llama-7b-code_easy / training_logs.json
joshuaclymer's picture
Upload folder using huggingface_hub
32f82ea
[
{
"loss": 0.7176,
"learning_rate": 0.0002,
"epoch": 0.02,
"step": 1
},
{
"loss": 0.626,
"learning_rate": 0.0002,
"epoch": 0.04,
"step": 2
},
{
"loss": 0.6406,
"learning_rate": 0.0002,
"epoch": 0.05,
"step": 3
},
{
"loss": 0.5599,
"learning_rate": 0.0002,
"epoch": 0.07,
"step": 4
},
{
"loss": 0.4681,
"learning_rate": 0.0002,
"epoch": 0.09,
"step": 5
},
{
"loss": 0.4553,
"learning_rate": 0.0002,
"epoch": 0.11,
"step": 6
},
{
"loss": 0.4104,
"learning_rate": 0.0002,
"epoch": 0.12,
"step": 7
},
{
"loss": 0.3978,
"learning_rate": 0.0002,
"epoch": 0.14,
"step": 8
},
{
"loss": 0.3848,
"learning_rate": 0.0002,
"epoch": 0.16,
"step": 9
},
{
"loss": 0.5363,
"learning_rate": 0.0002,
"epoch": 0.18,
"step": 10
},
{
"loss": 0.3722,
"learning_rate": 0.0002,
"epoch": 0.19,
"step": 11
},
{
"loss": 0.3675,
"learning_rate": 0.0002,
"epoch": 0.21,
"step": 12
},
{
"loss": 0.3061,
"learning_rate": 0.0002,
"epoch": 0.23,
"step": 13
},
{
"loss": 0.2357,
"learning_rate": 0.0002,
"epoch": 0.25,
"step": 14
},
{
"loss": 0.2983,
"learning_rate": 0.0002,
"epoch": 0.26,
"step": 15
},
{
"loss": 0.4426,
"learning_rate": 0.0002,
"epoch": 0.28,
"step": 16
},
{
"loss": 0.3633,
"learning_rate": 0.0002,
"epoch": 0.3,
"step": 17
},
{
"loss": 0.3673,
"learning_rate": 0.0002,
"epoch": 0.32,
"step": 18
},
{
"loss": 0.2728,
"learning_rate": 0.0002,
"epoch": 0.33,
"step": 19
},
{
"loss": 0.3937,
"learning_rate": 0.0002,
"epoch": 0.35,
"step": 20
},
{
"loss": 0.3076,
"learning_rate": 0.0002,
"epoch": 0.37,
"step": 21
},
{
"loss": 0.1257,
"learning_rate": 0.0002,
"epoch": 0.39,
"step": 22
},
{
"loss": 0.1323,
"learning_rate": 0.0002,
"epoch": 0.4,
"step": 23
},
{
"loss": 0.3184,
"learning_rate": 0.0002,
"epoch": 0.42,
"step": 24
},
{
"loss": 0.1576,
"learning_rate": 0.0002,
"epoch": 0.44,
"step": 25
},
{
"eval_code_easy_loss": 0.34121498465538025,
"eval_code_easy_score": -0.08921275287866592,
"eval_code_easy_brier_score": 0.08921275287866592,
"eval_code_easy_average_probability": 0.8383815884590149,
"eval_code_easy_accuracy": 0.91,
"eval_code_easy_probabilities": [
0.9973312616348267,
0.9983710646629333,
0.997231662273407,
0.9965489506721497,
0.9987871050834656,
0.9933956265449524,
0.6967766284942627,
0.7963165044784546,
0.8957101702690125,
0.8911858797073364,
0.6700810194015503,
0.887999415397644,
0.8513328433036804,
0.7391810417175293,
0.968675971031189,
0.8287560939788818,
0.8346096873283386,
0.9987896084785461,
0.7685402035713196,
0.6702771186828613,
0.6680448055267334,
0.9981842637062073,
0.9979216456413269,
0.9962202906608582,
0.999874472618103,
0.9479197263717651,
0.92351233959198,
0.9997767806053162,
0.9997504353523254,
0.9994015693664551,
0.8156358003616333,
0.8976724147796631,
0.35684335231781006,
0.6020401120185852,
0.695774257183075,
0.5311483144760132,
0.9899909496307373,
0.9998482465744019,
0.6876679062843323,
0.9594499468803406,
0.991589367389679,
0.7725859880447388,
0.9999998807907104,
0.9900687336921692,
0.9798843860626221,
0.019548967480659485,
0.10404925793409348,
0.03977573662996292,
0.638748049736023,
0.9745407700538635,
0.8131493330001831,
0.4198451638221741,
0.9959275126457214,
0.9906547665596008,
0.999859094619751,
0.999954104423523,
0.9999948740005493,
0.9941011071205139,
0.5232334733009338,
0.996921181678772,
0.9495030641555786,
0.9917147755622864,
0.9849343299865723,
0.9999825954437256,
0.999990701675415,
0.9999847412109375,
0.9999663829803467,
0.9981877207756042,
0.9100466370582581,
0.8915355801582336,
0.02915954403579235,
0.004325248766690493,
0.9999915361404419,
0.9999961853027344,
0.9984298348426819,
0.9361048936843872,
0.7488651871681213,
0.9966148734092712,
0.9871622920036316,
0.41953274607658386,
0.964065432548523,
0.6713405251502991,
0.7432451844215393,
0.5503003597259521,
0.9845474362373352,
0.9991044402122498,
0.9908291101455688,
0.78965824842453,
0.1398678869009018,
0.7981515526771545,
0.9919771552085876,
0.6074525117874146,
0.996113657951355,
0.9997654557228088,
0.9999492168426514,
0.9985866546630859,
0.9999914169311523,
0.9999855756759644,
0.9999876022338867,
0.9762148261070251
],
"eval_code_easy_runtime": 17.1306,
"eval_code_easy_samples_per_second": 5.838,
"eval_code_easy_steps_per_second": 0.117,
"epoch": 0.44,
"step": 25
},
{
"eval_code_hard_loss": 1.0086419582366943,
"eval_code_hard_score": -0.28139397501945496,
"eval_code_hard_brier_score": 0.28139397501945496,
"eval_code_hard_average_probability": 0.5282384157180786,
"eval_code_hard_accuracy": 0.61,
"eval_code_hard_probabilities": [
0.9940557479858398,
0.5578140020370483,
0.4544859528541565,
0.9417738914489746,
0.9485132098197937,
0.9686947464942932,
0.34841597080230713,
0.37181127071380615,
0.25785237550735474,
0.5207318067550659,
0.4167640209197998,
0.43461835384368896,
0.9888814091682434,
0.11123046278953552,
0.15088781714439392,
0.4391879737377167,
0.4488320052623749,
0.48593440651893616,
0.902611255645752,
0.8697171211242676,
0.7664279937744141,
0.5244131684303284,
0.6065243482589722,
0.6952157616615295,
0.6103460192680359,
0.5863346457481384,
0.45322948694229126,
0.15636862814426422,
0.7031698226928711,
0.6602889895439148,
0.6120182275772095,
0.5830819010734558,
0.6418420076370239,
0.6374729871749878,
0.6232181191444397,
0.6474714875221252,
0.7186492085456848,
0.5149539709091187,
0.27517759799957275,
0.5705553293228149,
0.6419432163238525,
0.7602654695510864,
0.6087383031845093,
0.6384655237197876,
0.6501816511154175,
0.7480051517486572,
0.57122802734375,
0.3884866237640381,
0.692376971244812,
0.6024836301803589,
0.591353178024292,
0.43030333518981934,
0.2442779242992401,
0.22508001327514648,
0.0015732977772131562,
0.001636496395803988,
0.0015261240769177675,
0.5446102023124695,
0.5613898038864136,
0.48322516679763794,
0.004009010270237923,
0.003701635170727968,
0.0036944146268069744,
0.5865771174430847,
0.4441525340080261,
0.8783551454544067,
0.3124151825904846,
0.9564208984375,
0.3860621452331543,
0.9370018243789673,
0.8526004552841187,
0.8547378182411194,
0.5418311953544617,
0.4870041012763977,
0.6209992170333862,
0.3137390911579132,
0.1501292884349823,
0.2207321971654892,
0.508880615234375,
0.4939533770084381,
0.5084103345870972,
0.5919321775436401,
0.6774353384971619,
0.5825659036636353,
0.5430614948272705,
0.5601528882980347,
0.6737014651298523,
0.45028504729270935,
0.5943098664283752,
0.4708901345729828,
0.5047590732574463,
0.5815862417221069,
0.5608630776405334,
0.2037791609764099,
0.20710068941116333,
0.1756637841463089,
0.48647746443748474,
0.9042562246322632,
0.706450343132019,
0.6964374780654907
],
"eval_code_hard_runtime": 57.8032,
"eval_code_hard_samples_per_second": 1.73,
"eval_code_hard_steps_per_second": 0.035,
"epoch": 0.44,
"step": 25
},
{
"loss": 0.1411,
"learning_rate": 0.0002,
"epoch": 0.46,
"step": 26
},
{
"loss": 0.2268,
"learning_rate": 0.0002,
"epoch": 0.47,
"step": 27
},
{
"loss": 0.1373,
"learning_rate": 0.0002,
"epoch": 0.49,
"step": 28
},
{
"loss": 0.2788,
"learning_rate": 0.0002,
"epoch": 0.51,
"step": 29
},
{
"loss": 0.0514,
"learning_rate": 0.0002,
"epoch": 0.53,
"step": 30
},
{
"loss": 0.174,
"learning_rate": 0.0002,
"epoch": 0.54,
"step": 31
},
{
"loss": 0.1099,
"learning_rate": 0.0002,
"epoch": 0.56,
"step": 32
},
{
"loss": 0.181,
"learning_rate": 0.0002,
"epoch": 0.58,
"step": 33
},
{
"loss": 0.0913,
"learning_rate": 0.0002,
"epoch": 0.6,
"step": 34
},
{
"loss": 0.1685,
"learning_rate": 0.0002,
"epoch": 0.61,
"step": 35
},
{
"loss": 0.1701,
"learning_rate": 0.0002,
"epoch": 0.63,
"step": 36
},
{
"loss": 0.054,
"learning_rate": 0.0002,
"epoch": 0.65,
"step": 37
},
{
"loss": 0.0469,
"learning_rate": 0.0002,
"epoch": 0.67,
"step": 38
},
{
"loss": 0.0865,
"learning_rate": 0.0002,
"epoch": 0.68,
"step": 39
},
{
"loss": 0.1743,
"learning_rate": 0.0002,
"epoch": 0.7,
"step": 40
},
{
"loss": 0.072,
"learning_rate": 0.0002,
"epoch": 0.72,
"step": 41
},
{
"loss": 0.1192,
"learning_rate": 0.0002,
"epoch": 0.74,
"step": 42
},
{
"loss": 0.056,
"learning_rate": 0.0002,
"epoch": 0.75,
"step": 43
},
{
"loss": 0.1633,
"learning_rate": 0.0002,
"epoch": 0.77,
"step": 44
},
{
"loss": 0.0487,
"learning_rate": 0.0002,
"epoch": 0.79,
"step": 45
},
{
"loss": 0.111,
"learning_rate": 0.0002,
"epoch": 0.81,
"step": 46
},
{
"loss": 0.0998,
"learning_rate": 0.0002,
"epoch": 0.82,
"step": 47
},
{
"loss": 0.162,
"learning_rate": 0.0002,
"epoch": 0.84,
"step": 48
},
{
"loss": 0.1508,
"learning_rate": 0.0002,
"epoch": 0.86,
"step": 49
},
{
"loss": 0.158,
"learning_rate": 0.0002,
"epoch": 0.88,
"step": 50
},
{
"eval_code_easy_loss": 0.26072654128074646,
"eval_code_easy_score": -0.07467034459114075,
"eval_code_easy_brier_score": 0.07467034459114075,
"eval_code_easy_average_probability": 0.8786800503730774,
"eval_code_easy_accuracy": 0.91,
"eval_code_easy_probabilities": [
0.993358314037323,
0.9995765089988708,
0.9998908042907715,
0.9999998807907104,
0.9999995231628418,
0.9999985694885254,
0.8484190106391907,
0.9649071097373962,
0.9977992177009583,
0.6331314444541931,
0.7679887413978577,
0.7734430432319641,
0.9972902536392212,
0.9985353946685791,
0.9985615611076355,
0.9800190329551697,
0.9999395608901978,
0.9999872446060181,
0.9737911820411682,
0.36379221081733704,
0.8597760796546936,
1.0,
1.0,
0.9999998807907104,
0.9999985694885254,
0.9999895095825195,
0.9909597635269165,
0.9998130202293396,
0.9994999170303345,
0.9997158646583557,
0.7183305621147156,
0.6560869216918945,
0.5542047023773193,
0.9361991882324219,
0.9441348314285278,
0.7303879261016846,
0.9999759197235107,
0.9999983310699463,
0.9700189828872681,
0.9989641904830933,
0.9984664916992188,
0.9994432330131531,
1.0,
1.0,
0.9999915361404419,
0.3375416696071625,
0.09802956134080887,
0.1487502008676529,
0.9902018904685974,
0.9985565543174744,
0.9950182437896729,
0.147225022315979,
0.9999966621398926,
0.9999779462814331,
0.999680757522583,
0.9997058510780334,
0.9998592138290405,
0.9999998807907104,
0.5950238108634949,
0.999997615814209,
0.999997615814209,
1.0,
0.9999998807907104,
0.9999052286148071,
0.9999927282333374,
0.99998939037323,
0.9201897382736206,
0.9418489933013916,
0.8507129549980164,
0.609320342540741,
0.07032223790884018,
0.003869357518851757,
0.9999988079071045,
0.9999964237213135,
0.999927282333374,
0.9999841451644897,
0.9998486042022705,
0.9999936819076538,
0.9999700784683228,
0.9927895665168762,
0.9999749660491943,
0.9999265670776367,
0.9999915361404419,
0.9663752913475037,
0.7864801287651062,
0.8877244591712952,
0.8446969985961914,
0.968808114528656,
0.03765033930540085,
0.8669968247413635,
0.9999896287918091,
0.6672230958938599,
0.9992057681083679,
0.9999997615814209,
1.0,
0.9999957084655762,
1.0,
1.0,
1.0,
0.4983561038970947
],
"eval_code_easy_runtime": 17.1242,
"eval_code_easy_samples_per_second": 5.84,
"eval_code_easy_steps_per_second": 0.117,
"epoch": 0.88,
"step": 50
},
{
"eval_code_hard_loss": 0.6272040605545044,
"eval_code_hard_score": -0.2097242921590805,
"eval_code_hard_brier_score": 0.2097242921590805,
"eval_code_hard_average_probability": 0.6029007434844971,
"eval_code_hard_accuracy": 0.6,
"eval_code_hard_probabilities": [
0.9979154467582703,
0.7335461378097534,
0.4222542941570282,
0.9998534917831421,
0.9998674392700195,
0.9999492168426514,
0.5001015663146973,
0.5079646706581116,
0.5019187927246094,
0.6476748585700989,
0.3486829996109009,
0.38149967789649963,
0.9999423027038574,
0.9999504089355469,
0.9999474287033081,
0.4916134774684906,
0.5060228705406189,
0.49220144748687744,
0.9999898672103882,
0.9999915361404419,
0.9999905824661255,
0.8047722578048706,
0.8523899912834167,
0.8705809116363525,
0.3922067880630493,
0.4324615001678467,
0.3545004725456238,
0.15673233568668365,
0.6471846699714661,
0.6424362063407898,
0.510160505771637,
0.5128241181373596,
0.46828266978263855,
0.43302497267723083,
0.5793876051902771,
0.5684614181518555,
0.33457034826278687,
0.3091184198856354,
0.0008105101296678185,
0.37077412009239197,
0.4145500063896179,
0.4761010408401489,
0.7447249293327332,
0.438394695520401,
0.7413656115531921,
0.4960051476955414,
0.5138289928436279,
0.5173673033714294,
0.5924621820449829,
0.7587030529975891,
0.7857908606529236,
0.3681909143924713,
0.3770434856414795,
0.6424773931503296,
0.9851202368736267,
0.9859110116958618,
0.9855308532714844,
0.5370909571647644,
0.6129180192947388,
0.5251724123954773,
0.9999006986618042,
0.9998983144760132,
0.9998980760574341,
0.8960296511650085,
0.8455232977867126,
0.9984153509140015,
0.44011861085891724,
0.5573714971542358,
0.45769622921943665,
0.5591621994972229,
0.42212074995040894,
0.5001301765441895,
0.4822878837585449,
0.7320272326469421,
0.6362009048461914,
0.429107666015625,
0.41713282465934753,
0.4034157991409302,
0.5436654686927795,
0.47678494453430176,
0.5470160245895386,
0.5428171157836914,
0.5227158665657043,
0.5470259785652161,
0.4163595139980316,
0.35028159618377686,
0.5244665145874023,
0.40996211767196655,
0.9096313714981079,
0.37183111906051636,
0.47767525911331177,
0.47421643137931824,
0.46094876527786255,
0.47423312067985535,
0.4818762242794037,
0.4068610668182373,
0.497257262468338,
0.559116542339325,
0.5229799747467041,
0.725604772567749
],
"eval_code_hard_runtime": 57.8139,
"eval_code_hard_samples_per_second": 1.73,
"eval_code_hard_steps_per_second": 0.035,
"epoch": 0.88,
"step": 50
},
{
"loss": 0.1258,
"learning_rate": 0.0002,
"epoch": 0.89,
"step": 51
},
{
"loss": 0.0618,
"learning_rate": 0.0002,
"epoch": 0.91,
"step": 52
},
{
"loss": 0.2478,
"learning_rate": 0.0002,
"epoch": 0.93,
"step": 53
},
{
"loss": 0.4399,
"learning_rate": 0.0002,
"epoch": 0.95,
"step": 54
},
{
"loss": 0.1653,
"learning_rate": 0.0002,
"epoch": 0.96,
"step": 55
},
{
"loss": 0.0441,
"learning_rate": 0.0002,
"epoch": 0.98,
"step": 56
},
{
"loss": 0.0223,
"learning_rate": 0.0002,
"epoch": 1.0,
"step": 57
},
{
"loss": 0.0056,
"learning_rate": 0.0002,
"epoch": 1.02,
"step": 58
},
{
"loss": 0.1784,
"learning_rate": 0.0002,
"epoch": 1.04,
"step": 59
},
{
"loss": 0.0642,
"learning_rate": 0.0002,
"epoch": 1.05,
"step": 60
},
{
"loss": 0.1516,
"learning_rate": 0.0002,
"epoch": 1.07,
"step": 61
},
{
"loss": 0.0548,
"learning_rate": 0.0002,
"epoch": 1.09,
"step": 62
},
{
"loss": 0.0663,
"learning_rate": 0.0002,
"epoch": 1.11,
"step": 63
},
{
"loss": 0.1384,
"learning_rate": 0.0002,
"epoch": 1.12,
"step": 64
},
{
"loss": 0.0505,
"learning_rate": 0.0002,
"epoch": 1.14,
"step": 65
},
{
"loss": 0.0631,
"learning_rate": 0.0002,
"epoch": 1.16,
"step": 66
},
{
"loss": 0.0532,
"learning_rate": 0.0002,
"epoch": 1.18,
"step": 67
},
{
"loss": 0.028,
"learning_rate": 0.0002,
"epoch": 1.19,
"step": 68
},
{
"loss": 0.0891,
"learning_rate": 0.0002,
"epoch": 1.21,
"step": 69
},
{
"loss": 0.0735,
"learning_rate": 0.0002,
"epoch": 1.23,
"step": 70
},
{
"loss": 0.0397,
"learning_rate": 0.0002,
"epoch": 1.25,
"step": 71
},
{
"loss": 0.0342,
"learning_rate": 0.0002,
"epoch": 1.26,
"step": 72
},
{
"loss": 0.0234,
"learning_rate": 0.0002,
"epoch": 1.28,
"step": 73
},
{
"loss": 0.0108,
"learning_rate": 0.0002,
"epoch": 1.3,
"step": 74
},
{
"loss": 0.0525,
"learning_rate": 0.0002,
"epoch": 1.32,
"step": 75
},
{
"eval_code_easy_loss": 0.22187194228172302,
"eval_code_easy_score": -0.044156476855278015,
"eval_code_easy_brier_score": 0.044156476855278015,
"eval_code_easy_average_probability": 0.9226246476173401,
"eval_code_easy_accuracy": 0.95,
"eval_code_easy_probabilities": [
0.9997438788414001,
0.9999991655349731,
0.9999997615814209,
0.9999998807907104,
0.9999997615814209,
0.9999947547912598,
0.8262179493904114,
0.9995021820068359,
0.998319685459137,
0.988502025604248,
0.9749259352684021,
0.9808527827262878,
0.9996113181114197,
0.9993595480918884,
0.9999059438705444,
0.9027615189552307,
0.9773653745651245,
0.9999016523361206,
0.9789873957633972,
0.45622462034225464,
0.9388629198074341,
1.0,
1.0,
0.9999997615814209,
0.9999998807907104,
0.9999997615814209,
0.9993877410888672,
1.0,
0.9999998807907104,
0.9999998807907104,
0.7212077379226685,
0.7462506294250488,
0.5923525094985962,
0.9379615187644958,
0.9973413348197937,
0.881181001663208,
0.9957089424133301,
0.9999865293502808,
0.9746310114860535,
0.9999599456787109,
0.9995377063751221,
0.9997233748435974,
1.0,
1.0,
1.0,
0.7736052870750427,
0.15912780165672302,
0.6213001608848572,
0.9973452687263489,
0.9999977350234985,
0.9999613761901855,
0.5660737752914429,
1.0,
0.9999997615814209,
0.999991774559021,
0.9999953508377075,
0.999993085861206,
0.9999995231628418,
0.6532827615737915,
0.999998927116394,
0.9999794960021973,
0.9999988079071045,
0.999998927116394,
0.9999998807907104,
1.0,
1.0,
0.997825026512146,
0.999733030796051,
0.9979919195175171,
0.7770313024520874,
0.011424303986132145,
9.0804387582466e-05,
1.0,
1.0,
0.9999986886978149,
0.99936443567276,
0.9927093982696533,
0.9999185800552368,
0.9999991655349731,
0.9982544779777527,
0.9999996423721313,
0.9998847246170044,
0.9999909400939941,
0.9689223766326904,
0.9911876916885376,
0.9976316690444946,
0.9958218336105347,
0.993769109249115,
0.40141063928604126,
0.8972048759460449,
0.9999942779541016,
0.6255294680595398,
0.9922812581062317,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.9875762462615967
],
"eval_code_easy_runtime": 17.1185,
"eval_code_easy_samples_per_second": 5.842,
"eval_code_easy_steps_per_second": 0.117,
"epoch": 1.32,
"step": 75
},
{
"eval_code_hard_loss": 0.613934338092804,
"eval_code_hard_score": -0.20562565326690674,
"eval_code_hard_brier_score": 0.20562565326690674,
"eval_code_hard_average_probability": 0.6201408505439758,
"eval_code_hard_accuracy": 0.68,
"eval_code_hard_probabilities": [
0.9997522234916687,
0.6726481318473816,
0.4837990403175354,
0.9999945163726807,
0.9999929666519165,
0.9999992847442627,
0.5199088454246521,
0.5684794783592224,
0.5020670890808105,
0.6793079376220703,
0.3110067546367645,
0.4360009431838989,
0.999996542930603,
0.9999467134475708,
0.9999498128890991,
0.5130871534347534,
0.5327062010765076,
0.5177708864212036,
0.9999998807907104,
0.9999998807907104,
0.9999998807907104,
0.8731643557548523,
0.9360027313232422,
0.9267758727073669,
0.35389643907546997,
0.39809563755989075,
0.20879632234573364,
0.07375767827033997,
0.5250951051712036,
0.5149143934249878,
0.7871077656745911,
0.7780881524085999,
0.6657141447067261,
0.5761507749557495,
0.758147120475769,
0.6872865557670593,
0.29255661368370056,
0.1509549617767334,
0.004953588359057903,
0.3843843936920166,
0.4454701542854309,
0.5638160705566406,
0.6067726612091064,
0.3357914388179779,
0.5700987577438354,
0.8174778819084167,
0.5854097008705139,
0.5972737073898315,
0.6639817953109741,
0.8354654908180237,
0.8449712991714478,
0.35150453448295593,
0.25781506299972534,
0.7074410319328308,
0.9738778471946716,
0.9760010242462158,
0.9789164066314697,
0.5705886483192444,
0.642711341381073,
0.5328339338302612,
0.999995231628418,
0.9999946355819702,
0.9999946355819702,
0.9326982498168945,
0.7097592353820801,
0.9997654557228088,
0.446959525346756,
0.6458287835121155,
0.47074800729751587,
0.6868018507957458,
0.5448780059814453,
0.5729479789733887,
0.4683343768119812,
0.570155918598175,
0.5829461216926575,
0.4596071243286133,
0.407598614692688,
0.38017627596855164,
0.5608930587768555,
0.49683383107185364,
0.600445568561554,
0.6426953077316284,
0.8385844826698303,
0.6047258377075195,
0.2799336314201355,
0.2773086130619049,
0.6116948127746582,
0.43185508251190186,
0.8329448103904724,
0.4886557459831238,
0.4926038086414337,
0.4693112373352051,
0.4659242630004883,
0.321489155292511,
0.3335782587528229,
0.1558743566274643,
0.5026931762695312,
0.7897409200668335,
0.5024547576904297,
0.9441749453544617
],
"eval_code_hard_runtime": 57.778,
"eval_code_hard_samples_per_second": 1.731,
"eval_code_hard_steps_per_second": 0.035,
"epoch": 1.32,
"step": 75
},
{
"loss": 0.0344,
"learning_rate": 0.0002,
"epoch": 1.33,
"step": 76
},
{
"loss": 0.0514,
"learning_rate": 0.0002,
"epoch": 1.35,
"step": 77
},
{
"loss": 0.0065,
"learning_rate": 0.0002,
"epoch": 1.37,
"step": 78
},
{
"loss": 0.0893,
"learning_rate": 0.0002,
"epoch": 1.39,
"step": 79
},
{
"loss": 0.0431,
"learning_rate": 0.0002,
"epoch": 1.4,
"step": 80
},
{
"loss": 0.0101,
"learning_rate": 0.0002,
"epoch": 1.42,
"step": 81
},
{
"loss": 0.0622,
"learning_rate": 0.0002,
"epoch": 1.44,
"step": 82
},
{
"loss": 0.0007,
"learning_rate": 0.0002,
"epoch": 1.46,
"step": 83
},
{
"loss": 0.0084,
"learning_rate": 0.0002,
"epoch": 1.47,
"step": 84
},
{
"loss": 0.0307,
"learning_rate": 0.0002,
"epoch": 1.49,
"step": 85
},
{
"loss": 0.0005,
"learning_rate": 0.0002,
"epoch": 1.51,
"step": 86
},
{
"loss": 0.001,
"learning_rate": 0.0002,
"epoch": 1.53,
"step": 87
},
{
"loss": 0.0243,
"learning_rate": 0.0002,
"epoch": 1.54,
"step": 88
},
{
"loss": 0.0417,
"learning_rate": 0.0002,
"epoch": 1.56,
"step": 89
},
{
"loss": 0.0418,
"learning_rate": 0.0002,
"epoch": 1.58,
"step": 90
},
{
"loss": 0.0207,
"learning_rate": 0.0002,
"epoch": 1.6,
"step": 91
},
{
"loss": 0.2464,
"learning_rate": 0.0002,
"epoch": 1.61,
"step": 92
},
{
"loss": 0.1456,
"learning_rate": 0.0002,
"epoch": 1.63,
"step": 93
},
{
"loss": 0.0005,
"learning_rate": 0.0002,
"epoch": 1.65,
"step": 94
},
{
"loss": 0.0025,
"learning_rate": 0.0002,
"epoch": 1.67,
"step": 95
},
{
"loss": 0.0456,
"learning_rate": 0.0002,
"epoch": 1.68,
"step": 96
},
{
"loss": 0.0069,
"learning_rate": 0.0002,
"epoch": 1.7,
"step": 97
},
{
"loss": 0.0092,
"learning_rate": 0.0002,
"epoch": 1.72,
"step": 98
},
{
"loss": 0.1516,
"learning_rate": 0.0002,
"epoch": 1.74,
"step": 99
},
{
"loss": 0.0164,
"learning_rate": 0.0002,
"epoch": 1.75,
"step": 100
},
{
"eval_code_easy_loss": 0.37038832902908325,
"eval_code_easy_score": -0.03134830296039581,
"eval_code_easy_brier_score": 0.03134830296039581,
"eval_code_easy_average_probability": 0.9487545490264893,
"eval_code_easy_accuracy": 0.97,
"eval_code_easy_probabilities": [
1.0,
1.0,
1.0,
1.0,
1.0,
0.9999997615814209,
0.9346917271614075,
1.0,
0.9999998807907104,
0.9997040629386902,
0.9992467164993286,
0.9994949102401733,
0.9994587302207947,
0.9999791383743286,
0.9999996423721313,
0.9992595314979553,
0.9999998807907104,
1.0,
0.9998413324356079,
0.6475287675857544,
0.9992796778678894,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.8477630615234375,
0.863884687423706,
0.5967232584953308,
0.8738592267036438,
0.9956537485122681,
0.7420057058334351,
0.9940522313117981,
1.0,
0.9952139854431152,
0.9999994039535522,
1.0,
0.9999545812606812,
1.0,
1.0,
1.0,
0.9897578954696655,
0.2987537980079651,
0.9979766011238098,
0.9999840259552002,
1.0,
1.0,
0.9765376448631287,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.7584323883056641,
1.0,
1.0,
1.0,
0.9999982118606567,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.9797248840332031,
1.8104368791682646e-05,
2.9319591199339357e-10,
1.0,
1.0,
1.0,
1.0,
0.9998795986175537,
1.0,
1.0,
0.9999961853027344,
1.0,
0.9750629663467407,
0.9934146404266357,
0.9500468969345093,
1.0,
1.0,
1.0,
0.9999947547912598,
0.843275785446167,
0.9999544620513916,
1.0,
0.6250555515289307,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.999996542930603
],
"eval_code_easy_runtime": 17.1259,
"eval_code_easy_samples_per_second": 5.839,
"eval_code_easy_steps_per_second": 0.117,
"epoch": 1.75,
"step": 100
},
{
"eval_code_hard_loss": 1.3108476400375366,
"eval_code_hard_score": -0.238570898771286,
"eval_code_hard_brier_score": 0.238570898771286,
"eval_code_hard_average_probability": 0.637829601764679,
"eval_code_hard_accuracy": 0.69,
"eval_code_hard_probabilities": [
1.0,
0.8749839663505554,
0.6682276129722595,
1.0,
1.0,
1.0,
0.4749336838722229,
0.7184466123580933,
0.4824881851673126,
0.7342566251754761,
0.34355923533439636,
0.3212933838367462,
0.999998927116394,
0.11708329617977142,
0.2533932030200958,
0.4083555340766907,
0.49011072516441345,
0.5354903340339661,
0.9999998807907104,
0.9999995231628418,
0.9999921321868896,
0.9668300151824951,
0.9995100498199463,
0.9995787739753723,
0.5743376612663269,
0.6365594267845154,
0.19516201317310333,
5.900204200770531e-07,
0.6282064914703369,
0.6691750884056091,
0.9999256134033203,
0.9999665021896362,
0.9993390440940857,
0.9999998807907104,
0.9648662805557251,
0.7553066611289978,
0.15391865372657776,
0.0059105814434587955,
0.0041419873014092445,
0.6582466959953308,
0.6763702630996704,
0.8389391303062439,
0.8646246194839478,
0.7645048499107361,
0.873203456401825,
0.9956676959991455,
0.6906198263168335,
0.7347480058670044,
0.9999996423721313,
0.9999585151672363,
0.9992007613182068,
0.034449804574251175,
0.01096381526440382,
0.8154931664466858,
3.0823147767478076e-07,
3.748235144485079e-07,
3.7546107023445074e-07,
0.507102370262146,
0.7094148397445679,
0.6129729151725769,
0.8193482756614685,
0.6906271576881409,
0.7478002905845642,
0.9924389123916626,
0.7537370324134827,
1.0,
0.33823978900909424,
0.8392996788024902,
0.33018726110458374,
0.9978412389755249,
0.9927680492401123,
0.9661868214607239,
0.739952027797699,
0.757354199886322,
0.8434256911277771,
0.5655504465103149,
0.09736324101686478,
0.15342624485492706,
0.5846419930458069,
0.5172314047813416,
0.7640779614448547,
0.7927785515785217,
0.9958668947219849,
0.6929405331611633,
0.4140860140323639,
0.4588907063007355,
0.8826824426651001,
0.4302871525287628,
0.8620820045471191,
0.5388514995574951,
0.49043917655944824,
0.4464437961578369,
0.5461203455924988,
0.01837478019297123,
0.01984582096338272,
0.0083817383274436,
0.417593389749527,
0.9964190721511841,
0.5195274949073792,
0.9980224370956421
],
"eval_code_hard_runtime": 57.8261,
"eval_code_hard_samples_per_second": 1.729,
"eval_code_hard_steps_per_second": 0.035,
"epoch": 1.75,
"step": 100
},
{
"train_runtime": 1718.4456,
"train_samples_per_second": 1.862,
"train_steps_per_second": 0.058,
"total_flos": 0.0,
"train_loss": 0.1626638910168549,
"epoch": 1.75,
"step": 100
}
]