{
  "best_metric": 0.36184510588645935,
  "best_model_checkpoint": "model_output/e2e_opentable_5_way__inclusive__19684-shot__seed-42__roberta-base/checkpoint-750",
  "epoch": 4.87012987012987,
  "global_step": 750,
  "is_hyper_param_search": false,
  "is_local_process_zero": true,
  "is_world_process_zero": true,
  "log_history": [
    {
      "epoch": 0.13,
      "learning_rate": 4.9783362218370885e-05,
      "loss": 12.7993,
      "step": 20
    },
    {
      "epoch": 0.26,
      "learning_rate": 4.956672443674177e-05,
      "loss": 9.661,
      "step": 40
    },
    {
      "epoch": 0.32,
      "eval_distillation_accuracy_counterfactual": 0.5356593124679323,
      "eval_distillation_accuracy_factual": 0.7821959979476655,
      "eval_distillation_f1_counterfactual": 0.49056278645130524,
      "eval_distillation_f1_factual": 0.7607404659535162,
      "eval_groundtruth_accuracy_counterfactual": 0.4923037455105182,
      "eval_groundtruth_f1_counterfactual": 0.44063267805739714,
      "eval_groundtruth_f1_factual": 0.6494341569985087,
      "eval_icace_cosine": 0.43297919631004333,
      "eval_icace_l2": 0.6036356091499329,
      "eval_icace_normdiff": 0.47348305583000183,
      "eval_loss": 7.667526721954346,
      "eval_runtime": 12.9377,
      "eval_samples_per_second": 301.291,
      "eval_steps_per_second": 2.396,
      "step": 50
    },
    {
      "epoch": 0.39,
      "learning_rate": 4.935008665511265e-05,
      "loss": 8.5289,
      "step": 60
    },
    {
      "epoch": 0.52,
      "learning_rate": 4.913344887348354e-05,
      "loss": 7.4228,
      "step": 80
    },
    {
      "epoch": 0.65,
      "learning_rate": 4.891681109185442e-05,
      "loss": 6.5345,
      "step": 100
    },
    {
      "epoch": 0.65,
      "eval_distillation_accuracy_counterfactual": 0.6400718317085685,
      "eval_distillation_accuracy_factual": 0.8199076449461262,
      "eval_distillation_f1_counterfactual": 0.6023887287116597,
      "eval_distillation_f1_factual": 0.8061943464757944,
      "eval_groundtruth_accuracy_counterfactual": 0.5890200102616726,
      "eval_groundtruth_f1_counterfactual": 0.5451256022750195,
      "eval_groundtruth_f1_factual": 0.673398320776372,
      "eval_icace_cosine": 0.39177101850509644,
      "eval_icace_l2": 0.4908820688724518,
      "eval_icace_normdiff": 0.3136594295501709,
      "eval_loss": 6.298822402954102,
      "eval_runtime": 13.0823,
      "eval_samples_per_second": 297.959,
      "eval_steps_per_second": 2.37,
      "step": 100
    },
    {
      "epoch": 0.78,
      "learning_rate": 4.8700173310225307e-05,
      "loss": 6.3292,
      "step": 120
    },
    {
      "epoch": 0.91,
      "learning_rate": 4.848353552859619e-05,
      "loss": 5.8151,
      "step": 140
    },
    {
      "epoch": 0.97,
      "eval_distillation_accuracy_counterfactual": 0.64648537711647,
      "eval_distillation_accuracy_factual": 0.844792201128784,
      "eval_distillation_f1_counterfactual": 0.6093189989605211,
      "eval_distillation_f1_factual": 0.830785120452864,
      "eval_groundtruth_accuracy_counterfactual": 0.5890200102616726,
      "eval_groundtruth_f1_counterfactual": 0.5415275400851176,
      "eval_groundtruth_f1_factual": 0.6939511324529375,
      "eval_icace_cosine": 0.4003066420555115,
      "eval_icace_l2": 0.47722694277763367,
      "eval_icace_normdiff": 0.287803590297699,
      "eval_loss": 5.9511847496032715,
      "eval_runtime": 13.0483,
      "eval_samples_per_second": 298.736,
      "eval_steps_per_second": 2.376,
      "step": 150
    },
    {
      "epoch": 1.04,
      "learning_rate": 4.826689774696707e-05,
      "loss": 5.5867,
      "step": 160
    },
    {
      "epoch": 1.17,
      "learning_rate": 4.8050259965337955e-05,
      "loss": 5.2935,
      "step": 180
    },
    {
      "epoch": 1.3,
      "learning_rate": 4.7833622183708845e-05,
      "loss": 5.1949,
      "step": 200
    },
    {
      "epoch": 1.3,
      "eval_distillation_accuracy_counterfactual": 0.6559774243201641,
      "eval_distillation_accuracy_factual": 0.8460749102103643,
      "eval_distillation_f1_counterfactual": 0.6245235591434717,
      "eval_distillation_f1_factual": 0.8332957237840644,
      "eval_groundtruth_accuracy_counterfactual": 0.6021036428937917,
      "eval_groundtruth_f1_counterfactual": 0.5652103198044586,
      "eval_groundtruth_f1_factual": 0.6893514033709632,
      "eval_icace_cosine": 0.3755510449409485,
      "eval_icace_l2": 0.46113836765289307,
      "eval_icace_normdiff": 0.2758239805698395,
      "eval_loss": 5.779408931732178,
      "eval_runtime": 12.9764,
      "eval_samples_per_second": 300.392,
      "eval_steps_per_second": 2.389,
      "step": 200
    },
    {
      "epoch": 1.43,
      "learning_rate": 4.761698440207972e-05,
      "loss": 5.1505,
      "step": 220
    },
    {
      "epoch": 1.56,
      "learning_rate": 4.740034662045061e-05,
      "loss": 5.0401,
      "step": 240
    },
    {
      "epoch": 1.62,
      "eval_distillation_accuracy_counterfactual": 0.642124166239097,
      "eval_distillation_accuracy_factual": 0.8360697793740379,
      "eval_distillation_f1_counterfactual": 0.6001274813514513,
      "eval_distillation_f1_factual": 0.8234400842832201,
      "eval_groundtruth_accuracy_counterfactual": 0.5956900974858902,
      "eval_groundtruth_f1_counterfactual": 0.5433794665526103,
      "eval_groundtruth_f1_factual": 0.6779222154859431,
      "eval_icace_cosine": 0.3853909969329834,
      "eval_icace_l2": 0.4814322888851166,
      "eval_icace_normdiff": 0.28195297718048096,
      "eval_loss": 6.057939052581787,
      "eval_runtime": 12.9193,
      "eval_samples_per_second": 301.719,
      "eval_steps_per_second": 2.4,
      "step": 250
    },
    {
      "epoch": 1.69,
      "learning_rate": 4.7183708838821494e-05,
      "loss": 5.2737,
      "step": 260
    },
    {
      "epoch": 1.82,
      "learning_rate": 4.6967071057192376e-05,
      "loss": 4.709,
      "step": 280
    },
    {
      "epoch": 1.95,
      "learning_rate": 4.675043327556326e-05,
      "loss": 5.1205,
      "step": 300
    },
    {
      "epoch": 1.95,
      "eval_distillation_accuracy_counterfactual": 0.6536685479733196,
      "eval_distillation_accuracy_factual": 0.8558234992303746,
      "eval_distillation_f1_counterfactual": 0.6210189824287856,
      "eval_distillation_f1_factual": 0.8424842300771921,
      "eval_groundtruth_accuracy_counterfactual": 0.5946639302206259,
      "eval_groundtruth_f1_counterfactual": 0.5588206943074987,
      "eval_groundtruth_f1_factual": 0.7068947667463121,
      "eval_icace_cosine": 0.38478559255599976,
      "eval_icace_l2": 0.4652198255062103,
      "eval_icace_normdiff": 0.27682700753211975,
      "eval_loss": 5.684730052947998,
      "eval_runtime": 12.9473,
      "eval_samples_per_second": 301.067,
      "eval_steps_per_second": 2.394,
      "step": 300
    },
    {
      "epoch": 2.08,
      "learning_rate": 4.653379549393415e-05,
      "loss": 4.5402,
      "step": 320
    },
    {
      "epoch": 2.21,
      "learning_rate": 4.6317157712305025e-05,
      "loss": 4.2713,
      "step": 340
    },
    {
      "epoch": 2.27,
      "eval_distillation_accuracy_counterfactual": 0.6623909697280657,
      "eval_distillation_accuracy_factual": 0.8422267829656234,
      "eval_distillation_f1_counterfactual": 0.629941171815943,
      "eval_distillation_f1_factual": 0.8301058383190201,
      "eval_groundtruth_accuracy_counterfactual": 0.6139045664443304,
      "eval_groundtruth_f1_counterfactual": 0.5770124192575252,
      "eval_groundtruth_f1_factual": 0.6852970353343062,
      "eval_icace_cosine": 0.38777029514312744,
      "eval_icace_l2": 0.4553129971027374,
      "eval_icace_normdiff": 0.2668167054653168,
      "eval_loss": 5.9077653884887695,
      "eval_runtime": 13.1474,
      "eval_samples_per_second": 296.483,
      "eval_steps_per_second": 2.358,
      "step": 350
    },
    {
      "epoch": 2.34,
      "learning_rate": 4.6100519930675915e-05,
      "loss": 4.3929,
      "step": 360
    },
    {
      "epoch": 2.47,
      "learning_rate": 4.58838821490468e-05,
      "loss": 4.3321,
      "step": 380
    },
    {
      "epoch": 2.6,
      "learning_rate": 4.566724436741768e-05,
      "loss": 4.3156,
      "step": 400
    },
    {
      "epoch": 2.6,
      "eval_distillation_accuracy_counterfactual": 0.6706003078501795,
      "eval_distillation_accuracy_factual": 0.8683940482298614,
      "eval_distillation_f1_counterfactual": 0.6377329494351492,
      "eval_distillation_f1_factual": 0.8603493366805927,
      "eval_groundtruth_accuracy_counterfactual": 0.6131349409953822,
      "eval_groundtruth_f1_counterfactual": 0.576058451153896,
      "eval_groundtruth_f1_factual": 0.7079048939472437,
      "eval_icace_cosine": 0.37668994069099426,
      "eval_icace_l2": 0.446246474981308,
      "eval_icace_normdiff": 0.26775261759757996,
      "eval_loss": 5.3354644775390625,
      "eval_runtime": 12.9883,
      "eval_samples_per_second": 300.117,
      "eval_steps_per_second": 2.387,
      "step": 400
    },
    {
      "epoch": 2.73,
      "learning_rate": 4.5450606585788563e-05,
      "loss": 4.1218,
      "step": 420
    },
    {
      "epoch": 2.86,
      "learning_rate": 4.5233968804159446e-05,
      "loss": 3.951,
      "step": 440
    },
    {
      "epoch": 2.92,
      "eval_distillation_accuracy_counterfactual": 0.6782965623396614,
      "eval_distillation_accuracy_factual": 0.8550538737814264,
      "eval_distillation_f1_counterfactual": 0.6420700045289753,
      "eval_distillation_f1_factual": 0.838841980355596,
      "eval_groundtruth_accuracy_counterfactual": 0.6308363263211904,
      "eval_groundtruth_f1_counterfactual": 0.5946005250482415,
      "eval_groundtruth_f1_factual": 0.7064437054676487,
      "eval_icace_cosine": 0.3774036169052124,
      "eval_icace_l2": 0.43422719836235046,
      "eval_icace_normdiff": 0.26325178146362305,
      "eval_loss": 5.460549831390381,
      "eval_runtime": 12.8049,
      "eval_samples_per_second": 304.415,
      "eval_steps_per_second": 2.421,
      "step": 450
    },
    {
      "epoch": 2.99,
      "learning_rate": 4.501733102253033e-05,
      "loss": 4.1292,
      "step": 460
    },
    {
      "epoch": 3.12,
      "learning_rate": 4.480069324090121e-05,
      "loss": 3.6114,
      "step": 480
    },
    {
      "epoch": 3.25,
      "learning_rate": 4.45840554592721e-05,
      "loss": 3.6312,
      "step": 500
    },
    {
      "epoch": 3.25,
      "eval_distillation_accuracy_counterfactual": 0.6672652642380708,
      "eval_distillation_accuracy_factual": 0.8573627501282709,
      "eval_distillation_f1_counterfactual": 0.6349827827156049,
      "eval_distillation_f1_factual": 0.8441154905674914,
      "eval_groundtruth_accuracy_counterfactual": 0.612365315546434,
      "eval_groundtruth_f1_counterfactual": 0.5748755221946082,
      "eval_groundtruth_f1_factual": 0.7053883325563638,
      "eval_icace_cosine": 0.3803243041038513,
      "eval_icace_l2": 0.4419807195663452,
      "eval_icace_normdiff": 0.25063320994377136,
      "eval_loss": 5.637600421905518,
      "eval_runtime": 12.9083,
      "eval_samples_per_second": 301.977,
      "eval_steps_per_second": 2.402,
      "step": 500
    },
    {
      "epoch": 3.38,
      "learning_rate": 4.436741767764298e-05,
      "loss": 3.5906,
      "step": 520
    },
    {
      "epoch": 3.51,
      "learning_rate": 4.415077989601387e-05,
      "loss": 3.5046,
      "step": 540
    },
    {
      "epoch": 3.57,
      "eval_distillation_accuracy_counterfactual": 0.681118522319138,
      "eval_distillation_accuracy_factual": 0.8373524884556183,
      "eval_distillation_f1_counterfactual": 0.6456007756070636,
      "eval_distillation_f1_factual": 0.8262948785728061,
      "eval_groundtruth_accuracy_counterfactual": 0.6231400718317086,
      "eval_groundtruth_f1_counterfactual": 0.5852566556669566,
      "eval_groundtruth_f1_factual": 0.7081753089584231,
      "eval_icace_cosine": 0.3782615661621094,
      "eval_icace_l2": 0.4349997341632843,
      "eval_icace_normdiff": 0.260315477848053,
      "eval_loss": 5.667562007904053,
      "eval_runtime": 13.0477,
      "eval_samples_per_second": 298.75,
      "eval_steps_per_second": 2.376,
      "step": 550
    },
    {
      "epoch": 3.64,
      "learning_rate": 4.393414211438475e-05,
      "loss": 3.4758,
      "step": 560
    },
    {
      "epoch": 3.77,
      "learning_rate": 4.371750433275563e-05,
      "loss": 3.4736,
      "step": 580
    },
    {
      "epoch": 3.9,
      "learning_rate": 4.3500866551126516e-05,
      "loss": 3.5078,
      "step": 600
    },
    {
      "epoch": 3.9,
      "eval_distillation_accuracy_counterfactual": 0.6775269368907132,
      "eval_distillation_accuracy_factual": 0.8671113391482812,
      "eval_distillation_f1_counterfactual": 0.6443129742790921,
      "eval_distillation_f1_factual": 0.856068235660411,
      "eval_groundtruth_accuracy_counterfactual": 0.6180092355053873,
      "eval_groundtruth_f1_counterfactual": 0.5788738103142942,
      "eval_groundtruth_f1_factual": 0.7201005920079917,
      "eval_icace_cosine": 0.3698969781398773,
      "eval_icace_l2": 0.44297319650650024,
      "eval_icace_normdiff": 0.2465381622314453,
      "eval_loss": 5.7820048332214355,
      "eval_runtime": 12.9334,
      "eval_samples_per_second": 301.39,
      "eval_steps_per_second": 2.397,
      "step": 600
    },
    {
      "epoch": 4.03,
      "learning_rate": 4.3284228769497406e-05,
      "loss": 3.2757,
      "step": 620
    },
    {
      "epoch": 4.16,
      "learning_rate": 4.306759098786828e-05,
      "loss": 3.0079,
      "step": 640
    },
    {
      "epoch": 4.22,
      "eval_distillation_accuracy_counterfactual": 0.6739353514622883,
      "eval_distillation_accuracy_factual": 0.8604412519240636,
      "eval_distillation_f1_counterfactual": 0.6432037881949733,
      "eval_distillation_f1_factual": 0.847959783260203,
      "eval_groundtruth_accuracy_counterfactual": 0.6241662390969728,
      "eval_groundtruth_f1_counterfactual": 0.5908586556615842,
      "eval_groundtruth_f1_factual": 0.7152174183342359,
      "eval_icace_cosine": 0.37243351340293884,
      "eval_icace_l2": 0.4401131570339203,
      "eval_icace_normdiff": 0.24781018495559692,
      "eval_loss": 5.813790321350098,
      "eval_runtime": 12.8545,
      "eval_samples_per_second": 303.241,
      "eval_steps_per_second": 2.412,
      "step": 650
    },
    {
      "epoch": 4.29,
      "learning_rate": 4.285095320623917e-05,
      "loss": 2.8728,
      "step": 660
    },
    {
      "epoch": 4.42,
      "learning_rate": 4.2634315424610055e-05,
      "loss": 3.0153,
      "step": 680
    },
    {
      "epoch": 4.55,
      "learning_rate": 4.241767764298094e-05,
      "loss": 2.9833,
      "step": 700
    },
    {
      "epoch": 4.55,
      "eval_distillation_accuracy_counterfactual": 0.6847101077475628,
      "eval_distillation_accuracy_factual": 0.861723961005644,
      "eval_distillation_f1_counterfactual": 0.655880561232892,
      "eval_distillation_f1_factual": 0.8544377792240528,
      "eval_groundtruth_accuracy_counterfactual": 0.6169830682401232,
      "eval_groundtruth_f1_counterfactual": 0.5835509049462576,
      "eval_groundtruth_f1_factual": 0.7078647076565944,
      "eval_icace_cosine": 0.3621787428855896,
      "eval_icace_l2": 0.4256499409675598,
      "eval_icace_normdiff": 0.24053125083446503,
      "eval_loss": 5.67288875579834,
      "eval_runtime": 12.8728,
      "eval_samples_per_second": 302.81,
      "eval_steps_per_second": 2.408,
      "step": 700
    },
    {
      "epoch": 4.68,
      "learning_rate": 4.220103986135182e-05,
      "loss": 3.119,
      "step": 720
    },
    {
      "epoch": 4.81,
      "learning_rate": 4.198440207972271e-05,
      "loss": 3.0303,
      "step": 740
    },
    {
      "epoch": 4.87,
      "eval_distillation_accuracy_counterfactual": 0.6813750641354541,
      "eval_distillation_accuracy_factual": 0.853771164699846,
      "eval_distillation_f1_counterfactual": 0.6556449925274147,
      "eval_distillation_f1_factual": 0.843877666601483,
      "eval_groundtruth_accuracy_counterfactual": 0.6241662390969728,
      "eval_groundtruth_f1_counterfactual": 0.5937900551226225,
      "eval_groundtruth_f1_factual": 0.7217526332733881,
      "eval_icace_cosine": 0.36184510588645935,
      "eval_icace_l2": 0.43167001008987427,
      "eval_icace_normdiff": 0.24829088151454926,
      "eval_loss": 5.614396572113037,
      "eval_runtime": 13.0379,
      "eval_samples_per_second": 298.974,
      "eval_steps_per_second": 2.378,
      "step": 750
    }
  ],
  "max_steps": 4616,
  "num_train_epochs": 30,
  "total_flos": 6307468111147008.0,
  "trial_name": null,
  "trial_params": null
}