KarelDO's picture
commit files to HF hub
bed49ff
{
"best_metric": 0.4564709961414337,
"best_model_checkpoint": "model_output/e2e_opentable_5_way__approximate__0-shot__seed-66__bert-base-uncased/checkpoint-1400",
"epoch": 2.4013722126929675,
"global_step": 1400,
"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": 11.2983,
"step": 20
},
{
"epoch": 0.07,
"learning_rate": 4.956672443674177e-05,
"loss": 8.8977,
"step": 40
},
{
"epoch": 0.09,
"eval_distillation_accuracy_counterfactual": 0.5030785017957927,
"eval_distillation_accuracy_factual": 0.870702924576706,
"eval_distillation_f1_counterfactual": 0.36918123521791174,
"eval_distillation_f1_factual": 0.8518990750712729,
"eval_groundtruth_accuracy_counterfactual": 0.4763981528989225,
"eval_groundtruth_f1_counterfactual": 0.3533145590655197,
"eval_groundtruth_f1_factual": 0.6552258773293613,
"eval_icace_cosine": 0.4796179533004761,
"eval_icace_l2": 0.6570112705230713,
"eval_icace_normdiff": 0.4324696958065033,
"eval_loss": 6.444849967956543,
"eval_runtime": 22.056,
"eval_samples_per_second": 176.732,
"eval_steps_per_second": 2.766,
"step": 50
},
{
"epoch": 0.1,
"learning_rate": 4.935008665511265e-05,
"loss": 8.1157,
"step": 60
},
{
"epoch": 0.14,
"learning_rate": 4.913344887348354e-05,
"loss": 7.8853,
"step": 80
},
{
"epoch": 0.17,
"learning_rate": 4.891681109185442e-05,
"loss": 7.6564,
"step": 100
},
{
"epoch": 0.17,
"eval_distillation_accuracy_counterfactual": 0.5312981015905592,
"eval_distillation_accuracy_factual": 0.8776295536172396,
"eval_distillation_f1_counterfactual": 0.4177716134310396,
"eval_distillation_f1_factual": 0.8606344639589892,
"eval_groundtruth_accuracy_counterfactual": 0.5,
"eval_groundtruth_f1_counterfactual": 0.39828651286680133,
"eval_groundtruth_f1_factual": 0.6882338232741171,
"eval_icace_cosine": 0.4674900770187378,
"eval_icace_l2": 0.6195012331008911,
"eval_icace_normdiff": 0.3801978528499603,
"eval_loss": 6.348506450653076,
"eval_runtime": 22.6417,
"eval_samples_per_second": 172.16,
"eval_steps_per_second": 2.694,
"step": 100
},
{
"epoch": 0.21,
"learning_rate": 4.8700173310225307e-05,
"loss": 7.5574,
"step": 120
},
{
"epoch": 0.24,
"learning_rate": 4.848353552859619e-05,
"loss": 7.2643,
"step": 140
},
{
"epoch": 0.26,
"eval_distillation_accuracy_counterfactual": 0.5287326834273987,
"eval_distillation_accuracy_factual": 0.8694202154951257,
"eval_distillation_f1_counterfactual": 0.4118396244303133,
"eval_distillation_f1_factual": 0.8599084137519469,
"eval_groundtruth_accuracy_counterfactual": 0.5141097998973833,
"eval_groundtruth_f1_counterfactual": 0.4079251742601512,
"eval_groundtruth_f1_factual": 0.6810018997954198,
"eval_icace_cosine": 0.46454283595085144,
"eval_icace_l2": 0.6308194398880005,
"eval_icace_normdiff": 0.40617796778678894,
"eval_loss": 6.1462531089782715,
"eval_runtime": 22.1649,
"eval_samples_per_second": 175.863,
"eval_steps_per_second": 2.752,
"step": 150
},
{
"epoch": 0.27,
"learning_rate": 4.826689774696707e-05,
"loss": 7.3808,
"step": 160
},
{
"epoch": 0.31,
"learning_rate": 4.8050259965337955e-05,
"loss": 7.1507,
"step": 180
},
{
"epoch": 0.34,
"learning_rate": 4.7833622183708845e-05,
"loss": 7.0277,
"step": 200
},
{
"epoch": 0.34,
"eval_distillation_accuracy_counterfactual": 0.521806054386865,
"eval_distillation_accuracy_factual": 0.8594150846587993,
"eval_distillation_f1_counterfactual": 0.43996859803992977,
"eval_distillation_f1_factual": 0.8377228086695208,
"eval_groundtruth_accuracy_counterfactual": 0.5048742945100051,
"eval_groundtruth_f1_counterfactual": 0.42823671858777795,
"eval_groundtruth_f1_factual": 0.6661417838076318,
"eval_icace_cosine": 0.4762553572654724,
"eval_icace_l2": 0.6346240043640137,
"eval_icace_normdiff": 0.38476699590682983,
"eval_loss": 6.443660259246826,
"eval_runtime": 22.3822,
"eval_samples_per_second": 174.156,
"eval_steps_per_second": 2.725,
"step": 200
},
{
"epoch": 0.38,
"learning_rate": 4.761698440207972e-05,
"loss": 6.9694,
"step": 220
},
{
"epoch": 0.41,
"learning_rate": 4.740034662045061e-05,
"loss": 7.0702,
"step": 240
},
{
"epoch": 0.43,
"eval_distillation_accuracy_counterfactual": 0.5307850179579271,
"eval_distillation_accuracy_factual": 0.8640328373524885,
"eval_distillation_f1_counterfactual": 0.43581597774498526,
"eval_distillation_f1_factual": 0.8511773812631599,
"eval_groundtruth_accuracy_counterfactual": 0.5082093381221139,
"eval_groundtruth_f1_counterfactual": 0.4226816426807245,
"eval_groundtruth_f1_factual": 0.6714755980104481,
"eval_icace_cosine": 0.4607926905155182,
"eval_icace_l2": 0.6025251150131226,
"eval_icace_normdiff": 0.366618812084198,
"eval_loss": 6.052988052368164,
"eval_runtime": 22.0965,
"eval_samples_per_second": 176.408,
"eval_steps_per_second": 2.761,
"step": 250
},
{
"epoch": 0.45,
"learning_rate": 4.7183708838821494e-05,
"loss": 6.8827,
"step": 260
},
{
"epoch": 0.48,
"learning_rate": 4.6967071057192376e-05,
"loss": 6.8895,
"step": 280
},
{
"epoch": 0.51,
"learning_rate": 4.675043327556326e-05,
"loss": 6.744,
"step": 300
},
{
"epoch": 0.51,
"eval_distillation_accuracy_counterfactual": 0.5377116469984607,
"eval_distillation_accuracy_factual": 0.8612108773730118,
"eval_distillation_f1_counterfactual": 0.4223090065654679,
"eval_distillation_f1_factual": 0.8456621535049905,
"eval_groundtruth_accuracy_counterfactual": 0.5143663417136993,
"eval_groundtruth_f1_counterfactual": 0.40663553727210966,
"eval_groundtruth_f1_factual": 0.6775861228321817,
"eval_icace_cosine": 0.46002256870269775,
"eval_icace_l2": 0.5889779329299927,
"eval_icace_normdiff": 0.3519256114959717,
"eval_loss": 5.902462005615234,
"eval_runtime": 22.3125,
"eval_samples_per_second": 174.7,
"eval_steps_per_second": 2.734,
"step": 300
},
{
"epoch": 0.55,
"learning_rate": 4.653379549393415e-05,
"loss": 6.6897,
"step": 320
},
{
"epoch": 0.58,
"learning_rate": 4.6317157712305025e-05,
"loss": 6.7666,
"step": 340
},
{
"epoch": 0.6,
"eval_distillation_accuracy_counterfactual": 0.5548999486916367,
"eval_distillation_accuracy_factual": 0.8499230374551052,
"eval_distillation_f1_counterfactual": 0.49580886864510487,
"eval_distillation_f1_factual": 0.8361446774720787,
"eval_groundtruth_accuracy_counterfactual": 0.538737814263725,
"eval_groundtruth_f1_counterfactual": 0.4813622017124305,
"eval_groundtruth_f1_factual": 0.674748187131342,
"eval_icace_cosine": 0.46281909942626953,
"eval_icace_l2": 0.5994406342506409,
"eval_icace_normdiff": 0.3744681477546692,
"eval_loss": 5.933077335357666,
"eval_runtime": 22.1654,
"eval_samples_per_second": 175.86,
"eval_steps_per_second": 2.752,
"step": 350
},
{
"epoch": 0.62,
"learning_rate": 4.6100519930675915e-05,
"loss": 6.5721,
"step": 360
},
{
"epoch": 0.65,
"learning_rate": 4.58838821490468e-05,
"loss": 6.7086,
"step": 380
},
{
"epoch": 0.69,
"learning_rate": 4.566724436741768e-05,
"loss": 6.5311,
"step": 400
},
{
"epoch": 0.69,
"eval_distillation_accuracy_counterfactual": 0.5556695741405849,
"eval_distillation_accuracy_factual": 0.8724987172909184,
"eval_distillation_f1_counterfactual": 0.456210753567529,
"eval_distillation_f1_factual": 0.8597691779856307,
"eval_groundtruth_accuracy_counterfactual": 0.5333504361210878,
"eval_groundtruth_f1_counterfactual": 0.444792193959889,
"eval_groundtruth_f1_factual": 0.6708274089754076,
"eval_icace_cosine": 0.45892879366874695,
"eval_icace_l2": 0.5904973745346069,
"eval_icace_normdiff": 0.3636997938156128,
"eval_loss": 5.868222236633301,
"eval_runtime": 22.3242,
"eval_samples_per_second": 174.609,
"eval_steps_per_second": 2.732,
"step": 400
},
{
"epoch": 0.72,
"learning_rate": 4.5450606585788563e-05,
"loss": 6.5486,
"step": 420
},
{
"epoch": 0.75,
"learning_rate": 4.5233968804159446e-05,
"loss": 6.4615,
"step": 440
},
{
"epoch": 0.77,
"eval_distillation_accuracy_counterfactual": 0.5428424833247819,
"eval_distillation_accuracy_factual": 0.8601847101077476,
"eval_distillation_f1_counterfactual": 0.45475006277273583,
"eval_distillation_f1_factual": 0.8490794790111732,
"eval_groundtruth_accuracy_counterfactual": 0.5223191380194971,
"eval_groundtruth_f1_counterfactual": 0.44181890062982204,
"eval_groundtruth_f1_factual": 0.6911106055165974,
"eval_icace_cosine": 0.4593462646007538,
"eval_icace_l2": 0.5996768474578857,
"eval_icace_normdiff": 0.3729654550552368,
"eval_loss": 5.918853282928467,
"eval_runtime": 22.1691,
"eval_samples_per_second": 175.83,
"eval_steps_per_second": 2.752,
"step": 450
},
{
"epoch": 0.79,
"learning_rate": 4.501733102253033e-05,
"loss": 6.4495,
"step": 460
},
{
"epoch": 0.82,
"learning_rate": 4.480069324090121e-05,
"loss": 6.6214,
"step": 480
},
{
"epoch": 0.86,
"learning_rate": 4.45840554592721e-05,
"loss": 6.5091,
"step": 500
},
{
"epoch": 0.86,
"eval_distillation_accuracy_counterfactual": 0.5528476141611083,
"eval_distillation_accuracy_factual": 0.8606977937403797,
"eval_distillation_f1_counterfactual": 0.42193207949843314,
"eval_distillation_f1_factual": 0.8449657823376556,
"eval_groundtruth_accuracy_counterfactual": 0.5318111852231914,
"eval_groundtruth_f1_counterfactual": 0.4123797131290104,
"eval_groundtruth_f1_factual": 0.6754423477994138,
"eval_icace_cosine": 0.4626261293888092,
"eval_icace_l2": 0.6023271083831787,
"eval_icace_normdiff": 0.37480491399765015,
"eval_loss": 5.962100028991699,
"eval_runtime": 22.2411,
"eval_samples_per_second": 175.261,
"eval_steps_per_second": 2.743,
"step": 500
},
{
"epoch": 0.89,
"learning_rate": 4.436741767764298e-05,
"loss": 6.5299,
"step": 520
},
{
"epoch": 0.93,
"learning_rate": 4.415077989601387e-05,
"loss": 6.5688,
"step": 540
},
{
"epoch": 0.94,
"eval_distillation_accuracy_counterfactual": 0.5377116469984607,
"eval_distillation_accuracy_factual": 0.8583889173935352,
"eval_distillation_f1_counterfactual": 0.46729782076302406,
"eval_distillation_f1_factual": 0.8469926857732132,
"eval_groundtruth_accuracy_counterfactual": 0.5105182144689584,
"eval_groundtruth_f1_counterfactual": 0.44493052070683115,
"eval_groundtruth_f1_factual": 0.6879462817233734,
"eval_icace_cosine": 0.4675835072994232,
"eval_icace_l2": 0.6114413738250732,
"eval_icace_normdiff": 0.376953661441803,
"eval_loss": 6.076689720153809,
"eval_runtime": 22.1309,
"eval_samples_per_second": 176.134,
"eval_steps_per_second": 2.756,
"step": 550
},
{
"epoch": 0.96,
"learning_rate": 4.393414211438475e-05,
"loss": 6.5147,
"step": 560
},
{
"epoch": 0.99,
"learning_rate": 4.371750433275563e-05,
"loss": 6.5651,
"step": 580
},
{
"epoch": 1.03,
"learning_rate": 4.3500866551126516e-05,
"loss": 6.2884,
"step": 600
},
{
"epoch": 1.03,
"eval_distillation_accuracy_counterfactual": 0.5315546434068753,
"eval_distillation_accuracy_factual": 0.8199076449461262,
"eval_distillation_f1_counterfactual": 0.4240354476459813,
"eval_distillation_f1_factual": 0.8110143323147028,
"eval_groundtruth_accuracy_counterfactual": 0.5141097998973833,
"eval_groundtruth_f1_counterfactual": 0.4161445570845627,
"eval_groundtruth_f1_factual": 0.6701739679915534,
"eval_icace_cosine": 0.4710710346698761,
"eval_icace_l2": 0.6157806515693665,
"eval_icace_normdiff": 0.36949479579925537,
"eval_loss": 6.440975666046143,
"eval_runtime": 22.0879,
"eval_samples_per_second": 176.477,
"eval_steps_per_second": 2.762,
"step": 600
},
{
"epoch": 1.06,
"learning_rate": 4.3284228769497406e-05,
"loss": 6.3874,
"step": 620
},
{
"epoch": 1.1,
"learning_rate": 4.306759098786828e-05,
"loss": 6.2439,
"step": 640
},
{
"epoch": 1.11,
"eval_distillation_accuracy_counterfactual": 0.5572088250384812,
"eval_distillation_accuracy_factual": 0.8345305284761416,
"eval_distillation_f1_counterfactual": 0.4681040280988178,
"eval_distillation_f1_factual": 0.8289204998070453,
"eval_groundtruth_accuracy_counterfactual": 0.5428424833247819,
"eval_groundtruth_f1_counterfactual": 0.4604017658322358,
"eval_groundtruth_f1_factual": 0.6775906371456756,
"eval_icace_cosine": 0.4579065144062042,
"eval_icace_l2": 0.5784377455711365,
"eval_icace_normdiff": 0.3504405617713928,
"eval_loss": 5.936356544494629,
"eval_runtime": 22.1397,
"eval_samples_per_second": 176.064,
"eval_steps_per_second": 2.755,
"step": 650
},
{
"epoch": 1.13,
"learning_rate": 4.285095320623917e-05,
"loss": 6.2876,
"step": 660
},
{
"epoch": 1.17,
"learning_rate": 4.2634315424610055e-05,
"loss": 6.3896,
"step": 680
},
{
"epoch": 1.2,
"learning_rate": 4.241767764298094e-05,
"loss": 6.1169,
"step": 700
},
{
"epoch": 1.2,
"eval_distillation_accuracy_counterfactual": 0.5479733196511031,
"eval_distillation_accuracy_factual": 0.8129810159055926,
"eval_distillation_f1_counterfactual": 0.48600933269771956,
"eval_distillation_f1_factual": 0.7860512885448971,
"eval_groundtruth_accuracy_counterfactual": 0.5336069779374037,
"eval_groundtruth_f1_counterfactual": 0.47412523543027546,
"eval_groundtruth_f1_factual": 0.6485794965038972,
"eval_icace_cosine": 0.4659746587276459,
"eval_icace_l2": 0.5945326089859009,
"eval_icace_normdiff": 0.36559173464775085,
"eval_loss": 6.143641948699951,
"eval_runtime": 22.3066,
"eval_samples_per_second": 174.747,
"eval_steps_per_second": 2.735,
"step": 700
},
{
"epoch": 1.23,
"learning_rate": 4.220103986135182e-05,
"loss": 6.1743,
"step": 720
},
{
"epoch": 1.27,
"learning_rate": 4.198440207972271e-05,
"loss": 6.4308,
"step": 740
},
{
"epoch": 1.29,
"eval_distillation_accuracy_counterfactual": 0.5569522832221652,
"eval_distillation_accuracy_factual": 0.8193945613134941,
"eval_distillation_f1_counterfactual": 0.4527655231915041,
"eval_distillation_f1_factual": 0.8096871483467666,
"eval_groundtruth_accuracy_counterfactual": 0.5551564905079528,
"eval_groundtruth_f1_counterfactual": 0.45780758221444573,
"eval_groundtruth_f1_factual": 0.6628929940851196,
"eval_icace_cosine": 0.4578511118888855,
"eval_icace_l2": 0.5670422911643982,
"eval_icace_normdiff": 0.35481762886047363,
"eval_loss": 5.8051252365112305,
"eval_runtime": 22.076,
"eval_samples_per_second": 176.572,
"eval_steps_per_second": 2.763,
"step": 750
},
{
"epoch": 1.3,
"learning_rate": 4.1767764298093586e-05,
"loss": 6.1878,
"step": 760
},
{
"epoch": 1.34,
"learning_rate": 4.1551126516464476e-05,
"loss": 6.2818,
"step": 780
},
{
"epoch": 1.37,
"learning_rate": 4.133448873483536e-05,
"loss": 6.1387,
"step": 800
},
{
"epoch": 1.37,
"eval_distillation_accuracy_counterfactual": 0.5515649050795279,
"eval_distillation_accuracy_factual": 0.853771164699846,
"eval_distillation_f1_counterfactual": 0.46108787809547397,
"eval_distillation_f1_factual": 0.8408838366147668,
"eval_groundtruth_accuracy_counterfactual": 0.5328373524884557,
"eval_groundtruth_f1_counterfactual": 0.4439112446666075,
"eval_groundtruth_f1_factual": 0.6747449056620686,
"eval_icace_cosine": 0.45763716101646423,
"eval_icace_l2": 0.5753128528594971,
"eval_icace_normdiff": 0.3466692864894867,
"eval_loss": 5.853026866912842,
"eval_runtime": 22.2637,
"eval_samples_per_second": 175.083,
"eval_steps_per_second": 2.74,
"step": 800
},
{
"epoch": 1.41,
"learning_rate": 4.111785095320624e-05,
"loss": 6.3818,
"step": 820
},
{
"epoch": 1.44,
"learning_rate": 4.0901213171577124e-05,
"loss": 6.3433,
"step": 840
},
{
"epoch": 1.46,
"eval_distillation_accuracy_counterfactual": 0.5448948178553105,
"eval_distillation_accuracy_factual": 0.8476141611082606,
"eval_distillation_f1_counterfactual": 0.4361535399476864,
"eval_distillation_f1_factual": 0.8329766547479427,
"eval_groundtruth_accuracy_counterfactual": 0.5351462288353002,
"eval_groundtruth_f1_counterfactual": 0.42831401054324375,
"eval_groundtruth_f1_factual": 0.6756139171118364,
"eval_icace_cosine": 0.4597473442554474,
"eval_icace_l2": 0.583962619304657,
"eval_icace_normdiff": 0.3553682863712311,
"eval_loss": 6.0229010581970215,
"eval_runtime": 22.1925,
"eval_samples_per_second": 175.645,
"eval_steps_per_second": 2.749,
"step": 850
},
{
"epoch": 1.48,
"learning_rate": 4.068457538994801e-05,
"loss": 6.2674,
"step": 860
},
{
"epoch": 1.51,
"learning_rate": 4.04679376083189e-05,
"loss": 6.5452,
"step": 880
},
{
"epoch": 1.54,
"learning_rate": 4.025129982668977e-05,
"loss": 6.2361,
"step": 900
},
{
"epoch": 1.54,
"eval_distillation_accuracy_counterfactual": 0.5454079014879425,
"eval_distillation_accuracy_factual": 0.8186249358645459,
"eval_distillation_f1_counterfactual": 0.4448454069324789,
"eval_distillation_f1_factual": 0.8057071203974917,
"eval_groundtruth_accuracy_counterfactual": 0.5287326834273987,
"eval_groundtruth_f1_counterfactual": 0.42960759215150884,
"eval_groundtruth_f1_factual": 0.6670500492904399,
"eval_icace_cosine": 0.46398869156837463,
"eval_icace_l2": 0.5982965230941772,
"eval_icace_normdiff": 0.3660006821155548,
"eval_loss": 6.118747234344482,
"eval_runtime": 22.1991,
"eval_samples_per_second": 175.593,
"eval_steps_per_second": 2.748,
"step": 900
},
{
"epoch": 1.58,
"learning_rate": 4.003466204506066e-05,
"loss": 6.3136,
"step": 920
},
{
"epoch": 1.61,
"learning_rate": 3.9818024263431546e-05,
"loss": 6.3279,
"step": 940
},
{
"epoch": 1.63,
"eval_distillation_accuracy_counterfactual": 0.5451513596716264,
"eval_distillation_accuracy_factual": 0.8355566957414059,
"eval_distillation_f1_counterfactual": 0.4168175870039869,
"eval_distillation_f1_factual": 0.8267012706170487,
"eval_groundtruth_accuracy_counterfactual": 0.5274499743458184,
"eval_groundtruth_f1_counterfactual": 0.41075982637956343,
"eval_groundtruth_f1_factual": 0.6781751045050392,
"eval_icace_cosine": 0.45960184931755066,
"eval_icace_l2": 0.5789852142333984,
"eval_icace_normdiff": 0.35161492228507996,
"eval_loss": 5.898173809051514,
"eval_runtime": 22.1328,
"eval_samples_per_second": 176.119,
"eval_steps_per_second": 2.756,
"step": 950
},
{
"epoch": 1.65,
"learning_rate": 3.960138648180243e-05,
"loss": 6.349,
"step": 960
},
{
"epoch": 1.68,
"learning_rate": 3.938474870017331e-05,
"loss": 6.2648,
"step": 980
},
{
"epoch": 1.72,
"learning_rate": 3.91681109185442e-05,
"loss": 6.307,
"step": 1000
},
{
"epoch": 1.72,
"eval_distillation_accuracy_counterfactual": 0.547460236018471,
"eval_distillation_accuracy_factual": 0.8311954848640328,
"eval_distillation_f1_counterfactual": 0.4400284488718956,
"eval_distillation_f1_factual": 0.819341898820874,
"eval_groundtruth_accuracy_counterfactual": 0.5328373524884557,
"eval_groundtruth_f1_counterfactual": 0.431004456933043,
"eval_groundtruth_f1_factual": 0.6724529205225047,
"eval_icace_cosine": 0.4630897343158722,
"eval_icace_l2": 0.5932038426399231,
"eval_icace_normdiff": 0.3649563491344452,
"eval_loss": 5.96807861328125,
"eval_runtime": 22.2743,
"eval_samples_per_second": 175.0,
"eval_steps_per_second": 2.739,
"step": 1000
},
{
"epoch": 1.75,
"learning_rate": 3.895147313691508e-05,
"loss": 6.3078,
"step": 1020
},
{
"epoch": 1.78,
"learning_rate": 3.873483535528597e-05,
"loss": 6.299,
"step": 1040
},
{
"epoch": 1.8,
"eval_distillation_accuracy_counterfactual": 0.538481272447409,
"eval_distillation_accuracy_factual": 0.8358132375577219,
"eval_distillation_f1_counterfactual": 0.43944572973337176,
"eval_distillation_f1_factual": 0.8137606391644893,
"eval_groundtruth_accuracy_counterfactual": 0.5210364289379169,
"eval_groundtruth_f1_counterfactual": 0.4266673104908225,
"eval_groundtruth_f1_factual": 0.6679874290517764,
"eval_icace_cosine": 0.4646293520927429,
"eval_icace_l2": 0.5929473042488098,
"eval_icace_normdiff": 0.35142672061920166,
"eval_loss": 6.2012529373168945,
"eval_runtime": 22.1799,
"eval_samples_per_second": 175.745,
"eval_steps_per_second": 2.75,
"step": 1050
},
{
"epoch": 1.82,
"learning_rate": 3.851819757365685e-05,
"loss": 6.1563,
"step": 1060
},
{
"epoch": 1.85,
"learning_rate": 3.830155979202773e-05,
"loss": 6.2483,
"step": 1080
},
{
"epoch": 1.89,
"learning_rate": 3.8084922010398616e-05,
"loss": 6.2315,
"step": 1100
},
{
"epoch": 1.89,
"eval_distillation_accuracy_counterfactual": 0.5502821959979477,
"eval_distillation_accuracy_factual": 0.8283735248845562,
"eval_distillation_f1_counterfactual": 0.4492053235764312,
"eval_distillation_f1_factual": 0.8130071646506035,
"eval_groundtruth_accuracy_counterfactual": 0.5438686505900462,
"eval_groundtruth_f1_counterfactual": 0.4459935654992118,
"eval_groundtruth_f1_factual": 0.6759199159298241,
"eval_icace_cosine": 0.4584900140762329,
"eval_icace_l2": 0.5902418494224548,
"eval_icace_normdiff": 0.3675350546836853,
"eval_loss": 6.030585765838623,
"eval_runtime": 22.2302,
"eval_samples_per_second": 175.347,
"eval_steps_per_second": 2.744,
"step": 1100
},
{
"epoch": 1.92,
"learning_rate": 3.78682842287695e-05,
"loss": 6.0839,
"step": 1120
},
{
"epoch": 1.96,
"learning_rate": 3.765164644714038e-05,
"loss": 6.4395,
"step": 1140
},
{
"epoch": 1.97,
"eval_distillation_accuracy_counterfactual": 0.5420728578758338,
"eval_distillation_accuracy_factual": 0.8350436121087738,
"eval_distillation_f1_counterfactual": 0.44687319238313783,
"eval_distillation_f1_factual": 0.8227278561647481,
"eval_groundtruth_accuracy_counterfactual": 0.5279630579784504,
"eval_groundtruth_f1_counterfactual": 0.4362421678695392,
"eval_groundtruth_f1_factual": 0.6689379744755434,
"eval_icace_cosine": 0.4657202959060669,
"eval_icace_l2": 0.5986185073852539,
"eval_icace_normdiff": 0.37084171175956726,
"eval_loss": 6.077243328094482,
"eval_runtime": 22.1,
"eval_samples_per_second": 176.38,
"eval_steps_per_second": 2.76,
"step": 1150
},
{
"epoch": 1.99,
"learning_rate": 3.7435008665511264e-05,
"loss": 6.4188,
"step": 1160
},
{
"epoch": 2.02,
"learning_rate": 3.7218370883882154e-05,
"loss": 6.2773,
"step": 1180
},
{
"epoch": 2.06,
"learning_rate": 3.700173310225303e-05,
"loss": 6.0979,
"step": 1200
},
{
"epoch": 2.06,
"eval_distillation_accuracy_counterfactual": 0.5554130323242689,
"eval_distillation_accuracy_factual": 0.8304258594150846,
"eval_distillation_f1_counterfactual": 0.4772175429149318,
"eval_distillation_f1_factual": 0.808134058844989,
"eval_groundtruth_accuracy_counterfactual": 0.5389943560800411,
"eval_groundtruth_f1_counterfactual": 0.4650970297326843,
"eval_groundtruth_f1_factual": 0.6534003932896711,
"eval_icace_cosine": 0.4595852196216583,
"eval_icace_l2": 0.5800026059150696,
"eval_icace_normdiff": 0.3539329171180725,
"eval_loss": 5.965435981750488,
"eval_runtime": 22.2847,
"eval_samples_per_second": 174.919,
"eval_steps_per_second": 2.737,
"step": 1200
},
{
"epoch": 2.09,
"learning_rate": 3.678509532062392e-05,
"loss": 6.067,
"step": 1220
},
{
"epoch": 2.13,
"learning_rate": 3.65684575389948e-05,
"loss": 6.2014,
"step": 1240
},
{
"epoch": 2.14,
"eval_distillation_accuracy_counterfactual": 0.5477167778347871,
"eval_distillation_accuracy_factual": 0.827860441251924,
"eval_distillation_f1_counterfactual": 0.45873512192151045,
"eval_distillation_f1_factual": 0.8182707512505438,
"eval_groundtruth_accuracy_counterfactual": 0.5371985633658286,
"eval_groundtruth_f1_counterfactual": 0.45174422882383675,
"eval_groundtruth_f1_factual": 0.6686107185288309,
"eval_icace_cosine": 0.4629949629306793,
"eval_icace_l2": 0.5902252197265625,
"eval_icace_normdiff": 0.36104583740234375,
"eval_loss": 6.057925224304199,
"eval_runtime": 22.207,
"eval_samples_per_second": 175.53,
"eval_steps_per_second": 2.747,
"step": 1250
},
{
"epoch": 2.16,
"learning_rate": 3.6351819757365686e-05,
"loss": 6.0228,
"step": 1260
},
{
"epoch": 2.2,
"learning_rate": 3.613518197573657e-05,
"loss": 6.077,
"step": 1280
},
{
"epoch": 2.23,
"learning_rate": 3.591854419410746e-05,
"loss": 6.0282,
"step": 1300
},
{
"epoch": 2.23,
"eval_distillation_accuracy_counterfactual": 0.5507952796305798,
"eval_distillation_accuracy_factual": 0.8165726013340174,
"eval_distillation_f1_counterfactual": 0.4631377949704046,
"eval_distillation_f1_factual": 0.8093111197445927,
"eval_groundtruth_accuracy_counterfactual": 0.5284761416110826,
"eval_groundtruth_f1_counterfactual": 0.4474398103025621,
"eval_groundtruth_f1_factual": 0.669141158800666,
"eval_icace_cosine": 0.4620840549468994,
"eval_icace_l2": 0.5971253514289856,
"eval_icace_normdiff": 0.3704546391963959,
"eval_loss": 6.050764083862305,
"eval_runtime": 22.2574,
"eval_samples_per_second": 175.133,
"eval_steps_per_second": 2.741,
"step": 1300
},
{
"epoch": 2.26,
"learning_rate": 3.5701906412478334e-05,
"loss": 6.1973,
"step": 1320
},
{
"epoch": 2.3,
"learning_rate": 3.5485268630849224e-05,
"loss": 6.0877,
"step": 1340
},
{
"epoch": 2.32,
"eval_distillation_accuracy_counterfactual": 0.5064135454079015,
"eval_distillation_accuracy_factual": 0.8309389430477168,
"eval_distillation_f1_counterfactual": 0.4306802598777498,
"eval_distillation_f1_factual": 0.8171042534137397,
"eval_groundtruth_accuracy_counterfactual": 0.48460749102103645,
"eval_groundtruth_f1_counterfactual": 0.4162277530517664,
"eval_groundtruth_f1_factual": 0.6732278788009838,
"eval_icace_cosine": 0.47389736771583557,
"eval_icace_l2": 0.6207495331764221,
"eval_icace_normdiff": 0.3774127960205078,
"eval_loss": 6.461216926574707,
"eval_runtime": 22.063,
"eval_samples_per_second": 176.676,
"eval_steps_per_second": 2.765,
"step": 1350
},
{
"epoch": 2.33,
"learning_rate": 3.526863084922011e-05,
"loss": 6.0128,
"step": 1360
},
{
"epoch": 2.37,
"learning_rate": 3.505199306759099e-05,
"loss": 6.0307,
"step": 1380
},
{
"epoch": 2.4,
"learning_rate": 3.483535528596187e-05,
"loss": 6.1284,
"step": 1400
},
{
"epoch": 2.4,
"eval_distillation_accuracy_counterfactual": 0.5510518214468958,
"eval_distillation_accuracy_factual": 0.8055413032324269,
"eval_distillation_f1_counterfactual": 0.4448362647873446,
"eval_distillation_f1_factual": 0.7931010998998913,
"eval_groundtruth_accuracy_counterfactual": 0.5266803488968702,
"eval_groundtruth_f1_counterfactual": 0.4296197215462649,
"eval_groundtruth_f1_factual": 0.6675776163244647,
"eval_icace_cosine": 0.4564709961414337,
"eval_icace_l2": 0.5935506820678711,
"eval_icace_normdiff": 0.36554762721061707,
"eval_loss": 6.071511268615723,
"eval_runtime": 22.0977,
"eval_samples_per_second": 176.398,
"eval_steps_per_second": 2.76,
"step": 1400
}
],
"max_steps": 4616,
"num_train_epochs": 8,
"total_flos": 1.18623500497152e+16,
"trial_name": null,
"trial_params": null
}