openllama-3b-code_easy / training_logs.json
joshuaclymer's picture
Upload folder using huggingface_hub
11cc438
[
{
"loss": 0.7143,
"learning_rate": 7.2e-05,
"epoch": 0.02,
"step": 1
},
{
"loss": 0.6911,
"learning_rate": 7.2e-05,
"epoch": 0.04,
"step": 2
},
{
"loss": 0.6963,
"learning_rate": 7.2e-05,
"epoch": 0.05,
"step": 3
},
{
"loss": 0.678,
"learning_rate": 7.2e-05,
"epoch": 0.07,
"step": 4
},
{
"loss": 0.6763,
"learning_rate": 7.2e-05,
"epoch": 0.09,
"step": 5
},
{
"loss": 0.6766,
"learning_rate": 7.2e-05,
"epoch": 0.11,
"step": 6
},
{
"loss": 0.6475,
"learning_rate": 7.2e-05,
"epoch": 0.12,
"step": 7
},
{
"loss": 0.6429,
"learning_rate": 7.2e-05,
"epoch": 0.14,
"step": 8
},
{
"loss": 0.6151,
"learning_rate": 7.2e-05,
"epoch": 0.16,
"step": 9
},
{
"loss": 0.6278,
"learning_rate": 7.2e-05,
"epoch": 0.18,
"step": 10
},
{
"loss": 0.6226,
"learning_rate": 7.2e-05,
"epoch": 0.19,
"step": 11
},
{
"loss": 0.6017,
"learning_rate": 7.2e-05,
"epoch": 0.21,
"step": 12
},
{
"loss": 0.6445,
"learning_rate": 7.2e-05,
"epoch": 0.23,
"step": 13
},
{
"loss": 0.5827,
"learning_rate": 7.2e-05,
"epoch": 0.25,
"step": 14
},
{
"loss": 0.5437,
"learning_rate": 7.2e-05,
"epoch": 0.26,
"step": 15
},
{
"loss": 0.5466,
"learning_rate": 7.2e-05,
"epoch": 0.28,
"step": 16
},
{
"loss": 0.5452,
"learning_rate": 7.2e-05,
"epoch": 0.3,
"step": 17
},
{
"loss": 0.5531,
"learning_rate": 7.2e-05,
"epoch": 0.32,
"step": 18
},
{
"loss": 0.5615,
"learning_rate": 7.2e-05,
"epoch": 0.33,
"step": 19
},
{
"loss": 0.5586,
"learning_rate": 7.2e-05,
"epoch": 0.35,
"step": 20
},
{
"loss": 0.5116,
"learning_rate": 7.2e-05,
"epoch": 0.37,
"step": 21
},
{
"loss": 0.4264,
"learning_rate": 7.2e-05,
"epoch": 0.39,
"step": 22
},
{
"loss": 0.3866,
"learning_rate": 7.2e-05,
"epoch": 0.4,
"step": 23
},
{
"loss": 0.5063,
"learning_rate": 7.2e-05,
"epoch": 0.42,
"step": 24
},
{
"loss": 0.4082,
"learning_rate": 7.2e-05,
"epoch": 0.44,
"step": 25
},
{
"eval_code_easy_loss": 0.4812281131744385,
"eval_code_easy_score": -0.14910461008548737,
"eval_code_easy_brier_score": 0.14910461008548737,
"eval_code_easy_average_probability": 0.6539512872695923,
"eval_code_easy_accuracy": 0.85,
"eval_code_easy_probabilities": [
0.7876701354980469,
0.7523844242095947,
0.6472867131233215,
0.7906888723373413,
0.8586155772209167,
0.6918468475341797,
0.4946542978286743,
0.578050434589386,
0.5266337990760803,
0.5107197761535645,
0.505793035030365,
0.45479127764701843,
0.4748489260673523,
0.5174643397331238,
0.4289233088493347,
0.5851827263832092,
0.8242431879043579,
0.9364657402038574,
0.5783424377441406,
0.48775342106819153,
0.4637766480445862,
0.8675320148468018,
0.7406874299049377,
0.7409279346466064,
0.7196134924888611,
0.6729203462600708,
0.7667352557182312,
0.7618107795715332,
0.5047318935394287,
0.8554519414901733,
0.48701024055480957,
0.5184937119483948,
0.26061391830444336,
0.6388080716133118,
0.557035505771637,
0.541324257850647,
0.7926197052001953,
0.9460112452507019,
0.548197865486145,
0.7201337814331055,
0.7818894982337952,
0.6181257367134094,
0.6234219670295715,
0.6032276749610901,
0.5777942538261414,
0.7809272408485413,
0.41414424777030945,
0.5810375809669495,
0.5093526244163513,
0.5954176783561707,
0.5138558745384216,
0.43075305223464966,
0.6626322865486145,
0.6255874037742615,
0.7210639119148254,
0.9159994125366211,
0.8485010266304016,
0.8202032446861267,
0.5092943906784058,
0.8338488936424255,
0.5957618355751038,
0.6853424906730652,
0.6718733310699463,
0.9674582481384277,
0.9621556401252747,
0.9909061193466187,
0.5818647742271423,
0.5197274088859558,
0.5295705795288086,
0.5335188508033752,
0.35347408056259155,
0.20989884436130524,
0.8785598278045654,
0.9439873695373535,
0.8205457329750061,
0.5045193433761597,
0.5271017551422119,
0.7975461483001709,
0.5877244472503662,
0.5780063271522522,
0.4087151885032654,
0.502793550491333,
0.5167631506919861,
0.5050711035728455,
0.6043635010719299,
0.6559382677078247,
0.7178013920783997,
0.6761601567268372,
0.4603518843650818,
0.4613737463951111,
0.8092925548553467,
0.5021629929542542,
0.8891451358795166,
0.8492684960365295,
0.8078585863113403,
0.7930970191955566,
0.9665018916130066,
0.920259416103363,
0.9725727438926697,
0.6322199702262878
],
"eval_code_easy_runtime": 12.1001,
"eval_code_easy_samples_per_second": 8.264,
"eval_code_easy_steps_per_second": 0.165,
"epoch": 0.44,
"step": 25
},
{
"eval_code_hard_loss": 0.6784273386001587,
"eval_code_hard_score": -0.23437322676181793,
"eval_code_hard_brier_score": 0.23437322676181793,
"eval_code_hard_average_probability": 0.5313071608543396,
"eval_code_hard_accuracy": 0.63,
"eval_code_hard_probabilities": [
0.6153736114501953,
0.5013247132301331,
0.5044454336166382,
0.9211879968643188,
0.8884113430976868,
0.9285586476325989,
0.4261794984340668,
0.4360635578632355,
0.5060375928878784,
0.4633253514766693,
0.49218645691871643,
0.5567600727081299,
0.878082811832428,
0.8190033435821533,
0.8284511566162109,
0.5461653470993042,
0.5189316868782043,
0.535224974155426,
0.42352545261383057,
0.37337490916252136,
0.3578466475009918,
0.5396072268486023,
0.5384335517883301,
0.5416436195373535,
0.4356071949005127,
0.5550840497016907,
0.38927942514419556,
0.19190631806850433,
0.525556743144989,
0.563071072101593,
0.5183289647102356,
0.5080891847610474,
0.4980936646461487,
0.4459080696105957,
0.4500540792942047,
0.44474658370018005,
0.35595303773880005,
0.3355133831501007,
0.7131393551826477,
0.5476568937301636,
0.6050811409950256,
0.6322606801986694,
0.5295588374137878,
0.5178883075714111,
0.5216538906097412,
0.6891106367111206,
0.4794466197490692,
0.5025456547737122,
0.7308195233345032,
0.37424957752227783,
0.37053611874580383,
0.5303871631622314,
0.5796085000038147,
0.7150017619132996,
0.4277779161930084,
0.4290866553783417,
0.430403470993042,
0.5324957370758057,
0.5010586380958557,
0.520031750202179,
0.6008161902427673,
0.5889564156532288,
0.5865479111671448,
0.5397530794143677,
0.5804182887077332,
0.44839367270469666,
0.5159912109375,
0.7570854425430298,
0.49450379610061646,
0.6515185236930847,
0.6266943216323853,
0.4965268075466156,
0.5120733976364136,
0.540576696395874,
0.5366629958152771,
0.4482656717300415,
0.46240511536598206,
0.5012778639793396,
0.4999728500843048,
0.4921656548976898,
0.5284487009048462,
0.5068923234939575,
0.46083176136016846,
0.5137518644332886,
0.5790409445762634,
0.515843391418457,
0.542677640914917,
0.5102853178977966,
0.4863445460796356,
0.49948781728744507,
0.5162002444267273,
0.5039041042327881,
0.4971469044685364,
0.4223201274871826,
0.4179053008556366,
0.43632903695106506,
0.5268281102180481,
0.5249642729759216,
0.4914223849773407,
0.532348096370697
],
"eval_code_hard_runtime": 53.2874,
"eval_code_hard_samples_per_second": 1.877,
"eval_code_hard_steps_per_second": 0.038,
"epoch": 0.44,
"step": 25
},
{
"loss": 0.4493,
"learning_rate": 7.2e-05,
"epoch": 0.46,
"step": 26
},
{
"loss": 0.4695,
"learning_rate": 7.2e-05,
"epoch": 0.47,
"step": 27
},
{
"loss": 0.3858,
"learning_rate": 7.2e-05,
"epoch": 0.49,
"step": 28
},
{
"loss": 0.3683,
"learning_rate": 7.2e-05,
"epoch": 0.51,
"step": 29
},
{
"loss": 0.4011,
"learning_rate": 7.2e-05,
"epoch": 0.53,
"step": 30
},
{
"loss": 0.346,
"learning_rate": 7.2e-05,
"epoch": 0.54,
"step": 31
},
{
"loss": 0.2892,
"learning_rate": 7.2e-05,
"epoch": 0.56,
"step": 32
},
{
"loss": 0.4798,
"learning_rate": 7.2e-05,
"epoch": 0.58,
"step": 33
},
{
"loss": 0.3073,
"learning_rate": 7.2e-05,
"epoch": 0.6,
"step": 34
},
{
"loss": 0.3376,
"learning_rate": 7.2e-05,
"epoch": 0.61,
"step": 35
},
{
"loss": 0.2565,
"learning_rate": 7.2e-05,
"epoch": 0.63,
"step": 36
},
{
"loss": 0.1994,
"learning_rate": 7.2e-05,
"epoch": 0.65,
"step": 37
},
{
"loss": 0.3296,
"learning_rate": 7.2e-05,
"epoch": 0.67,
"step": 38
},
{
"loss": 0.2029,
"learning_rate": 7.2e-05,
"epoch": 0.68,
"step": 39
},
{
"loss": 0.2739,
"learning_rate": 7.2e-05,
"epoch": 0.7,
"step": 40
},
{
"loss": 0.1792,
"learning_rate": 7.2e-05,
"epoch": 0.72,
"step": 41
},
{
"loss": 0.1962,
"learning_rate": 7.2e-05,
"epoch": 0.74,
"step": 42
},
{
"loss": 0.2281,
"learning_rate": 7.2e-05,
"epoch": 0.75,
"step": 43
},
{
"loss": 0.1416,
"learning_rate": 7.2e-05,
"epoch": 0.77,
"step": 44
},
{
"loss": 0.2981,
"learning_rate": 7.2e-05,
"epoch": 0.79,
"step": 45
},
{
"loss": 0.2767,
"learning_rate": 7.2e-05,
"epoch": 0.81,
"step": 46
},
{
"loss": 0.2888,
"learning_rate": 7.2e-05,
"epoch": 0.82,
"step": 47
},
{
"loss": 0.3354,
"learning_rate": 7.2e-05,
"epoch": 0.84,
"step": 48
},
{
"loss": 0.2017,
"learning_rate": 7.2e-05,
"epoch": 0.86,
"step": 49
},
{
"loss": 0.1569,
"learning_rate": 7.2e-05,
"epoch": 0.88,
"step": 50
},
{
"eval_code_easy_loss": 0.366973876953125,
"eval_code_easy_score": -0.11782870441675186,
"eval_code_easy_brier_score": 0.11782870441675186,
"eval_code_easy_average_probability": 0.7787162065505981,
"eval_code_easy_accuracy": 0.82,
"eval_code_easy_probabilities": [
0.9949461817741394,
0.9871755242347717,
0.9638071060180664,
0.9957897067070007,
0.9965255856513977,
0.9852939248085022,
0.6930150985717773,
0.9989522695541382,
0.8558874130249023,
0.4613070785999298,
0.510718822479248,
0.2324236035346985,
0.3973901569843292,
0.7853201031684875,
0.5421555042266846,
0.7354709506034851,
0.9784907698631287,
0.9970309734344482,
0.5461677312850952,
0.41464078426361084,
0.30067187547683716,
0.9925328493118286,
0.9674832224845886,
0.9843401908874512,
0.9914712309837341,
0.9825579524040222,
0.9267136454582214,
0.9984757304191589,
0.9873777031898499,
0.9998621940612793,
0.47989073395729065,
0.6140918731689453,
0.39963647723197937,
0.7302553653717041,
0.6465033292770386,
0.5506258606910706,
0.9697403311729431,
0.961764395236969,
0.4946594834327698,
0.8810425400733948,
0.8602923154830933,
0.7874042987823486,
0.9162008166313171,
0.9948663711547852,
0.9439269304275513,
0.9370085000991821,
0.34556975960731506,
0.10199112445116043,
0.6533010005950928,
0.9573982954025269,
0.7248471975326538,
0.2611381411552429,
0.9998348951339722,
0.9983940720558167,
0.9555704593658447,
0.999995231628418,
0.999896764755249,
0.9813823699951172,
0.5570517778396606,
0.9944754242897034,
0.6928525567054749,
0.9708505272865295,
0.9566367268562317,
0.999937891960144,
0.9999886751174927,
0.9999959468841553,
0.9741113781929016,
0.7828025221824646,
0.8869524002075195,
0.6031803488731384,
0.13089396059513092,
0.015088181011378765,
0.9989626407623291,
0.9993138313293457,
0.9666383862495422,
0.7539190649986267,
0.6800207495689392,
0.7904426455497742,
0.9921445250511169,
0.8796939849853516,
0.9624642133712769,
0.48169296979904175,
0.5490749478340149,
0.5619942545890808,
0.7675090432167053,
0.7946746349334717,
0.8194403648376465,
0.787605881690979,
0.2022314816713333,
0.3003339469432831,
0.999944806098938,
0.3906439542770386,
0.9998724460601807,
0.9990270137786865,
0.9992762207984924,
0.9981435537338257,
0.9994109869003296,
0.9930939674377441,
0.9995924830436707,
0.2884075939655304
],
"eval_code_easy_runtime": 12.0912,
"eval_code_easy_samples_per_second": 8.27,
"eval_code_easy_steps_per_second": 0.165,
"epoch": 0.88,
"step": 50
},
{
"eval_code_hard_loss": 0.7648758888244629,
"eval_code_hard_score": -0.2455763965845108,
"eval_code_hard_brier_score": 0.2455763965845108,
"eval_code_hard_average_probability": 0.5492634773254395,
"eval_code_hard_accuracy": 0.58,
"eval_code_hard_probabilities": [
0.9692575335502625,
0.5152681469917297,
0.5391058921813965,
0.9931036829948425,
0.9909272193908691,
0.9973450303077698,
0.37631839513778687,
0.33356037735939026,
0.5611117482185364,
0.4919166564941406,
0.4896278977394104,
0.546688973903656,
0.9802632927894592,
0.9941868782043457,
0.9953175783157349,
0.5917196869850159,
0.556675374507904,
0.5831913352012634,
0.7293359041213989,
0.47196894884109497,
0.4489763081073761,
0.768726646900177,
0.5400140285491943,
0.6983548998832703,
0.38701897859573364,
0.6672759056091309,
0.38983210921287537,
0.005684987176209688,
0.6010024547576904,
0.6203393936157227,
0.590334415435791,
0.5500776767730713,
0.4848686158657074,
0.49481847882270813,
0.4460141360759735,
0.41459348797798157,
0.2456727772951126,
0.24025699496269226,
0.5689725875854492,
0.4383637607097626,
0.5424249172210693,
0.4941108524799347,
0.5239763855934143,
0.4846762716770172,
0.4753747582435608,
0.9676743149757385,
0.40851080417633057,
0.4941249489784241,
0.8557487726211548,
0.8221828937530518,
0.8013899922370911,
0.49702978134155273,
0.3039790391921997,
0.30129650235176086,
0.02055172249674797,
0.021210504695773125,
0.021720753982663155,
0.575912594795227,
0.5162267088890076,
0.5338138341903687,
0.8233920335769653,
0.8342075347900391,
0.8165725469589233,
0.6269811391830444,
0.6718840599060059,
0.5006963014602661,
0.3764708340167999,
0.48138299584388733,
0.33396115899086,
0.749821126461029,
0.6757022142410278,
0.48911863565444946,
0.5064474940299988,
0.62013179063797,
0.5993181467056274,
0.4782550632953644,
0.4940183758735657,
0.5602287650108337,
0.5086959004402161,
0.46642473340034485,
0.5594733953475952,
0.5183232426643372,
0.5857313275337219,
0.5656166076660156,
0.5952458381652832,
0.5976989269256592,
0.5781928300857544,
0.48465627431869507,
0.5290895700454712,
0.5022053718566895,
0.47955581545829773,
0.465986043214798,
0.46290040016174316,
0.33050066232681274,
0.34588003158569336,
0.387125700712204,
0.6185724139213562,
0.683625340461731,
0.32840457558631897,
0.723832905292511
],
"eval_code_hard_runtime": 53.3211,
"eval_code_hard_samples_per_second": 1.875,
"eval_code_hard_steps_per_second": 0.038,
"epoch": 0.88,
"step": 50
},
{
"loss": 0.2551,
"learning_rate": 7.2e-05,
"epoch": 0.89,
"step": 51
},
{
"loss": 0.1634,
"learning_rate": 7.2e-05,
"epoch": 0.91,
"step": 52
},
{
"loss": 0.3866,
"learning_rate": 7.2e-05,
"epoch": 0.93,
"step": 53
},
{
"loss": 0.4277,
"learning_rate": 7.2e-05,
"epoch": 0.95,
"step": 54
},
{
"loss": 0.2225,
"learning_rate": 7.2e-05,
"epoch": 0.96,
"step": 55
},
{
"loss": 0.1702,
"learning_rate": 7.2e-05,
"epoch": 0.98,
"step": 56
},
{
"loss": 0.0772,
"learning_rate": 7.2e-05,
"epoch": 1.0,
"step": 57
},
{
"loss": 0.082,
"learning_rate": 7.2e-05,
"epoch": 1.02,
"step": 58
},
{
"loss": 0.3056,
"learning_rate": 7.2e-05,
"epoch": 1.04,
"step": 59
},
{
"loss": 0.1856,
"learning_rate": 7.2e-05,
"epoch": 1.05,
"step": 60
},
{
"loss": 0.1862,
"learning_rate": 7.2e-05,
"epoch": 1.07,
"step": 61
},
{
"loss": 0.1939,
"learning_rate": 7.2e-05,
"epoch": 1.09,
"step": 62
},
{
"loss": 0.1331,
"learning_rate": 7.2e-05,
"epoch": 1.11,
"step": 63
},
{
"loss": 0.2338,
"learning_rate": 7.2e-05,
"epoch": 1.12,
"step": 64
},
{
"loss": 0.1673,
"learning_rate": 7.2e-05,
"epoch": 1.14,
"step": 65
},
{
"loss": 0.3696,
"learning_rate": 7.2e-05,
"epoch": 1.16,
"step": 66
},
{
"loss": 0.1056,
"learning_rate": 7.2e-05,
"epoch": 1.18,
"step": 67
},
{
"loss": 0.1094,
"learning_rate": 7.2e-05,
"epoch": 1.19,
"step": 68
},
{
"loss": 0.0796,
"learning_rate": 7.2e-05,
"epoch": 1.21,
"step": 69
},
{
"loss": 0.2261,
"learning_rate": 7.2e-05,
"epoch": 1.23,
"step": 70
},
{
"loss": 0.1961,
"learning_rate": 7.2e-05,
"epoch": 1.25,
"step": 71
},
{
"loss": 0.1452,
"learning_rate": 7.2e-05,
"epoch": 1.26,
"step": 72
},
{
"loss": 0.1716,
"learning_rate": 7.2e-05,
"epoch": 1.28,
"step": 73
},
{
"loss": 0.1421,
"learning_rate": 7.2e-05,
"epoch": 1.3,
"step": 74
},
{
"loss": 0.3305,
"learning_rate": 7.2e-05,
"epoch": 1.32,
"step": 75
},
{
"eval_code_easy_loss": 0.32117652893066406,
"eval_code_easy_score": -0.09214462339878082,
"eval_code_easy_brier_score": 0.09214462339878082,
"eval_code_easy_average_probability": 0.8216841220855713,
"eval_code_easy_accuracy": 0.88,
"eval_code_easy_probabilities": [
0.9994999170303345,
0.9912821054458618,
0.9577344059944153,
0.9980210065841675,
0.9973551034927368,
0.9871847033500671,
0.6745311617851257,
0.9999939203262329,
0.9304938912391663,
0.5086526870727539,
0.5627270340919495,
0.2480248361825943,
0.49823904037475586,
0.8804905414581299,
0.5820022821426392,
0.748561680316925,
0.9955660700798035,
0.9997114539146423,
0.8057821393013,
0.5374618768692017,
0.2833597958087921,
0.9967227578163147,
0.9863321781158447,
0.9948417544364929,
0.9996953010559082,
0.9988725781440735,
0.9955350160598755,
0.9988835453987122,
0.9974863529205322,
0.9998946189880371,
0.8057632446289062,
0.7699967622756958,
0.6971409916877747,
0.7715607285499573,
0.6907534599304199,
0.5492137670516968,
0.9612501263618469,
0.9800459146499634,
0.6639916896820068,
0.992854654788971,
0.9864367842674255,
0.9040735363960266,
0.9259136319160461,
0.9870310425758362,
0.8883805274963379,
0.9945270419120789,
0.48652422428131104,
0.044553183019161224,
0.6629449129104614,
0.9678350687026978,
0.7551115155220032,
0.35034820437431335,
0.9999849796295166,
0.9999747276306152,
0.9967097043991089,
0.9999997615814209,
0.9999991655349731,
0.9840741157531738,
0.5545885562896729,
0.9983722567558289,
0.6852497458457947,
0.9241886138916016,
0.8779072165489197,
0.9999986886978149,
0.9999998807907104,
1.0,
0.9829393625259399,
0.9291936755180359,
0.9386154413223267,
0.6534070372581482,
0.10347169637680054,
0.0036570930387824774,
0.9999821186065674,
0.9999948740005493,
0.9956468939781189,
0.9652343988418579,
0.9181939363479614,
0.9902083873748779,
0.9989393353462219,
0.853695273399353,
0.9876751899719238,
0.4735547602176666,
0.5543529391288757,
0.5561679601669312,
0.8694782853126526,
0.9327576756477356,
0.9288151264190674,
0.8425791263580322,
0.2500445544719696,
0.4530153274536133,
0.9999988079071045,
0.4624493420124054,
0.9999938011169434,
0.9999393224716187,
0.9999752044677734,
0.9998465776443481,
0.999996542930603,
0.9989469647407532,
0.999995231628418,
0.5134104490280151
],
"eval_code_easy_runtime": 12.0923,
"eval_code_easy_samples_per_second": 8.27,
"eval_code_easy_steps_per_second": 0.165,
"epoch": 1.32,
"step": 75
},
{
"eval_code_hard_loss": 0.8514172434806824,
"eval_code_hard_score": -0.2549550235271454,
"eval_code_hard_brier_score": 0.2549550235271454,
"eval_code_hard_average_probability": 0.5471080541610718,
"eval_code_hard_accuracy": 0.63,
"eval_code_hard_probabilities": [
0.9908287525177002,
0.5119895339012146,
0.5428490042686462,
0.9895175099372864,
0.9891449809074402,
0.9949464201927185,
0.38589784502983093,
0.2820265591144562,
0.5328211784362793,
0.5014193058013916,
0.5086326003074646,
0.5764115452766418,
0.9859844446182251,
0.9957466721534729,
0.9967710375785828,
0.5794450640678406,
0.5391579270362854,
0.6133323311805725,
0.6606585383415222,
0.34289708733558655,
0.2988581657409668,
0.788067102432251,
0.542411744594574,
0.7060100436210632,
0.42542076110839844,
0.7323933839797974,
0.42784348130226135,
0.0037993849255144596,
0.6944400072097778,
0.7083581686019897,
0.5933657288551331,
0.5459165573120117,
0.5039231181144714,
0.4072609841823578,
0.3503766357898712,
0.35972559452056885,
0.1359257698059082,
0.13460229337215424,
0.4615738093852997,
0.38321685791015625,
0.37138813734054565,
0.5392828583717346,
0.5346622467041016,
0.4690147042274475,
0.4728128910064697,
0.9811822772026062,
0.36270952224731445,
0.5393079519271851,
0.8588457703590393,
0.9537001252174377,
0.9503171443939209,
0.4992208480834961,
0.16286799311637878,
0.3865070939064026,
0.0036291300784796476,
0.0037814583629369736,
0.003941776696592569,
0.6453168392181396,
0.5069921016693115,
0.5409190654754639,
0.620768666267395,
0.6463491320610046,
0.6067190766334534,
0.6744793653488159,
0.7356851696968079,
0.6201030611991882,
0.34589684009552,
0.4320986270904541,
0.2923872470855713,
0.8295382857322693,
0.7723568677902222,
0.48901495337486267,
0.5491488575935364,
0.6467846035957336,
0.6319132447242737,
0.5377576351165771,
0.5089011788368225,
0.6081604957580566,
0.5122814774513245,
0.45855069160461426,
0.5822516083717346,
0.5393768548965454,
0.579654335975647,
0.5982511043548584,
0.6504716873168945,
0.658900797367096,
0.6075130105018616,
0.4958891272544861,
0.5496628880500793,
0.5090362429618835,
0.46470561623573303,
0.4267897605895996,
0.43224194645881653,
0.3274325728416443,
0.35687658190727234,
0.4236689805984497,
0.6108702421188354,
0.7473148107528687,
0.3625641465187073,
0.7580684423446655
],
"eval_code_hard_runtime": 53.2977,
"eval_code_hard_samples_per_second": 1.876,
"eval_code_hard_steps_per_second": 0.038,
"epoch": 1.32,
"step": 75
},
{
"loss": 0.149,
"learning_rate": 7.2e-05,
"epoch": 1.33,
"step": 76
},
{
"loss": 0.2295,
"learning_rate": 7.2e-05,
"epoch": 1.35,
"step": 77
},
{
"loss": 0.1289,
"learning_rate": 7.2e-05,
"epoch": 1.37,
"step": 78
},
{
"loss": 0.0937,
"learning_rate": 7.2e-05,
"epoch": 1.39,
"step": 79
},
{
"loss": 0.2041,
"learning_rate": 7.2e-05,
"epoch": 1.4,
"step": 80
},
{
"loss": 0.2389,
"learning_rate": 7.2e-05,
"epoch": 1.42,
"step": 81
},
{
"loss": 0.1343,
"learning_rate": 7.2e-05,
"epoch": 1.44,
"step": 82
},
{
"loss": 0.0283,
"learning_rate": 7.2e-05,
"epoch": 1.46,
"step": 83
},
{
"loss": 0.0841,
"learning_rate": 7.2e-05,
"epoch": 1.47,
"step": 84
},
{
"loss": 0.3316,
"learning_rate": 7.2e-05,
"epoch": 1.49,
"step": 85
},
{
"loss": 0.0393,
"learning_rate": 7.2e-05,
"epoch": 1.51,
"step": 86
},
{
"loss": 0.0755,
"learning_rate": 7.2e-05,
"epoch": 1.53,
"step": 87
},
{
"loss": 0.1595,
"learning_rate": 7.2e-05,
"epoch": 1.54,
"step": 88
},
{
"loss": 0.1434,
"learning_rate": 7.2e-05,
"epoch": 1.56,
"step": 89
},
{
"loss": 0.1001,
"learning_rate": 7.2e-05,
"epoch": 1.58,
"step": 90
},
{
"loss": 0.1532,
"learning_rate": 7.2e-05,
"epoch": 1.6,
"step": 91
},
{
"loss": 0.1773,
"learning_rate": 7.2e-05,
"epoch": 1.61,
"step": 92
},
{
"loss": 0.1351,
"learning_rate": 7.2e-05,
"epoch": 1.63,
"step": 93
},
{
"loss": 0.0533,
"learning_rate": 7.2e-05,
"epoch": 1.65,
"step": 94
},
{
"loss": 0.1834,
"learning_rate": 7.2e-05,
"epoch": 1.67,
"step": 95
},
{
"loss": 0.1847,
"learning_rate": 7.2e-05,
"epoch": 1.68,
"step": 96
},
{
"loss": 0.072,
"learning_rate": 7.2e-05,
"epoch": 1.7,
"step": 97
},
{
"loss": 0.0431,
"learning_rate": 7.2e-05,
"epoch": 1.72,
"step": 98
},
{
"loss": 0.0796,
"learning_rate": 7.2e-05,
"epoch": 1.74,
"step": 99
},
{
"loss": 0.1992,
"learning_rate": 7.2e-05,
"epoch": 1.75,
"step": 100
},
{
"eval_code_easy_loss": 0.3075363039970398,
"eval_code_easy_score": -0.07319054007530212,
"eval_code_easy_brier_score": 0.07319054007530212,
"eval_code_easy_average_probability": 0.8647841811180115,
"eval_code_easy_accuracy": 0.92,
"eval_code_easy_probabilities": [
0.9999574422836304,
0.9999752044677734,
0.9896870255470276,
0.9995444416999817,
0.9996813535690308,
0.9912283420562744,
0.8707631230354309,
1.0,
0.9232298135757446,
0.8876208066940308,
0.7544382810592651,
0.8350661993026733,
0.8029730319976807,
0.9684423804283142,
0.8641867637634277,
0.840520977973938,
0.9997034668922424,
0.9998855590820312,
0.8682356476783752,
0.5141866207122803,
0.25204265117645264,
0.9994714856147766,
0.99544358253479,
0.9964646100997925,
0.9997562766075134,
0.999864935874939,
0.9721921682357788,
0.9998952150344849,
0.9750569462776184,
0.9999854564666748,
0.7429996132850647,
0.822045624256134,
0.7596055865287781,
0.8471194505691528,
0.7925592660903931,
0.795616865158081,
0.976042628288269,
0.9971067309379578,
0.5366230607032776,
0.9954087138175964,
0.977104127407074,
0.9664822816848755,
0.9948425889015198,
0.9999990463256836,
0.9974696636199951,
0.9896783828735352,
0.7055782675743103,
0.010203487239778042,
0.9344508051872253,
0.999211311340332,
0.9515018463134766,
0.5253821611404419,
1.0,
0.9999994039535522,
0.9999967813491821,
1.0,
1.0,
0.9986414313316345,
0.5329476594924927,
0.9997962117195129,
0.9566650986671448,
0.9977226853370667,
0.9972772002220154,
1.0,
1.0,
1.0,
0.9998082518577576,
0.9984453320503235,
0.9989882111549377,
0.6647514700889587,
0.06555651873350143,
0.0002535696839913726,
0.9999556541442871,
0.9999914169311523,
0.9952349066734314,
0.9737288951873779,
0.9164130091667175,
0.9624624252319336,
0.9998660087585449,
0.9491363763809204,
0.9998786449432373,
0.4198010265827179,
0.5238525867462158,
0.5886799693107605,
0.8423507213592529,
0.9458127617835999,
0.9036057591438293,
0.9418210387229919,
0.18787764012813568,
0.38167285919189453,
1.0,
0.3317376375198364,
0.9999998807907104,
0.9999892711639404,
0.9999970197677612,
0.9999692440032959,
0.9997808337211609,
0.9896361231803894,
0.9996919631958008,
0.7980904579162598
],
"eval_code_easy_runtime": 12.0952,
"eval_code_easy_samples_per_second": 8.268,
"eval_code_easy_steps_per_second": 0.165,
"epoch": 1.75,
"step": 100
},
{
"eval_code_hard_loss": 0.8713796734809875,
"eval_code_hard_score": -0.23470951616764069,
"eval_code_hard_brier_score": 0.23470951616764069,
"eval_code_hard_average_probability": 0.5856627821922302,
"eval_code_hard_accuracy": 0.65,
"eval_code_hard_probabilities": [
0.9855136275291443,
0.49629148840904236,
0.5387540459632874,
0.9978917241096497,
0.9978112578392029,
0.9988699555397034,
0.47865962982177734,
0.3495783805847168,
0.606380045413971,
0.5135536789894104,
0.5702087879180908,
0.6042059659957886,
0.9980809688568115,
0.99955815076828,
0.9996216297149658,
0.5243353247642517,
0.48392897844314575,
0.6072529554367065,
0.9629855751991272,
0.8894397616386414,
0.8946234583854675,
0.8234318494796753,
0.5752303004264832,
0.7489052414894104,
0.4938078820705414,
0.764939546585083,
0.49246543645858765,
0.0005044332938268781,
0.6989362835884094,
0.6961206197738647,
0.695743203163147,
0.5966794490814209,
0.4954136610031128,
0.428554505109787,
0.3025328516960144,
0.31304794549942017,
0.05157984420657158,
0.057667434215545654,
0.295099675655365,
0.4117719233036041,
0.4178787171840668,
0.44591596722602844,
0.5260471701622009,
0.46641460061073303,
0.44238483905792236,
0.9921270608901978,
0.3312618136405945,
0.49680960178375244,
0.8725456595420837,
0.990997314453125,
0.9901727437973022,
0.42896220088005066,
0.10010050982236862,
0.30535706877708435,
0.000553447927813977,
0.000573858036659658,
0.000650162051897496,
0.6373094916343689,
0.5249813199043274,
0.5481424331665039,
0.9771735668182373,
0.9809019565582275,
0.9766740798950195,
0.7092797756195068,
0.7087807059288025,
0.700412392616272,
0.4039199948310852,
0.626448929309845,
0.3165664076805115,
0.8700190782546997,
0.7991524934768677,
0.5164951086044312,
0.5391688346862793,
0.6079928874969482,
0.6740086674690247,
0.5694684386253357,
0.5098953247070312,
0.6005474925041199,
0.5051020979881287,
0.4350145161151886,
0.6045871376991272,
0.5573720932006836,
0.763244092464447,
0.6566476225852966,
0.655601978302002,
0.6529899835586548,
0.6917677521705627,
0.4615822732448578,
0.5680032968521118,
0.5078142285346985,
0.48479217290878296,
0.437853068113327,
0.4693388044834137,
0.5179882645606995,
0.5434749126434326,
0.650637686252594,
0.6607488393783569,
0.7874378561973572,
0.1389518529176712,
0.76926189661026
],
"eval_code_hard_runtime": 53.3137,
"eval_code_hard_samples_per_second": 1.876,
"eval_code_hard_steps_per_second": 0.038,
"epoch": 1.75,
"step": 100
},
{
"train_runtime": 1129.7886,
"train_samples_per_second": 2.832,
"train_steps_per_second": 0.089,
"total_flos": 0.0,
"train_loss": 0.30551302678883074,
"epoch": 1.75,
"step": 100
}
]