GENIES/results/evaluations/zero_shot/openllama-3b-code/us_history.json

{
    "model_dir": "models/openllama-3b",
    "distribution_id": "us_history",
    "timestamp": "2023-10-20 01:24:16",
    "eval_loss": 0.6691340208053589,
    "eval_score": -0.23976899683475494,
    "eval_brier_score": 0.23976901092271624,
    "eval_average_probability": 0.5122894008159637,
    "eval_accuracy": 0.6493333333333333,
    "eval_probabilities": [
        0.5682207942008972,
        0.5524383783340454,
        0.5527743697166443,
        0.6282231211662292,
        0.6205865144729614,
        0.6281256675720215,
        0.4948121905326843,
        0.5089397430419922,
        0.5207363963127136,
        0.603369951248169,
        0.603807806968689,
        0.5962295532226562,
        0.48740890622138977,
        0.4844508469104767,
        0.47564274072647095,
        0.5404852032661438,
        0.5406257510185242,
        0.555126428604126,
        0.6065139174461365,
        0.5948545336723328,
        0.5728118419647217,
        0.49452200531959534,
        0.5123866200447083,
        0.43495631217956543,
        0.49767258763313293,
        0.5095402598381042,
        0.5133123397827148,
        0.41805294156074524,
        0.42625030875205994,
        0.42486366629600525,
        0.5939502120018005,
        0.6311264634132385,
        0.6140653491020203,
        0.4655592143535614,
        0.45416060090065,
        0.4467146098613739,
        0.4845709800720215,
        0.4975309371948242,
        0.49294009804725647,
        0.5167430639266968,
        0.5015862584114075,
        0.5138328671455383,
        0.49964824318885803,
        0.4971540570259094,
        0.5040467381477356,
        0.5457506775856018,
        0.5554729700088501,
        0.5490840077400208,
        0.5072984099388123,
        0.5236062407493591,
        0.5211291313171387,
        0.47636517882347107,
        0.5175215005874634,
        0.44194963574409485,
        0.5297903418540955,
        0.4944212734699249,
        0.511172890663147,
        0.5105995535850525,
        0.5177040100097656,
        0.5137489438056946,
        0.5113855600357056,
        0.516503632068634,
        0.5434898138046265,
        0.47789326310157776,
        0.5063337087631226,
        0.4674566090106964,
        0.5547204613685608,
        0.5512599349021912,
        0.5821061134338379,
        0.6095503568649292,
        0.5962245464324951,
        0.5899627804756165,
        0.5872119665145874,
        0.6050300002098083,
        0.5320652723312378,
        0.4996505081653595,
        0.46085891127586365,
        0.5016400814056396,
        0.484977126121521,
        0.4759133756160736,
        0.46948254108428955,
        0.6150104999542236,
        0.6090078353881836,
        0.5799394845962524,
        0.5925146341323853,
        0.5728327631950378,
        0.5371448397636414,
        0.5196922421455383,
        0.5188688635826111,
        0.5051478743553162,
        0.4709605872631073,
        0.4673032760620117,
        0.4653334319591522,
        0.47317516803741455,
        0.4496268332004547,
        0.4515853822231293,
        0.5617247223854065,
        0.5806660056114197,
        0.556975781917572,
        0.5028976798057556,
        0.49992385506629944,
        0.5173745155334473,
        0.470426082611084,
        0.4399070143699646,
        0.45083755254745483,
        0.5250056982040405,
        0.5489998459815979,
        0.5483051538467407,
        0.4967859387397766,
        0.5032984614372253,
        0.49278882145881653,
        0.5704954266548157,
        0.5423449277877808,
        0.5719475150108337,
        0.6220575571060181,
        0.6182311177253723,
        0.5959829688072205,
        0.5712084770202637,
        0.551188051700592,
        0.5560855269432068,
        0.48938998579978943,
        0.4956319332122803,
        0.496985524892807,
        0.49886345863342285,
        0.5055021047592163,
        0.512376606464386,
        0.4169715642929077,
        0.42551568150520325,
        0.4222303628921509,
        0.5125755071640015,
        0.500823438167572,
        0.5069188475608826,
        0.480791836977005,
        0.47227057814598083,
        0.4808599352836609,
        0.5200278759002686,
        0.5132653117179871,
        0.4794500470161438,
        0.5918850302696228,
        0.5792427062988281,
        0.5741116404533386,
        0.49508723616600037,
        0.49890443682670593,
        0.5023145079612732,
        0.5186350345611572,
        0.4841456711292267,
        0.49816441535949707,
        0.4829530715942383,
        0.4789375960826874,
        0.4804285764694214,
        0.5066986680030823,
        0.4968593418598175,
        0.49812352657318115,
        0.5266085267066956,
        0.5106497406959534,
        0.5165297985076904,
        0.5260772109031677,
        0.5233297944068909,
        0.5758408904075623,
        0.44542619585990906,
        0.43315792083740234,
        0.4240458607673645,
        0.5072693228721619,
        0.502674400806427,
        0.49394193291664124,
        0.6015763282775879,
        0.6164434552192688,
        0.5907071232795715,
        0.5149018168449402,
        0.509438157081604,
        0.5121998190879822,
        0.5303035378456116,
        0.5354618430137634,
        0.5224488377571106,
        0.5102747082710266,
        0.5028833150863647,
        0.50312739610672,
        0.510371208190918,
        0.516396701335907,
        0.5237711668014526,
        0.47373539209365845,
        0.4958321452140808,
        0.4878159463405609,
        0.4951622784137726,
        0.4913522005081177,
        0.4856465756893158,
        0.4906422197818756,
        0.49810436367988586,
        0.5036133527755737,
        0.4617498517036438,
        0.5191302299499512,
        0.49073681235313416,
        0.5001580715179443,
        0.4983668625354767,
        0.5058532357215881,
        0.5895419120788574,
        0.545975923538208,
        0.6258460879325867,
        0.45666852593421936,
        0.4480047821998596,
        0.4362683594226837,
        0.49598172307014465,
        0.47700148820877075,
        0.46995168924331665,
        0.511577308177948,
        0.5136296153068542,
        0.5206295847892761,
        0.5048058032989502,
        0.5101553797721863,
        0.4974718391895294,
        0.5006347894668579,
        0.5023354887962341,
        0.5081377029418945,
        0.5168053507804871,
        0.5144450664520264,
        0.5152599811553955,
        0.5406501889228821,
        0.5368693470954895,
        0.5221444964408875,
        0.49383366107940674,
        0.49631717801094055,
        0.48102670907974243,
        0.5094847083091736,
        0.5004937052726746,
        0.5085771083831787,
        0.42393290996551514,
        0.39194256067276,
        0.42215195298194885,
        0.524935781955719,
        0.5213128924369812,
        0.521791934967041,
        0.541603147983551,
        0.5421897172927856,
        0.5410572290420532,
        0.4814724326133728,
        0.47767576575279236,
        0.4728677272796631,
        0.4894760847091675,
        0.522022545337677,
        0.5369196534156799,
        0.5125432014465332,
        0.5021210312843323,
        0.5009797811508179,
        0.522920548915863,
        0.5170161724090576,
        0.5468817949295044,
        0.5647842884063721,
        0.5693818926811218,
        0.5655085444450378,
        0.5321863293647766,
        0.49910682439804077,
        0.5234453082084656,
        0.4628234803676605,
        0.4503444731235504,
        0.47892147302627563,
        0.5549036264419556,
        0.6062313914299011,
        0.5898271799087524,
        0.5250076055526733,
        0.5124813914299011,
        0.5152792930603027,
        0.47009021043777466,
        0.476891428232193,
        0.47205662727355957,
        0.487435907125473,
        0.4948550760746002,
        0.48252177238464355,
        0.4577889144420624,
        0.4830418527126312,
        0.4206792712211609,
        0.4974518418312073,
        0.4948914349079132,
        0.49761781096458435,
        0.4169645607471466,
        0.4440787136554718,
        0.44131380319595337,
        0.5070161819458008,
        0.49705079197883606,
        0.4978044927120209,
        0.49653491377830505,
        0.4974987208843231,
        0.522584855556488,
        0.5073515772819519,
        0.512330174446106,
        0.5156213045120239,
        0.5089912414550781,
        0.5153915882110596,
        0.5192590951919556,
        0.5092111825942993,
        0.5075430870056152,
        0.5094568133354187,
        0.4429752826690674,
        0.4853109121322632,
        0.4676140546798706,
        0.49446970224380493,
        0.5012256503105164,
        0.4973224103450775,
        0.5399768948554993,
        0.4991808235645294,
        0.5012447237968445,
        0.5698264837265015,
        0.5484668612480164,
        0.5210722088813782,
        0.4943370819091797,
        0.5011413097381592,
        0.5007398724555969,
        0.49397215247154236,
        0.5082046389579773,
        0.5042240619659424,
        0.44791722297668457,
        0.44826769828796387,
        0.46703189611434937,
        0.5074156522750854,
        0.5016489624977112,
        0.5018003582954407,
        0.5019153356552124,
        0.5052042007446289,
        0.5015776753425598,
        0.5106533169746399,
        0.5077043175697327,
        0.5023712515830994,
        0.5134146809577942,
        0.5129215717315674,
        0.5151938796043396,
        0.5198009014129639,
        0.4907361567020416,
        0.5175072550773621,
        0.5115537643432617,
        0.5083064436912537,
        0.5243549346923828,
        0.5321025252342224,
        0.5121833682060242,
        0.5271683931350708,
        0.5432240962982178,
        0.5987139940261841,
        0.5914070010185242,
        0.5037053227424622,
        0.49722665548324585,
        0.5000015497207642,
        0.5413800477981567,
        0.5513725876808167,
        0.5355043411254883,
        0.5109012126922607,
        0.5330449342727661,
        0.5216690301895142,
        0.48434945940971375,
        0.4600626826286316,
        0.47221657633781433,
        0.4853765368461609,
        0.4913669526576996,
        0.4743122458457947,
        0.509884238243103,
        0.5052781105041504,
        0.5006000399589539,
        0.48430415987968445,
        0.511134147644043,
        0.4976176619529724,
        0.5939187407493591,
        0.6085214614868164,
        0.6199467778205872,
        0.5050240159034729,
        0.5080088376998901,
        0.5157637596130371,
        0.5357295870780945,
        0.5330662727355957,
        0.521038293838501,
        0.5611529350280762,
        0.5741451382637024,
        0.5513404011726379,
        0.4962119162082672,
        0.490121990442276,
        0.509446382522583,
        0.497160941362381,
        0.505458652973175,
        0.5020835995674133,
        0.5216751098632812,
        0.5120912194252014,
        0.5206372737884521,
        0.5173282027244568,
        0.5335246920585632,
        0.5046808123588562,
        0.5083194971084595,
        0.49712368845939636,
        0.4940822720527649,
        0.5132601857185364,
        0.5220094919204712,
        0.5681636333465576,
        0.5109795331954956,
        0.5068950653076172,
        0.5096277594566345,
        0.4923011064529419,
        0.5162477493286133,
        0.518042266368866,
        0.5140049457550049,
        0.5102848410606384,
        0.496192067861557,
        0.567232608795166,
        0.5623913407325745,
        0.5691769123077393,
        0.49946242570877075,
        0.4994986355304718,
        0.4939383566379547,
        0.5262704491615295,
        0.538678765296936,
        0.5120803117752075,
        0.5018618106842041,
        0.5017870664596558,
        0.5015931129455566,
        0.48532432317733765,
        0.5416369438171387,
        0.5165446400642395,
        0.66316819190979,
        0.6627864241600037,
        0.6338526010513306,
        0.5182491540908813,
        0.5263990163803101,
        0.524048924446106,
        0.5017751455307007,
        0.5031629204750061,
        0.5074881911277771,
        0.5018927454948425,
        0.5079371929168701,
        0.4964205324649811,
        0.5144279599189758,
        0.5238584876060486,
        0.5211344957351685,
        0.5411505699157715,
        0.5206217765808105,
        0.5067276954650879,
        0.43451744318008423,
        0.4586152136325836,
        0.4607141613960266,
        0.5111634135246277,
        0.4994567334651947,
        0.4900016784667969,
        0.49982431530952454,
        0.5149359107017517,
        0.49920544028282166,
        0.494920015335083,
        0.5020367503166199,
        0.504837691783905,
        0.5000603199005127,
        0.5050713419914246,
        0.509554922580719,
        0.6277421712875366,
        0.6120190024375916,
        0.6154093742370605,
        0.4935853183269501,
        0.49616965651512146,
        0.49244531989097595,
        0.5399248600006104,
        0.5701155066490173,
        0.535513162612915,
        0.44078290462493896,
        0.46000850200653076,
        0.46633461117744446,
        0.5791850686073303,
        0.5774042010307312,
        0.5818692445755005,
        0.5055866241455078,
        0.5186561346054077,
        0.511214017868042,
        0.45469290018081665,
        0.4666598439216614,
        0.464605450630188,
        0.6251022219657898,
        0.602652370929718,
        0.6022546291351318,
        0.4883878529071808,
        0.42277687788009644,
        0.46406516432762146,
        0.4887782633304596,
        0.4899451434612274,
        0.49312421679496765,
        0.5044760704040527,
        0.49792447686195374,
        0.5069671869277954,
        0.4787726104259491,
        0.469808965921402,
        0.47069862484931946,
        0.5131545662879944,
        0.4976878762245178,
        0.5071484446525574,
        0.47176966071128845,
        0.4853769540786743,
        0.46412035822868347,
        0.503024697303772,
        0.5074666142463684,
        0.5105546116828918,
        0.4784400761127472,
        0.5132582187652588,
        0.47739025950431824,
        0.5275378823280334,
        0.5148946642875671,
        0.5132803320884705,
        0.5081470608711243,
        0.5168777108192444,
        0.5128099322319031,
        0.5493481755256653,
        0.558631420135498,
        0.5315588712692261,
        0.48212572932243347,
        0.49468353390693665,
        0.5011990070343018,
        0.5875621438026428,
        0.5991061925888062,
        0.56623375415802,
        0.5055972337722778,
        0.5155515074729919,
        0.48452141880989075,
        0.492844820022583,
        0.49876609444618225,
        0.47960329055786133,
        0.49972647428512573,
        0.5086188316345215,
        0.5151358246803284,
        0.3475954234600067,
        0.3430183231830597,
        0.3414551615715027,
        0.39604073762893677,
        0.404378205537796,
        0.39594799280166626,
        0.5015736222267151,
        0.49897176027297974,
        0.49956220388412476,
        0.5128662586212158,
        0.5095438957214355,
        0.5173943042755127,
        0.5162033438682556,
        0.5041149854660034,
        0.534008264541626,
        0.49908217787742615,
        0.5007685422897339,
        0.5118399858474731,
        0.517787516117096,
        0.5185492038726807,
        0.5121815204620361,
        0.4523976147174835,
        0.4877399504184723,
        0.47317245602607727,
        0.5110830664634705,
        0.51609867811203,
        0.5250146389007568,
        0.5476350784301758,
        0.5499494671821594,
        0.5619015097618103,
        0.5517339706420898,
        0.5704801082611084,
        0.530616044998169,
        0.5074082612991333,
        0.5381484031677246,
        0.5263859033584595,
        0.4244441092014313,
        0.4091035723686218,
        0.4186810851097107,
        0.5120778679847717,
        0.5094859004020691,
        0.5140645503997803,
        0.4976401627063751,
        0.5235224366188049,
        0.5144694447517395,
        0.5492826104164124,
        0.5228663086891174,
        0.5366529822349548,
        0.4923694133758545,
        0.4856140911579132,
        0.4933948218822479,
        0.5197526216506958,
        0.532535970211029,
        0.5182756185531616,
        0.5027977228164673,
        0.5011717677116394,
        0.5011762976646423,
        0.49992701411247253,
        0.5003582835197449,
        0.5049194097518921,
        0.38544586300849915,
        0.3821794092655182,
        0.4180629551410675,
        0.4963180124759674,
        0.49633345007896423,
        0.49666014313697815,
        0.49451354146003723,
        0.4981456696987152,
        0.49853000044822693,
        0.5206111669540405,
        0.5217689275741577,
        0.5109986662864685,
        0.5283075571060181,
        0.5414741039276123,
        0.5616273283958435,
        0.618395209312439,
        0.6232250332832336,
        0.6326168179512024,
        0.5371412634849548,
        0.5254169702529907,
        0.5353584885597229,
        0.5104691386222839,
        0.5209883451461792,
        0.5234472751617432,
        0.5277793407440186,
        0.5439046025276184,
        0.5438912510871887,
        0.5045725703239441,
        0.4919312298297882,
        0.4823627769947052,
        0.5147815346717834,
        0.5273908376693726,
        0.5208408236503601,
        0.5197887420654297,
        0.4930064082145691,
        0.48449572920799255,
        0.549476146697998,
        0.5602967143058777,
        0.5319384932518005,
        0.5139468908309937,
        0.522526741027832,
        0.50569748878479,
        0.40324047207832336,
        0.4256508946418762,
        0.40353626012802124,
        0.5128648281097412,
        0.47129032015800476,
        0.49060341715812683,
        0.5039991736412048,
        0.49952489137649536,
        0.5070732235908508,
        0.5591380000114441,
        0.5196278691291809,
        0.5702155828475952,
        0.5216491222381592,
        0.519769549369812,
        0.5172757506370544,
        0.5008880496025085,
        0.5001219511032104,
        0.5001779794692993,
        0.5008786916732788,
        0.5012657046318054,
        0.5050615668296814,
        0.550430953502655,
        0.5495932698249817,
        0.5766720771789551,
        0.49401021003723145,
        0.48237502574920654,
        0.5035213232040405,
        0.5724731087684631,
        0.5619921684265137,
        0.5645650625228882,
        0.5964047908782959,
        0.6001353859901428,
        0.5621862411499023,
        0.44778433442115784,
        0.47060221433639526,
        0.4718515872955322,
        0.5730733871459961,
        0.5744489431381226,
        0.5367487072944641,
        0.4170621633529663,
        0.4312669634819031,
        0.4308631718158722,
        0.49964094161987305,
        0.4944249391555786,
        0.4808571934700012,
        0.4556325376033783,
        0.4561145007610321,
        0.42659854888916016,
        0.47014811635017395,
        0.4612431526184082,
        0.49073177576065063,
        0.5102331638336182,
        0.5184701085090637,
        0.521766722202301,
        0.49913284182548523,
        0.47587478160858154,
        0.5048987865447998,
        0.5177756547927856,
        0.5156026482582092,
        0.5143850445747375,
        0.48052722215652466,
        0.49964776635169983,
        0.4856102466583252,
        0.4953272342681885,
        0.505927562713623,
        0.5026220083236694,
        0.4955656826496124,
        0.507610559463501,
        0.48941734433174133,
        0.5131369233131409,
        0.5216907858848572,
        0.5214042663574219,
        0.5298634171485901,
        0.5309587717056274,
        0.5289116501808167,
        0.5170915722846985,
        0.5151835083961487,
        0.5230124592781067,
        0.5222985744476318,
        0.5238386392593384,
        0.522898256778717,
        0.5497962832450867,
        0.5364474058151245,
        0.5607157349586487,
        0.5410718321800232,
        0.5469083189964294,
        0.5368506908416748,
        0.5288794040679932,
        0.5127543210983276,
        0.5570014119148254,
        0.5767610669136047,
        0.5829082727432251,
        0.5718637108802795,
        0.49822744727134705,
        0.4999799430370331,
        0.5007001757621765,
        0.5225939750671387,
        0.5257173180580139,
        0.49963563680648804,
        0.4259820282459259,
        0.46162474155426025,
        0.4770861268043518,
        0.5232495665550232,
        0.5010129809379578,
        0.5211510062217712,
        0.5226557850837708,
        0.5137490630149841,
        0.5108913779258728,
        0.5198879837989807,
        0.5000721216201782,
        0.4976511001586914,
        0.5349262356758118,
        0.5321810245513916,
        0.521889865398407,
        0.6040803790092468,
        0.5971738696098328,
        0.6116356253623962,
        0.5340478420257568,
        0.5248314142227173,
        0.5162411332130432,
        0.5296908617019653,
        0.5145148038864136,
        0.48725625872612,
        0.5203425884246826,
        0.5149979591369629,
        0.5206071734428406,
        0.5174234509468079,
        0.5370599031448364,
        0.5065783858299255,
        0.47213637828826904,
        0.45470839738845825,
        0.47088170051574707
    ],
    "eval_runtime": 47.5351,
    "eval_samples_per_second": 15.778,
    "eval_steps_per_second": 0.505,
    "eval_rms_calibration_error": 0.09769978776219478
}