GENIES/results/evaluations/cra/llama-7b-code_low_quality/code.json

{
    "model_dir": "models/llama-7b",
    "distribution_id": "code",
    "timestamp": "2023-10-16 02:52:43",
    "eval_score": -0.1990952731689067,
    "eval_brier_score": 0.1990952731689067,
    "eval_average_probability": 0.6739371965107741,
    "eval_accuracy": 0.7066666666666667,
    "eval_probabilities": [
        0.8261599082303105,
        0.9613865305712365,
        0.8456792266250042,
        0.9559445158252088,
        0.9856790620037408,
        0.9802381669299809,
        0.3887173233121327,
        0.9811128174898878,
        0.5970503265293547,
        0.3099091604684607,
        0.23105821785519118,
        0.8870598447589886,
        0.8757296102234315,
        0.927076766943273,
        0.47276428499309747,
        0.5319007524689617,
        0.6218830256611401,
        0.2626430877540834,
        0.9895161064235585,
        0.9926468096804757,
        0.9773357064505221,
        0.4875580187222388,
        0.9909181069650362,
        0.9936608348966685,
        0.6537381037757862,
        0.6220389815982427,
        0.5792518024948528,
        0.5767283554142759,
        0.16235847910319404,
        0.5274560381158662,
        0.875412023399895,
        0.9709890007385622,
        0.9007275905202858,
        0.7095597944415273,
        0.9250585569422473,
        0.7598902269324564,
        0.8382326591418662,
        0.7709891008202625,
        0.77652166873765,
        0.2913628609509686,
        0.8325784630264429,
        0.9660187230685698,
        0.7031554309500523,
        0.9461235817376542,
        0.447270198947196,
        0.7365172516770083,
        0.9773181268386988,
        0.9818177374786206,
        0.9417167168583116,
        0.6456681828340821,
        0.8970905402240562,
        0.8962606501812664,
        0.999240315532188,
        0.9978009887260273,
        0.4040052949374581,
        0.5749270325932683,
        0.5983389585719081,
        0.7135061588329161,
        0.532078324504017,
        0.5119629478359545,
        0.9724217666493853,
        0.9768933188800528,
        0.9782358114147616,
        0.9805968087944023,
        0.6899933278117446,
        0.20393194051282326,
        0.9852410817592173,
        0.7853678505949077,
        0.7297666254239592,
        0.8549827922099511,
        0.9125591607888373,
        0.989204968672235,
        0.24787912056190364,
        0.43961944778545503,
        0.5902198465125953,
        0.9956100426787939,
        0.999591801341156,
        0.9883551590878289,
        0.9823697570340688,
        0.8581892621040192,
        0.9636857856878571,
        0.47077713458572074,
        0.57838618282038,
        0.3425466886392755,
        0.9454609478285816,
        0.9859332863679421,
        0.996127408605775,
        0.9615707082077012,
        0.9801269313235864,
        0.7485296513564433,
        0.8078521129760378,
        0.5578350060007613,
        0.10594480579634244,
        0.0768502985894978,
        0.15274618639280046,
        0.24895374824100183,
        0.6514278195456976,
        0.3570943199097625,
        0.06744475677451543,
        0.7566968854386871,
        0.711548248479666,
        0.22876459748350197,
        0.9416710713518216,
        0.9967830995051937,
        0.8817702064460253,
        0.9025235903002129,
        0.9290418721153959,
        0.7450278615749213,
        0.88971909177716,
        0.9103175270447743,
        0.9414626787228215,
        0.9828877675360342,
        0.8415969993751674,
        0.9351339823369981,
        0.8106914563425627,
        0.7355167375508584,
        0.7668537205035918,
        0.01918055741351563,
        0.025357696323194054,
        0.03109553482545413,
        0.9512122940670349,
        0.9591991795251662,
        0.9845280680134081,
        0.9680397885644868,
        0.9761715583242215,
        0.06710782933010595,
        0.9251273115500961,
        0.8499273776221481,
        0.7195695264520999,
        0.9664103875519524,
        0.97935945984049,
        0.8051636509846677,
        0.697157604212265,
        0.19310394718206503,
        0.41934066278081383,
        0.9824011978687824,
        0.965357203208419,
        0.9273823284223723,
        0.011347318550911879,
        0.9292547067300796,
        0.39399742915666347,
        0.6990076201371255,
        0.636539221867777,
        0.625655032245459,
        0.7038732356168222,
        0.44832350433010626,
        0.911531047840975,
        0.9182154887884415,
        0.7360564769300946,
        0.042494406264258634,
        0.4833951749085734,
        0.2080660271817345,
        0.5861831671617352,
        0.497713696047846,
        0.30223679698787276,
        0.6957761652696851,
        0.8606076292885771,
        0.8576197587148876,
        0.9032727120992919,
        0.5637396279829724,
        0.007790920419398044,
        0.037448937234235875,
        0.8053625672194554,
        0.9602700457697569,
        0.9476643548489522,
        0.9959212376982168,
        0.9632965569737584,
        0.9287091448662588,
        0.8884064926384443,
        0.8560690042221517,
        0.1483646570721186,
        0.9855808490419624,
        0.8916821191296994,
        0.8407364254096642,
        0.8972832778672016,
        0.33469281232861603,
        0.9936848039629556,
        0.1821219934956645,
        0.9991353760503767,
        0.9840277131049112,
        0.18968805665235733,
        0.19968391055993698,
        0.3513471532052282,
        0.6973176978088889,
        0.8242818776735712,
        0.1938171421661573,
        0.09876809119048745,
        0.04660645177832964,
        0.09987868371403723,
        0.11581502942309314,
        0.9589029573401768,
        0.8782195274268914,
        0.9041092287815714,
        0.7696366813045816,
        0.9571987570687192,
        0.9543084828141585,
        0.05910707388968737,
        0.9056604613723579,
        0.24887208718637677,
        0.05972103257711445,
        0.22192673509726274,
        0.9718604983519734,
        0.9573346815632662,
        0.9532149869289266,
        0.2786621790798043,
        0.9420031060541134,
        0.7090604002819765,
        0.2627217888960621,
        0.3289369514890189,
        0.06898571714277935,
        0.789271645883016,
        0.7820500075794989,
        0.911667997907592,
        0.2980071549281083,
        0.44541992707308964,
        0.30000979300108344,
        0.13992506476441846,
        0.36517719825851747,
        0.13242715718903186,
        0.8379440540993681,
        0.6651984751698129,
        0.8865130806843333,
        0.6288644492619275,
        0.9639243457399759,
        0.0926936208022292,
        0.1297545344409832,
        0.5341903719346905,
        0.48151285344411127,
        0.5719113320502678,
        0.9680217451727546,
        0.8785877126255114,
        0.03418556197286094,
        0.3925659368841102,
        0.20094662981707984,
        0.5307952975781011,
        0.750519366435504,
        0.5121692289053428,
        0.8167729912028,
        0.8824591586678694,
        0.22173503007822853,
        0.845552161255935,
        0.8870330517644915,
        0.9491240444055408,
        0.9875989507032176,
        0.37107947800170404,
        0.7020440210625387,
        0.5599053466129813,
        0.8365111678907999,
        0.9801101742969451,
        0.998773419810716,
        0.5428840961486654,
        0.9987736659302094,
        0.751880814722655,
        0.4627657871907995,
        0.6051359381342747,
        0.9246879595217791,
        0.935581397376554,
        0.7982679603248033,
        0.7283673114728288,
        0.8358341824830011,
        0.3484401336931825,
        0.9808661998012829,
        0.47232544088932493,
        0.6268382299226611,
        0.41215895057993407,
        0.3453037975751839,
        0.6967638237501859,
        0.19532873624470737,
        0.15567252377617158,
        0.1733204533994014,
        0.44162911065484295,
        0.9513643817826793,
        0.6345192276526681,
        0.8102623010430967,
        0.45325443172794,
        0.8819790533091751,
        0.9949934350489731,
        0.9962735753958055,
        0.965208978771498,
        0.9390981894909785,
        0.8955953409449458,
        0.9458097044655595,
        0.9943059091563724,
        0.9264875938213528,
        0.18243083599679444,
        0.9980312247165815,
        0.9975187374649124,
        0.9968568794458261,
        0.49278651688448905,
        0.5524765517866245,
        0.35087590589194256,
        0.8555357696585881,
        0.9242619089476773,
        0.7000359283090197,
        0.9251626282499816,
        0.9290597314312321,
        0.9797713447511058,
        0.4441890334935523,
        0.8406955283848204,
        0.22505254610803754,
        0.20322093510749825,
        0.24275015606956396,
        0.7840265665859258,
        0.9163202243495415,
        0.978769755989052,
        0.9138223641188381,
        0.5493627333151792,
        0.4813760501198634,
        0.9437654298573327,
        0.38078427808524073,
        0.2180367843484247,
        0.5716652075007645,
        0.978120403658401,
        0.9141348439160151,
        0.7133083778220976,
        0.3362225583271674,
        0.43649649603061447,
        0.7250167593640092,
        0.929809078624208,
        0.7971308919319552,
        0.37770817177754956,
        0.7959235710148109,
        0.9075671258572465,
        0.6877970296078184,
        0.5755122280550035,
        0.9897170636379979,
        0.23669282977061804,
        0.9179940481465746,
        0.43329493835932953,
        0.8731412758427773,
        0.6189978016883613,
        0.3864945213305316,
        0.5262806105459621,
        0.0889272325948526,
        0.17559624577507385,
        0.24054195508247656,
        0.05990984076450567,
        0.9923111900489684,
        0.01886571975216149,
        0.01762132412508454,
        0.04146304867051234,
        0.014504467199395488,
        0.7250192592544925,
        0.4311347779354858,
        0.7039980966223911,
        0.9845597082148454,
        0.7895544221050583,
        0.7835966121956629,
        0.6194594847914887,
        0.352645395101584,
        0.19365780397106608,
        0.5167773955027415,
        0.7379701094860113,
        0.5712926606766744,
        0.9194309815962051,
        0.5597285163834167,
        0.8731150528564666,
        0.6504718062338086,
        0.9453667292850421,
        0.5650541602870831,
        0.8929776490113202,
        0.5549687922485634,
        0.9248996124129194,
        0.25965558849729586,
        0.925971398371249,
        0.9971227381635335,
        0.9924610346145042,
        0.9667306497673657,
        0.9974537900773788,
        0.8674923357670672,
        0.9569849449708666,
        0.5942479575681734,
        0.7237905061881531,
        0.19337959441475172,
        0.1553504962418526,
        0.09023719645397321,
        0.04597278909554314,
        0.9501299035638924,
        0.5677033578820837,
        0.9303306131257948,
        0.8955280411199671,
        0.008331415486763541,
        0.9805023305237445,
        0.8636734277828687,
        0.3612203295124312,
        0.3299436132484676,
        0.2991050078208438,
        0.978416730055266,
        0.9365197108399468,
        0.9535909248871624,
        0.9607911186753462,
        0.9131442317179577,
        0.9726151242981758,
        0.9921350391098729,
        0.8039466662526251,
        0.4137888596478917,
        0.994876593402628,
        0.909886765765102,
        0.5670756750545547,
        0.7742617915161534,
        0.869407675129604,
        0.6467339187269806,
        0.260457828413196,
        0.508768217785272,
        0.22792185943595603,
        0.9117174671025963,
        0.7721948406058913,
        0.7607394322461483,
        0.9663859029563708,
        0.18271835445486553,
        0.973496397559788,
        0.3937205886302959,
        0.19587089069552643,
        0.31812662332086605,
        0.7718939752754781,
        0.3158214821620108,
        0.4415721787640611,
        0.6484040391526154,
        0.8166769699120411,
        0.7238847080117701,
        0.9854487950533313,
        0.7708642024156037,
        0.9801657429804524,
        0.8376083146137583,
        0.9875790611420958,
        0.2648261983287554,
        0.2578448006845722,
        0.30546556801424984,
        0.24403426256650848,
        0.8419523821300677,
        0.8403741224176838,
        0.9761357268101121,
        0.0363156461778576,
        0.09688622693115112,
        0.0882695469554281,
        0.8083781563044492,
        0.44499872728089007,
        0.9897951042327994,
        0.961146596992504,
        0.4163226086775549,
        0.9832459694279124,
        0.983477622550664,
        0.9887130278388838,
        0.9839277311135031,
        0.7654843079629597,
        0.3873527722070573,
        0.8904834645556298,
        0.7871634179421719,
        0.6544269233133719,
        0.5608272732077855,
        0.7795390348337266,
        0.9926819021896116,
        0.9901392121773245,
        0.679011397084688,
        0.09979956088023495,
        0.6428768291235616,
        0.5576749423826017,
        0.026940753183226977,
        0.7759998776686785,
        0.49369740865703443,
        0.2632963209687734,
        0.4087141118730105,
        0.9498566397568637,
        0.9662896048565837,
        0.9705267968295886,
        0.18700236191558417,
        0.4678256607969856,
        0.40682936888398286,
        0.5386351097303093,
        0.8353622658121842,
        0.9210025581597767,
        0.23971045624725623,
        0.2370208992264059,
        0.997943304917924,
        0.5886844318779203,
        0.16418426487646842,
        0.7069085351348092,
        0.9693295352270909,
        0.9078357309947029,
        0.9209603318520936,
        0.009496259570009546,
        0.002522062779975843,
        0.976960251032024,
        0.7791681682443836,
        0.9746508613545504,
        0.576033305480707,
        0.48864715656903335,
        0.9630787149528075,
        0.9879228723038632,
        0.9793109019473317,
        0.6538355105086604,
        0.7994766383952121,
        0.9774725719395089,
        0.9609617938058141,
        0.8319703796156217,
        0.6055667197743149,
        0.5733478545056476,
        0.9972153489641178,
        0.595083444335381,
        0.9532159644102677,
        0.9387874723727971,
        0.45303282421326396,
        0.7288298010859544,
        0.4672825467498686,
        0.6322498992335392,
        0.008536794237013923,
        0.878905192527331,
        0.38499448170610473,
        0.4792135124277507,
        0.7468613844815453,
        0.9812377547925926,
        0.9683025340858272,
        0.9806134533767101,
        0.8671571298664256,
        0.3760736762597616,
        0.7858116036555924,
        0.9982523056197921,
        0.9925693675978459,
        0.9957961932022648,
        0.3986096233582069,
        0.8491194703091954,
        0.7340992726944909,
        0.2572892960275479,
        0.1677486054655243,
        0.8600015874406864,
        0.5072423367316427,
        0.36709034780183997,
        0.371464236229058,
        0.8861453245066445,
        0.7080611273821411,
        0.7687416840680367,
        0.981646956705275,
        0.9864242616984681,
        0.969501076360403,
        0.5421184947684461,
        0.8434915167624688,
        0.5557464187189036,
        0.04785049870746768,
        0.0275317138581954,
        0.9010180630071742,
        0.9956551276287149,
        0.9612840984760533,
        0.9863142112239495,
        0.7870955768985258,
        0.29414539876325185,
        0.9925099657320761,
        0.8943321791842225,
        0.9700116613923557,
        0.9470477562087934,
        0.7696983560287516,
        0.6085394784063041,
        0.6796268784135621,
        0.372579834940694,
        0.6687044596930984,
        0.19583312942996237,
        0.10426442024113658,
        0.240641004865363,
        0.10611884400886358,
        0.43206753136196363,
        0.4243078930893506,
        0.2544212554056111,
        0.8696500399841304,
        0.8333034463122564,
        0.7588692110192059,
        0.8570920552566528,
        0.3546367628853292,
        0.3937877642306684,
        0.8518104737034131,
        0.9016712580094725,
        0.9156054213414117,
        0.8958876466664467,
        0.18296255070803594,
        0.241802036623187,
        0.8815849272334484,
        0.8289707259312333,
        0.2489715299975257,
        0.9934326780750194,
        0.9938315646028019,
        0.9925373662165015,
        0.9942108787341917,
        0.9827058862331169,
        0.9920393360464366,
        0.14587172332081233,
        0.3301518548239288,
        0.5801273148936232,
        0.9734648258546961,
        0.9733416679559512,
        0.9734971357951379,
        0.3029096555491996,
        0.9943351787909721,
        0.03449580606829489,
        0.32321376554428194,
        0.2499415077113586,
        0.35607446868019665,
        0.9889659907038972,
        0.8199055726375338,
        0.9711712077995187,
        0.994340272056146,
        0.9684668982480209,
        0.21023417843949094,
        0.07818562713405212,
        0.9625738108134914,
        0.6594139045319494,
        0.9022275827407878,
        0.7695389960858929,
        0.7688651852517147,
        0.8729786155787117,
        0.514283661781338,
        0.6353683830327825,
        0.9973837332096162,
        0.9791323232123207,
        0.9379452531534263,
        0.9686353658376441,
        0.5334009030851201,
        0.43228698865674864,
        0.9727215254760786,
        0.9391857916375398,
        0.9530012354665834,
        0.1408114159976877,
        0.885804038231567,
        0.9894696247087365,
        0.2883037649385759,
        0.4220660700033644,
        0.23274992888626359,
        0.938909292225737,
        0.8921767891547857,
        0.8123376212601313,
        0.13358975132983206,
        0.05910008548628569,
        0.753346558259534,
        0.934571220417344,
        0.8029844666293502,
        0.7458890677877879,
        0.8720574515832811,
        0.8600072171885633,
        0.9385598510700237,
        0.929013112242562,
        0.7212301341812735,
        0.21841034816193017,
        0.1252035097215829,
        0.19087385015007527,
        0.28478891923483673,
        0.2797122808622296,
        0.4179931220653514,
        0.9416129348835274,
        0.9984840562989752,
        0.9010087446876591,
        0.988882162764888,
        0.992776708215628,
        0.9414253014299179,
        0.6415687292921404,
        0.9861905041443062,
        0.9480375653099151,
        0.9555872563903032,
        0.9818073524551088,
        0.989553800185807,
        0.9830906055039494,
        0.6840275728255781,
        0.2765248127241509,
        0.9719830010905789,
        0.9055646590829395,
        0.9204876316201217,
        0.7851288078851572,
        0.11628593611775763,
        0.9110414981163869,
        0.9928274364553935,
        0.9263180175386581,
        0.9553546507788078,
        0.9673924215277765,
        0.9753606805351003,
        0.6987779370274958,
        0.8748148274885872,
        0.6139104421961685,
        0.05515282576696893,
        0.4102532581156028,
        0.5015421448451772,
        0.8701735770266114,
        0.2555657819692758,
        0.4597925309269672,
        0.1377446657167162,
        0.9089692602656557,
        0.7815562040411516,
        0.9766982412557031,
        0.9845157969681777,
        0.9699110131928477,
        0.9955221444792641,
        0.5934079274833085,
        0.9873176584909962,
        0.9598578147366562,
        0.9745857928874311,
        0.8828904397168152,
        0.9195971987669994,
        0.7798619603436755,
        0.9482762858780373,
        0.9765680552537767,
        0.1945927707418501,
        0.1512048741113301,
        0.1028811339968105,
        0.19612487562622505,
        0.9869332909374455,
        0.9821416217436497,
        0.98701826829483,
        0.48951837520223673,
        0.44772108103892705,
        0.5574907438477025,
        0.9857777543792743,
        0.6319698968010806,
        0.4977294337709304,
        0.9458405152404867,
        0.596116613985449,
        0.7722542132568895,
        0.8013354397747623,
        0.8501624417058871,
        0.1338547600819352,
        0.5303542245923771,
        0.42591348828208575,
        0.4774246635651762,
        0.6845992534293384,
        0.3155686002066973,
        0.5772561973380921,
        0.9414830963722602,
        0.8428227293524768,
        0.5303763969102657,
        0.6865823678373628,
        0.39845515541363075,
        0.8489344149211113,
        0.9113390192745376,
        0.9532444331770598,
        0.5673884383857527,
        0.8566076757580409,
        0.8952395719531413,
        0.15573457064971974,
        0.9493386411852349,
        0.9778795132224738,
        0.9586611130577353,
        0.18523930487678694,
        0.8656520867008911,
        0.7372933765480418,
        0.7600261958752772,
        0.9746086234012937,
        0.9633191534208317,
        0.7409487140824431,
        0.751902872928591,
        0.885312825271159,
        0.9640443824914028,
        0.6725951177721265,
        0.9238139458209225,
        0.8990548832022701,
        0.6315429258105376,
        0.8048421980805764
    ],
    "eval_rms_calibration_error": 0.3004103783351849
}