GENIES/results/evaluations/zero_shot/openllama-3b-code_easy/code_easy.json

{
    "model_dir": "models/openllama-3b",
    "distribution_id": "code_easy",
    "timestamp": "2023-10-20 01:21:46",
    "eval_loss": 0.6655530333518982,
    "eval_score": -0.2363140881061554,
    "eval_brier_score": 0.2363140881061554,
    "eval_average_probability": 0.5139952898025513,
    "eval_accuracy": 0.9106666666666666,
    "eval_probabilities": [
        0.5168652534484863,
        0.5133365392684937,
        0.5155653357505798,
        0.5151981115341187,
        0.512463390827179,
        0.505691409111023,
        0.5030257701873779,
        0.5155941247940063,
        0.5105094909667969,
        0.5234545469284058,
        0.5111677646636963,
        0.5146099925041199,
        0.5062571167945862,
        0.5314632058143616,
        0.520572304725647,
        0.500825822353363,
        0.5176051259040833,
        0.5097552537918091,
        0.5205347537994385,
        0.5144368410110474,
        0.5020245313644409,
        0.5052919983863831,
        0.520614743232727,
        0.5055747628211975,
        0.5164650678634644,
        0.5128494501113892,
        0.5145873427391052,
        0.510335385799408,
        0.5136173367500305,
        0.5185327529907227,
        0.5066341757774353,
        0.506339967250824,
        0.5109091401100159,
        0.5385568141937256,
        0.5234999060630798,
        0.5260714292526245,
        0.47542423009872437,
        0.48395073413848877,
        0.5000907182693481,
        0.5199007391929626,
        0.513775110244751,
        0.5105067491531372,
        0.5152577757835388,
        0.5225973725318909,
        0.5188552141189575,
        0.5126706957817078,
        0.49589803814888,
        0.5006158947944641,
        0.5126441121101379,
        0.516075611114502,
        0.5137516856193542,
        0.49833330512046814,
        0.5131268501281738,
        0.5110664963722229,
        0.5208512544631958,
        0.5303130745887756,
        0.532974898815155,
        0.506251871585846,
        0.5017001032829285,
        0.5035861730575562,
        0.5095774531364441,
        0.5116519331932068,
        0.5132797360420227,
        0.5353983640670776,
        0.5170761942863464,
        0.5326884388923645,
        0.5019939541816711,
        0.5185239315032959,
        0.5155539512634277,
        0.5045585632324219,
        0.4929264783859253,
        0.48992228507995605,
        0.5152061581611633,
        0.5194900631904602,
        0.5144708752632141,
        0.5099828839302063,
        0.5072536468505859,
        0.5074618458747864,
        0.5128169655799866,
        0.5022762417793274,
        0.5167270302772522,
        0.5078072547912598,
        0.5127918720245361,
        0.5112931132316589,
        0.5073410272598267,
        0.5099706053733826,
        0.5110816359519958,
        0.5059965252876282,
        0.5118052959442139,
        0.5137434601783752,
        0.5315967202186584,
        0.5059131383895874,
        0.5012301802635193,
        0.5305178165435791,
        0.5098488926887512,
        0.5242260098457336,
        0.5421270132064819,
        0.5130367279052734,
        0.5262084603309631,
        0.5203452706336975,
        0.5038136839866638,
        0.5173119306564331,
        0.521480917930603,
        0.5200851559638977,
        0.5147807002067566,
        0.5082964897155762,
        0.5238143801689148,
        0.52323979139328,
        0.51823490858078,
        0.5138131976127625,
        0.5314403772354126,
        0.5039888024330139,
        0.5222987532615662,
        0.505521833896637,
        0.513688325881958,
        0.49647653102874756,
        0.5119937658309937,
        0.49921560287475586,
        0.5320820212364197,
        0.5098838210105896,
        0.5100706815719604,
        0.5136504173278809,
        0.507882297039032,
        0.5354253649711609,
        0.5133380889892578,
        0.5159346461296082,
        0.5136502385139465,
        0.516739547252655,
        0.5149708390235901,
        0.5129197835922241,
        0.5166390538215637,
        0.5168036222457886,
        0.5300845503807068,
        0.5148807764053345,
        0.515875518321991,
        0.49943140149116516,
        0.5041342377662659,
        0.5040654540061951,
        0.5059065818786621,
        0.5059065818786621,
        0.5159474611282349,
        0.528287947177887,
        0.5295425057411194,
        0.5487002730369568,
        0.5230045914649963,
        0.509340226650238,
        0.5126523375511169,
        0.5106199979782104,
        0.53548264503479,
        0.5254026651382446,
        0.5057913661003113,
        0.49761682748794556,
        0.5129525661468506,
        0.5143794417381287,
        0.5075833201408386,
        0.4945555627346039,
        0.5298880338668823,
        0.5155590772628784,
        0.5131466388702393,
        0.5105268359184265,
        0.5123952031135559,
        0.5089564919471741,
        0.5026221871376038,
        0.5052317976951599,
        0.49915045499801636,
        0.5088129639625549,
        0.5051700472831726,
        0.5020691752433777,
        0.5195232629776001,
        0.49858972430229187,
        0.5257859826087952,
        0.5074553489685059,
        0.5151191353797913,
        0.516984224319458,
        0.5084666609764099,
        0.5055198073387146,
        0.5074424743652344,
        0.5097139477729797,
        0.5142449736595154,
        0.5136273503303528,
        0.5068559646606445,
        0.5184626579284668,
        0.5254793763160706,
        0.5126082897186279,
        0.5186235904693604,
        0.51517653465271,
        0.512076199054718,
        0.5454112887382507,
        0.506937563419342,
        0.5034267902374268,
        0.5143859386444092,
        0.5166438817977905,
        0.5260981917381287,
        0.5091431736946106,
        0.5214728713035583,
        0.5091848969459534,
        0.5066021680831909,
        0.4988032877445221,
        0.5172826051712036,
        0.5166453123092651,
        0.5134665369987488,
        0.5163140892982483,
        0.494120717048645,
        0.497835636138916,
        0.5135676264762878,
        0.515080451965332,
        0.5159626603126526,
        0.5154498219490051,
        0.5246013402938843,
        0.5191601514816284,
        0.5129184722900391,
        0.5050089359283447,
        0.5024225115776062,
        0.5209040641784668,
        0.5093562006950378,
        0.5255982279777527,
        0.5135701298713684,
        0.5082201957702637,
        0.5056025385856628,
        0.50942063331604,
        0.5147661566734314,
        0.5086358785629272,
        0.5061967372894287,
        0.510221004486084,
        0.5019175410270691,
        0.5230907201766968,
        0.5155262351036072,
        0.5254929065704346,
        0.5097371935844421,
        0.5066647529602051,
        0.5143868923187256,
        0.5171395540237427,
        0.5268906950950623,
        0.5200892090797424,
        0.5325561165809631,
        0.5121995210647583,
        0.5207148790359497,
        0.5201473236083984,
        0.5083786845207214,
        0.5129202604293823,
        0.5066179037094116,
        0.49678075313568115,
        0.5163722038269043,
        0.5182770490646362,
        0.5037958025932312,
        0.5322604179382324,
        0.5094074606895447,
        0.5403903722763062,
        0.5443368554115295,
        0.5093566179275513,
        0.5060123205184937,
        0.49956372380256653,
        0.5211423635482788,
        0.5055837035179138,
        0.5118573904037476,
        0.5323495864868164,
        0.4972348213195801,
        0.5127651691436768,
        0.5071895718574524,
        0.5096885561943054,
        0.4970099925994873,
        0.540378987789154,
        0.5375296473503113,
        0.531768798828125,
        0.5203022956848145,
        0.5085334777832031,
        0.5009279251098633,
        0.5192045569419861,
        0.5169415473937988,
        0.5129841566085815,
        0.5263469219207764,
        0.5247766375541687,
        0.5053591132164001,
        0.5058020949363708,
        0.5122199058532715,
        0.5079251527786255,
        0.51786869764328,
        0.5171554684638977,
        0.5073345303535461,
        0.505974531173706,
        0.5153393745422363,
        0.512198269367218,
        0.5067218542098999,
        0.5018500685691833,
        0.5112144947052002,
        0.5102890729904175,
        0.5082063674926758,
        0.5138794183731079,
        0.5382654070854187,
        0.5310537815093994,
        0.5403749346733093,
        0.5113850235939026,
        0.520185649394989,
        0.49896836280822754,
        0.5187456607818604,
        0.5242207646369934,
        0.5301237106323242,
        0.5060312747955322,
        0.5076606869697571,
        0.5061904788017273,
        0.5053616762161255,
        0.5182081460952759,
        0.521888792514801,
        0.5260858535766602,
        0.5267900824546814,
        0.5282099843025208,
        0.5164710879325867,
        0.5329036712646484,
        0.5181478261947632,
        0.5220378637313843,
        0.5179284811019897,
        0.5252305865287781,
        0.5280243158340454,
        0.5096965432167053,
        0.511328399181366,
        0.5326597690582275,
        0.5195102095603943,
        0.5224713683128357,
        0.5248386263847351,
        0.5229876637458801,
        0.5293217301368713,
        0.5197476744651794,
        0.5180314183235168,
        0.5263015627861023,
        0.5141007900238037,
        0.510242760181427,
        0.5161868929862976,
        0.5068742036819458,
        0.5029698014259338,
        0.5133624076843262,
        0.5243682861328125,
        0.5116703510284424,
        0.5202710628509521,
        0.5258780121803284,
        0.5392179489135742,
        0.5283312797546387,
        0.5094785094261169,
        0.5097296833992004,
        0.5012132525444031,
        0.5181065797805786,
        0.5251685976982117,
        0.5478435754776001,
        0.50618577003479,
        0.5145664811134338,
        0.5033037662506104,
        0.5105252265930176,
        0.5092099905014038,
        0.506291925907135,
        0.49711668491363525,
        0.5068477988243103,
        0.4934193789958954,
        0.5081856846809387,
        0.5065469145774841,
        0.5109930038452148,
        0.5139183402061462,
        0.5122271180152893,
        0.5248093605041504,
        0.5021740198135376,
        0.4985562562942505,
        0.508455216884613,
        0.5034996271133423,
        0.5155467987060547,
        0.5131726861000061,
        0.5013397932052612,
        0.5109596252441406,
        0.517383337020874,
        0.5089787840843201,
        0.5186073184013367,
        0.5123409628868103,
        0.5295681953430176,
        0.5040345788002014,
        0.5062264800071716,
        0.5280243158340454,
        0.5096965432167053,
        0.511328399181366,
        0.5224023461341858,
        0.5334970355033875,
        0.5119549632072449,
        0.5121527314186096,
        0.5060974359512329,
        0.5269674062728882,
        0.5137227177619934,
        0.5341827273368835,
        0.5172703266143799,
        0.5171635746955872,
        0.5111781358718872,
        0.5238491296768188,
        0.5268793702125549,
        0.5185365676879883,
        0.5212042927742004,
        0.5186237096786499,
        0.5192621350288391,
        0.5203912258148193,
        0.5211423635482788,
        0.5055837035179138,
        0.5118573904037476,
        0.502635657787323,
        0.498963326215744,
        0.4981263279914856,
        0.5028484463691711,
        0.5177129507064819,
        0.5220735669136047,
        0.5161548256874084,
        0.5227852463722229,
        0.5137733221054077,
        0.5221861600875854,
        0.5126166939735413,
        0.5232270956039429,
        0.4987533390522003,
        0.5260781645774841,
        0.5191270709037781,
        0.5210315585136414,
        0.5309420824050903,
        0.5408578515052795,
        0.5168924927711487,
        0.48262372612953186,
        0.517401933670044,
        0.510762631893158,
        0.5125330686569214,
        0.5179653763771057,
        0.5536504983901978,
        0.5227938294410706,
        0.5091719627380371,
        0.5020998120307922,
        0.5076441168785095,
        0.5173007249832153,
        0.5237111449241638,
        0.5124632120132446,
        0.514725923538208,
        0.5102843642234802,
        0.5331273078918457,
        0.5214481353759766,
        0.4866237938404083,
        0.49429893493652344,
        0.501456618309021,
        0.5196465253829956,
        0.5162537097930908,
        0.5134389996528625,
        0.5107799768447876,
        0.5201864838600159,
        0.5130090117454529,
        0.5208926796913147,
        0.5320008397102356,
        0.5198190212249756,
        0.5071419477462769,
        0.5078704953193665,
        0.5094708800315857,
        0.5298539400100708,
        0.5343999266624451,
        0.5240078568458557,
        0.5214256048202515,
        0.5178483724594116,
        0.5155824422836304,
        0.5153807997703552,
        0.5175819396972656,
        0.4998304545879364,
        0.5180754065513611,
        0.513190507888794,
        0.504591166973114,
        0.5022510290145874,
        0.5060615539550781,
        0.5190649032592773,
        0.5114405751228333,
        0.5241142511367798,
        0.5133840441703796,
        0.52033531665802,
        0.5323126912117004,
        0.5126071572303772,
        0.5144022107124329,
        0.5202900767326355,
        0.5214312672615051,
        0.5034793615341187,
        0.5103001594543457,
        0.5157377123832703,
        0.5323289632797241,
        0.520614743232727,
        0.5161120891571045,
        0.5115314722061157,
        0.5170944333076477,
        0.5170946717262268,
        0.5170031189918518,
        0.5147608518600464,
        0.5225780010223389,
        0.4971117377281189,
        0.5102503895759583,
        0.5094176530838013,
        0.5029498338699341,
        0.493847131729126,
        0.49959081411361694,
        0.510095477104187,
        0.5149405598640442,
        0.5036489367485046,
        0.48623988032341003,
        0.5037034749984741,
        0.4943508803844452,
        0.5210509300231934,
        0.5053063035011292,
        0.5126445889472961,
        0.5177347660064697,
        0.5047171115875244,
        0.5101521015167236,
        0.5109193325042725,
        0.5155804753303528,
        0.5259255170822144,
        0.497514009475708,
        0.50489741563797,
        0.4960957467556,
        0.5243540406227112,
        0.5216076970100403,
        0.5160923004150391,
        0.4914449453353882,
        0.488570898771286,
        0.49754735827445984,
        0.5264379978179932,
        0.4988703429698944,
        0.5246639251708984,
        0.49550682306289673,
        0.5080474019050598,
        0.50968998670578,
        0.5152894258499146,
        0.5114553570747375,
        0.5308259129524231,
        0.5158771872520447,
        0.5104707479476929,
        0.5034286379814148,
        0.52997887134552,
        0.5240810513496399,
        0.49129122495651245,
        0.5216582417488098,
        0.5033354163169861,
        0.5234576463699341,
        0.5186834335327148,
        0.5076455473899841,
        0.5068708658218384,
        0.5134397745132446,
        0.506188690662384,
        0.49746236205101013,
        0.5146307945251465,
        0.509185254573822,
        0.5207423567771912,
        0.5183326601982117,
        0.5166122317314148,
        0.5189650654792786,
        0.5132843255996704,
        0.5278602838516235,
        0.5152167677879333,
        0.5398840308189392,
        0.512366771697998,
        0.510562002658844,
        0.5164754390716553,
        0.5195456743240356,
        0.5107898712158203,
        0.509115993976593,
        0.5152239203453064,
        0.5087786912918091,
        0.5083650946617126,
        0.5184675455093384,
        0.5174605250358582,
        0.5269154906272888,
        0.5167929530143738,
        0.5142407417297363,
        0.5276424884796143,
        0.5250715017318726,
        0.516278862953186,
        0.49858972430229187,
        0.5142964124679565,
        0.5290735363960266,
        0.508661150932312,
        0.5070846080780029,
        0.5136452913284302,
        0.5145067572593689,
        0.5200774073600769,
        0.5128711462020874,
        0.5120267868041992,
        0.522773802280426,
        0.5085974335670471,
        0.5055769085884094,
        0.5245469212532043,
        0.5044512748718262,
        0.5196340680122375,
        0.5168160796165466,
        0.5124096870422363,
        0.5025730133056641,
        0.5087857842445374,
        0.5065606832504272,
        0.513995349407196,
        0.5269889235496521,
        0.5194478034973145,
        0.4992101490497589,
        0.522911787033081,
        0.5164030194282532,
        0.5016297698020935,
        0.517503023147583,
        0.5238389372825623,
        0.5112683773040771,
        0.521609365940094,
        0.5163987278938293,
        0.5121042728424072,
        0.5106320977210999,
        0.5084097981452942,
        0.5098064541816711,
        0.5289312601089478,
        0.5202357769012451,
        0.5168652534484863,
        0.49865248799324036,
        0.5097970366477966,
        0.5057938098907471,
        0.5120413899421692,
        0.4978764057159424,
        0.5182434916496277,
        0.5158467888832092,
        0.506808876991272,
        0.5146021842956543,
        0.512444019317627,
        0.5213906168937683,
        0.49921944737434387,
        0.5151411890983582,
        0.4939548671245575,
        0.504797637462616,
        0.4963148236274719,
        0.5164219737052917,
        0.5149368643760681,
        0.5188109278678894,
        0.5600021481513977,
        0.52180016040802,
        0.5060176849365234,
        0.5171722173690796,
        0.5137356519699097,
        0.5032011270523071,
        0.519280731678009,
        0.5076045989990234,
        0.5105989575386047,
        0.49712371826171875,
        0.5231624245643616,
        0.5216660499572754,
        0.5002707242965698,
        0.507478654384613,
        0.5270940065383911,
        0.5097835063934326,
        0.5124707221984863,
        0.5119192600250244,
        0.514817476272583,
        0.5256346464157104,
        0.528793215751648,
        0.5170072317123413,
        0.5190739631652832,
        0.5149672031402588,
        0.5088659524917603,
        0.5148844718933105,
        0.5434024930000305,
        0.5204606652259827,
        0.5314525961875916,
        0.5099261403083801,
        0.5402162671089172,
        0.5206576585769653,
        0.5219948887825012,
        0.5143062472343445,
        0.5045440196990967,
        0.501183271408081,
        0.4930412769317627,
        0.5223007798194885,
        0.5119546055793762,
        0.5213318467140198,
        0.47925126552581787,
        0.48674196004867554,
        0.49304574728012085,
        0.5101749300956726,
        0.514652669429779,
        0.5231984257698059,
        0.4979628920555115,
        0.4982250928878784,
        0.5010420083999634,
        0.4908541440963745,
        0.5164387822151184,
        0.5080543160438538,
        0.5184482932090759,
        0.49743586778640747,
        0.5134158134460449,
        0.5093085169792175,
        0.5193994045257568,
        0.5227612257003784,
        0.5278310179710388,
        0.5218878984451294,
        0.5206042528152466,
        0.515024721622467,
        0.5156537294387817,
        0.5176782608032227,
        0.5158219337463379,
        0.5063881874084473,
        0.5004536509513855,
        0.5194088816642761,
        0.5306379199028015,
        0.5153797268867493,
        0.5129148364067078,
        0.5281099081039429,
        0.5273489952087402,
        0.5343055129051208,
        0.5243300795555115,
        0.5342591404914856,
        0.4993523359298706,
        0.5001094937324524,
        0.49681684374809265,
        0.5183809995651245,
        0.5155224204063416,
        0.5099737048149109,
        0.5145843625068665,
        0.5020134449005127,
        0.5413309335708618,
        0.512254536151886,
        0.5002139210700989,
        0.5075558423995972,
        0.513963520526886,
        0.5198788642883301,
        0.5132453441619873,
        0.5236412882804871,
        0.513387143611908,
        0.5288215279579163,
        0.5075415372848511,
        0.5135637521743774,
        0.505742609500885,
        0.5042794346809387,
        0.5052470564842224,
        0.506701648235321,
        0.4985116720199585,
        0.4989016652107239,
        0.5204498171806335,
        0.5144628286361694,
        0.5166880488395691,
        0.5041576623916626,
        0.515058159828186,
        0.519608199596405,
        0.4977074861526489,
        0.5075705051422119,
        0.4912630319595337,
        0.5048036575317383,
        0.5288702845573425,
        0.5408480763435364,
        0.510444700717926,
        0.5029346942901611,
        0.5030338764190674,
        0.5101965665817261,
        0.5155835747718811,
        0.5249025821685791,
        0.5071493983268738,
        0.518950879573822,
        0.5201787948608398,
        0.5192499160766602,
        0.5168640613555908,
        0.5095548033714294,
        0.521274209022522
    ],
    "eval_runtime": 43.6329,
    "eval_samples_per_second": 17.189,
    "eval_steps_per_second": 0.55
}