distillation_f1_factual": 0.648085081022865,
      "eval_groundtruth_accuracy_counterfactual": 0.45459209851205745,
      "eval_groundtruth_f1_counterfactual": 0.3992732299014231,
      "eval_groundtruth_f1_factual": 0.5315401786684271,
      "eval_icace_cosine": 0.5196343660354614,
      "eval_icace_l2": 0.6643721461296082,
      "eval_icace_normdiff": 0.46649259328842163,
      "eval_loss": 11.92587947845459,
      "eval_runtime": 4.7274,
      "eval_samples_per_second": 824.557,
      "eval_steps_per_second": 6.558,
      "step": 200
    },
    {
      "epoch": 0.38,
      "learning_rate": 4.761698440207972e-05,
      "loss": 16.9263,
      "step": 220
    },
    {
      "epoch": 0.41,
      "learning_rate": 4.740034662045061e-05,
      "loss": 16.4577,
      "step": 240
    },
    {
      "epoch": 0.43,
      "eval_distillation_accuracy_counterfactual": 0.4953822473063109,
      "eval_distillation_accuracy_factual": 0.6990764494612622,
      "eval_distillation_f1_counterfactual": 0.4349405670052812,
      "eval_distillation_f1_factual": 0.6754484776025494,
      "eval_groundtruth_accuracy_counterfactual": 0.46587993842996406,
      "eval_groundtruth_f1_counterfactual": 0.4007888276500159,
      "eval_groundtruth_f1_factual": 0.537290774599058,
      "eval_icace_cosine": 0.5181905627250671,
      "eval_icace_l2": 0.6551650166511536,
      "eval_icace_normdiff": 0.4527762532234192,
      "eval_loss": 11.470577239990234,
      "eval_runtime": 4.5537,
      "eval_samples_per_second": 856.014,
      "eval_steps_per_second": 6.808,
      "step": 250
    },
    {
      "epoch": 0.45,
      "learning_rate": 4.7183708838821494e-05,
      "loss": 16.175,
      "step": 260
    },
    {
      "epoch": 0.48,
      "learning_rate": 4.6967071057192376e-05,
      "loss": 16.2998,
      "step": 280
    },
    {
      "epoch": 0.51,
      "learning_rate": 4.675043327556326e-05,
      "loss": 15.9865,
      "step": 300
    },
    {
      "epoch": 0.51,
      "eval_distillation_accuracy_counterfactual": 0.49384299640841456,
      "eval_distillation_accuracy_factual": 0.7421754746023602,
      "eval_distillation_f1_counterfactual": 0.43651743845854823,
      "eval_distillation_f1_factual": 0.7241733737284982,
      "eval_groundtruth_accuracy_counterfactual": 0.4674191893278604,
      "eval_groundtruth_f1_counterfactual": 0.4069323134051392,
      "eval_groundtruth_f1_factual": 0.5441331322978117,
      "eval_icace_cosine": 0.5107740163803101,
      "eval_icace_l2": 0.652337908744812,
      "eval_icace_normdiff": 0.44736814498901367,
      "eval_loss": 11.233296394348145,
      "eval_runtime": 7.2912,
      "eval_samples_per_second": 534.616,
      "eval_steps_per_second": 4.252,
      "step": 300
    },
    {
      "epoch": 0.55,
      "learning_rate": 4.653379549393415e-05,
      "loss": 15.9314,
      "step": 320
    },
    {
      "epoch": 0.58,
      "learning_rate": 4.6317157712305025e-05,
      "loss": 15.505,
      "step": 340
    },
    {
      "epoch": 0.6,
      "eval_distillation_accuracy_counterfactual": 0.5,
      "eval_distillation_accuracy_factual": 0.7637249871729092,
      "eval_distillation_f1_counterfactual": 0.44117078829235296,
      "eval_distillation_f1_factual": 0.7500667273075408,
      "eval_groundtruth_accuracy_counterfactual": 0.4879425346331452,
      "eval_groundtruth_f1_counterfactual": 0.42134689982561896,
      "eval_groundtruth_f1_factual": 0.562706144496911,
      "eval_icace_cosine": 0.5107312798500061,
      "eval_icace_l2": 0.6565550565719604,
      "eval_icace_normdiff": 0.4413979947566986,
      "eval_loss": 11.069193840026855,
      "eval_runtime": 6.2756,
      "eval_samples_per_second": 621.137,
      "eval_steps_per_second": 4.94,
      "step": 350
    },
    {
      "epoch": 0.62,
      "learning_rate": 4.6100519930675915e-05,
      "loss": 15.3031,
      "step": 360
    },
    {
      "epoch": 0.65,
      "learning_rate": 4.58838821490468e-05,
      "loss": 15.4206,
      "step": 380
    },
    {
      "epoch": 0.69,
      "learning_rate": 4.566724436741768e-05,
      "loss": 14.8304,
      "step": 400
    },
    {
      "epoch": 0.69,
      "eval_distillation_accuracy_counterfactual": 0.508722421754746,
      "eval_distillation_accuracy_factual": 0.771421241662391,
      "eval_distillation_f1_counterfactual": 0.44238377528971906,
      "eval_distillation_f1_factual": 0.7522621473444253,
      "eval_groundtruth_accuracy_counterfactual": 0.49923037455105185,
      "eval_groundtruth_f1_counterfactual": 0.4273841467316483,
      "eval_groundtruth_f1_factual": 0.5609154619878947,
      "eval_icace_cosine": 0.5019481182098389,
      "eval_icace_l2": 0.6451992988586426,
      "eval_icace_normdiff": 0.4310374855995178,
      "eval_loss": 10.662566184997559,
      "eval_runtime": 4.5814,
      "eval_samples_per_second": 850.836,
      "eval_steps_per_second": 6.767,
      "step": 400
    },
    {
      "epoch": 0.72,
      "learning_rate": 4.5450606585788563e-05,
      "loss": 14.9398,
      "step": 420
    },
    {
      "epoch": 0.75,
      "learning_rate": 4.5233968804159446e-05,
      "loss": 14.8756,
      "step": 440
    },
    {
      "epoch": 0.77,
      "eval_distillation_accuracy_counterfactual": 0.5071831708568497,
      "eval_distillation_accuracy_factual": 0.7806567470497691,
      "eval_distillation_f1_counterfactual": 0.4509914585662811,
      "eval_distillation_f1_factual": 0.7726447146201695,
      "eval_groundtruth_accuracy_counterfactual": 0.49640841457157514,
      "eval_groundtruth_f1_counterfactual": 0.4369358063962636,
      "eval_groundtruth_f1_factual": 0.5701105764409996,
      "eval_icace_cosine": 0.505915641784668,
      "eval_icace_l2": 0.6462938189506531,
      "eval_icace_normdiff": 0.4340927302837372,
      "eval_loss": 10.489557266235352,
      "eval_runtime": 4.6204,
      "eval_samples_per_second": 843.645,
      "eval_steps_per_second": 6.709,
      "step": 450
    },
    {
      "epoch": 0.79,
      "learning_rate": 4.501733102253033e-05,
      "loss": 14.6303,
      "step": 460
    },
    {
      "epoch": 0.82,
      "learning_rate": 4.480069324090121e-05,
      "loss": 14.6318,
      "step": 480
    },
    {
      "epoch": 0.86,
      "learning_rate": 4.45840554592721e-05,
      "loss": 14.3026,
      "step": 500
    },
    {
      "epoch": 0.86,
      "eval_distillation_accuracy_counterfactual": 0.512827090815803,
      "eval_distillation_accuracy_factual": 0.8009235505387378,
      "eval_distillation_f1_counterfactual": 0.45566955055093156,
      "eval_distillation_f1_factual": 0.7939347571605715,
      "eval_groundtruth_accuracy_counterfactual": 0.4987172909184197,
      "eval_groundtruth_f1_counterfactual": 0.4364792426591563,
      "eval_groundtruth_f1_factual": 0.5727123146480367,
      "eval_icace_cosine": 0.4989969730377197,
      "eval_icace_l2": 0.6415925025939941,
      "eval_icace_normdiff": 0.4321776330471039,
      "eval_loss": 10.163902282714844,
      "eval_runtime": 4.5879,
      "eval_samples_per_second": 849.631,
      "eval_steps_per_second": 6.757,
      "step": 500
    },
    {
      "epoch": 0.89,
      "learning_rate": 4.436741767764298e-05,
      "loss": 14.249,
      "step": 520
    },
    {
      "epoch": 0.93,
      "learning_rate": 4.415077989601387e-05,
      "loss": 14.2543,
      "step": 540
    },
    {
      "epoch": 0.94,
      "eval_distillation_accuracy_counterfactual": 0.5005130836326321,
      "eval_distillation_accuracy_factual": 0.793227296049256,
      "eval_distillation_f1_counterfactual": 0.4392802659881442,
      "eval_distillation_f1_factual": 0.7820837225108551,
      "eval_groundtruth_accuracy_counterfactual": 0.4935864545920985,
      "eval_groundtruth_f1_counterfactual": 0.4259921068619771,
      "eval_groundtruth_f1_factual": 0.568994158851058,
      "eval_icace_cosine": 0.5032749772071838,
      "eval_icace_l2": 0.6510319709777832,
      "eval_icace_normdiff": 0.4337131381034851,
      "eval_loss": 10.366103172302246,
      "eval_runtime": 4.5676,
      "eval_samples_per_second": 853.4,
      "eval_steps_per_second": 6.787,
      "step": 550
    },
    {
      "epoch": 0.96,
      "learning_rate": 4.393414211438475e-05,
      "loss": 14.1474,
      "step": 560
    },
    {
      "epoch": 0.99,
      "learning_rate": 4.371750433275563e-05,
      "loss": 13.8257,
      "step": 580
    },
    {
      "epoch": 1.03,
      "learning_rate": 4.3500866551126516e-05,
      "loss": 13.6754,
      "step": 600
    },
    {
      "epoch": 1.03,
      "eval_distillation_accuracy_counterfactual": 0.5151359671626475,
      "eval_distillation_accuracy_factual": 0.8099025141097999,
      "eval_distillation_f1_counterfactual": 0.4455005046191058,
      "eval_distillation_f1_factual": 0.8010898247683294,
      "eval_groundtruth_accuracy_counterfactual": 0.5064135454079015,
      "eval_groundtruth_f1_counterfactual": 0.43123052906317394,
      "eval_groundtruth_f1_factual": 0.5698126916133328,
      "eval_icace_cosine": 0.500103235244751,
      "eval_icace_l2": 0.6384768486022949,
      "eval_icace_normdiff": 0.42454108595848083,
      "eval_loss": 10.087526321411133,
      "eval_runtime": 4.6814,
      "eval_samples_per_second": 832.665,
      "eval_steps_per_second": 6.622,
      "step": 600
    },
    {
      "epoch": 1.06,
      "learning_rate": 4.3284228769497406e-05,
      "loss": 13.7083,
      "step": 620
    },
    {
      "epoch": 1.1,
      "learning_rate": 4.306759098786828e-05,
      "loss": 13.479,
      "step": 640
    },
    {
      "epoch": 1.11,
      "eval_distillation_accuracy_counterfactual": 0.5071831708568497,
      "eval_distillation_accuracy_factual": 0.8114417650076963,
      "eval_distillation_f1_counterfactual": 0.4452735168036456,
      "eval_distillation_f1_factual": 0.8041390027772044,
      "eval_groundtruth_accuracy_counterfactual": 0.5023088763468445,
      "eval_groundtruth_f1_counterfactual": 0.43961184729665465,
      "eval_groundtruth_f1_factual": 0.5743904963335431,
      "eval_icace_cosine": 0.4984728991985321,
      "eval_icace_l2": 0.6454970240592957,
      "eval_icace_normdiff": 0.4303584694862366,
      "eval_loss": 10.136197090148926,
      "eval_runtime": 4.5364,
      "eval_samples_per_second": 859.262,
      "eval_steps_per_second": 6.834,
      "step": 650
    },
    {
      "epoch": 1.13,
      "learning_rate": 4.285095320623917e-05,
      "loss": 13.5985,
      "step": 660
    },
    {
      "epoch": 1.17,
      "learning_rate": 4.2634315424610055e-05,
      "loss": 13.5509,
      "step": 680
    },
    {
      "epoch": 1.2,
      "learning_rate": 4.241767764298094e-05,
      "loss": 13.5415,
      "step": 700
    },
    {
      "epoch": 1.2,
      "eval_distillation_accuracy_counterfactual": 0.508722421754746,
      "eval_distillation_accuracy_factual": 0.8052847614161108,
      "eval_distillation_f1_counterfactual": 0.442105002572385,
      "eval_distillation_f1_factual": 0.7987637622171455,
      "eval_groundtruth_accuracy_counterfactual": 0.5030785017957927,
      "eval_groundtruth_f1_counterfactual": 0.43495956928902413,
      "eval_groundtruth_f1_factual": 0.5712315287508367,
      "eval_icace_cosine": 0.49780699610710144,
      "eval_icace_l2": 0.6452174782752991,
      "eval_icace_normdiff": 0.43159618973731995,
      "eval_loss": 10.0033597946167,
      "eval_runtime": 4.6939,
      "eval_samples_per_second": 830.434,
      "eval_steps_per_second": 6.604,
      "step": 700
    },
    {
      "epoch": 1.23,
      "learning_rate": 4.220103986135182e-05,
      "loss": 13.4033,
      "step": 720
    },
    {
      "epoch": 1.27,
      "learning_rate": 4.198440207972271e-05,
      "loss": 13.1605,
      "step": 740
    },
    {
      "epoch": 1.29,
      "eval_distillation_accuracy_counterfactual": 0.5135967162647511,
      "eval_distillation_accuracy_factual": 0.8155464340687532,
      "eval_distillation_f1_counterfactual": 0.44334782704039266,
      "eval_distillation_f1_factual": 0.8087207034484185,
      "eval_groundtruth_accuracy_counterfactual": 0.5100051308363264,
      "eval_groundtruth_f1_counterfactual": 0.43833016787486423,
      "eval_groundtruth_f1_factual": 0.5747378526464983,
      "eval_icace_cosine": 0.49590805172920227,
      "eval_icace_l2": 0.6382647752761841,
      "eval_icace_normdiff": 0.4257577061653137,
      "eval_loss": 9.923479080200195,
      "eval_runtime": 4.7429,
      "eval_samples_per_second": 821.861,
      "eval_steps_per_second": 6.536,
      "step": 750
    },
    {
      "epoch": 1.3,
      "learning_rate": 4.1767764298093586e-05,
      "loss": 13.2256,
      "step": 760
    },
    {
      "epoch": 1.34,
      "learning_rate": 4.1551126516464476e-05,
      "loss": 13.2255,
      "step": 780
    },
    {
      "epoch": 1.37,
      "learning_rate": 4.133448873483536e-05,
      "loss": 13.3642,
      "step": 800
    },
    {
      "epoch": 1.37,
      "eval_distillation_accuracy_counterfactual": 0.5110312981015905,
      "eval_distillation_accuracy_factual": 0.8170856849666496,
      "eval_distillation_f1_counterfactual": 0.450487562425549,
      "eval_distillation_f1_factual": 0.8126073239205303,
      "eval_groundtruth_accuracy_counterfactual": 0.5107747562852745,
      "eval_groundtruth_f1_counterfactual": 0.4463667818812921,
      "eval_groundtruth_f1_factual": 0.5773838596268437,
      "eval_icace_cosine": 0.4975035488605499,
      "eval_icace_l2": 0.6422339081764221,
      "eval_icace_normdiff": 0.4249454140663147,
      "eval_loss": 10.02309799194336,
      "eval_runtime": 4.6282,
      "eval_samples_per_second": 842.226,
      "eval_steps_per_second": 6.698,
      "step": 800
    },
    {
      "epoch": 1.41,
      "learning_rate": 4.111785095320624e-05,
      "loss": 13.4977,
      "step": 820
    },
    {
      "epoch": 1.44,
      "learning_rate": 4.0901213171577124e-05,
      "loss": 13.244,
      "step": 840
    },
    {
      "epoch": 1.46,
      "eval_distillation_accuracy_counterfactual": 0.5118009235505387,
      "eval_distillation_accuracy_factual": 0.8129810159055926,
      "eval_distillation_f1_counterfactual": 0.44921794601695686,
      "eval_distillation_f1_factual": 0.8068472467220802,
      "eval_groundtruth_accuracy_counterfactual": 0.504617752693689,
      "eval_groundtruth_f1_counterfactual": 0.4373446550684491,
      "eval_groundtruth_f1_factual": 0.5739812608335962,
      "eval_icace_cosine": 0.49849507212638855,
      "eval_icace_l2": 0.6440720558166504,
      "eval_icace_normdiff": 0.4274795949459076,
      "eval_loss": 10.06081485748291,
      "eval_runtime": 4.5658,
      "eval_samples_per_second": 853.743,
      "eval_steps_per_second": 6.79,
      "step": 850
    },
    {
      "epoch": 1.48,
      "learning_rate": 4.068457538994801e-05,
      "loss": 12.8572,
      "step": 860
    },
    {
      "epoch": 1.51,
      "learning_rate": 4.04679376083189e-05,
      "loss": 13.1029,
      "step": 880
    },
    {
      "epoch": 1.54,
      "learning_rate": 4.025129982668977e-05,
      "loss": 13.132,
      "step": 900
    },
    {
      "epoch": 1.54,
      "eval_distillation_accuracy_counterfactual": 0.5189840944073884,
      "eval_distillation_accuracy_factual": 0.8196511031298102,
      "eval_distillation_f1_counterfactual": 0.4513931354514297,
      "eval_distillation_f1_factual": 0.815584477795333,
      "eval_groundtruth_accuracy_counterfactual": 0.513083632632119,
      "eval_groundtruth_f1_counterfactual": 0.4442783591730798,
      "eval_groundtruth_f1_factual": 0.5677589854844752,
      "eval_icace_cosine": 0.4927329123020172,
      "eval_icace_l2": 0.6366254091262817,
      "eval_icace_normdiff": 0.42217281460762024,
      "eval_loss": 9.926761627197266,
      "eval_runtime": 4.5591,
      "eval_samples_per_second": 854.999,
      "eval_steps_per_second": 6.8,
      "step": 900
    }
  ],
  "max_steps": 4616,
  "num_train_epochs": 8,
  "total_flos": 135969650611200.0,
  "trial_name": null,
  "trial_params": null
}