GENIES/results/evaluations/mms/llama-13b-code/code.json

{
    "model_dir": "models/llama-13b",
    "distribution_id": "code",
    "timestamp": "2023-10-13 23:46:10",
    "eval_score": -0.09358970951839322,
    "eval_brier_score": 0.09358970951839322,
    "eval_average_probability": 0.8301823698711562,
    "eval_accuracy": 0.8693333333333333,
    "eval_probabilities": [
        0.8112808760125222,
        0.6597419970345558,
        0.8396240714188343,
        0.9994593051905339,
        0.8995442558145963,
        0.9996377040304124,
        0.9999284593289105,
        0.9932903006093015,
        0.9977978802815055,
        0.9944107334223516,
        0.9998587449972097,
        0.9996227235971284,
        0.9836807080452464,
        0.9912140461724868,
        0.9869073904905978,
        0.9828619588169444,
        0.2523671430785528,
        0.9933120889401434,
        0.920924129644211,
        0.637676334638906,
        0.3688570271696584,
        0.9575400877241839,
        0.9995003784989748,
        0.9994596683237144,
        0.16917279328677123,
        0.1052933916062655,
        0.12364048865445029,
        0.931199773599669,
        0.7563514120866255,
        0.9613035491916831,
        0.025018844229716408,
        0.9044314651237274,
        0.6775308112583052,
        0.0923771545371268,
        0.07408019946640533,
        0.08010974130099949,
        0.9778925836542949,
        0.9997550855362144,
        0.9963928042521462,
        0.9169388706158579,
        0.8519340250917988,
        0.6773106715135381,
        0.9994539943614531,
        0.9254589650051044,
        0.9988309232850414,
        0.9632712886789807,
        0.9859578765368737,
        0.9881047144770129,
        0.8596729439999686,
        0.4951080735121109,
        0.9929212194005944,
        0.9912386981251695,
        0.9994852739540661,
        0.9954746802362195,
        0.9856434452071121,
        0.9357119412372015,
        0.9467094605399061,
        0.9966945650391498,
        0.6665103738880531,
        0.9931588716182099,
        0.9716391324118947,
        0.9703622640552041,
        0.9656792393746778,
        0.5271323723990549,
        0.9913846276501022,
        0.9892751542155939,
        0.9737460319065143,
        0.3011772894103968,
        0.4479256291761263,
        0.7249154858864599,
        0.2162061015761041,
        0.9319732208738646,
        0.9201915595905441,
        0.974828329827474,
        0.6486270740885289,
        0.9984368042311621,
        0.9385716016733687,
        0.9976099386759623,
        0.9993651310493376,
        0.9971578134125066,
        0.9996166453487405,
        0.7654566734157155,
        0.9295736115805466,
        0.788622895293099,
        0.9320047265514534,
        0.9948857884090114,
        0.9746951191083727,
        0.8076069553801595,
        0.627717345744639,
        0.5404415585929763,
        0.915750198438163,
        0.9162986182052119,
        0.9465015992210485,
        0.9222147308911907,
        0.9943163434048434,
        0.8430493485412238,
        0.9453105369184402,
        0.49039886467839094,
        0.5187145187839453,
        0.9743645530575464,
        0.8786681497013874,
        0.0037337108245123485,
        0.9273825608118295,
        0.9271603555945538,
        0.9220425436296793,
        0.976731614915871,
        0.9880007024265484,
        0.9860359504553698,
        0.9840535389123842,
        0.9832830047364738,
        0.9796695511413616,
        0.9918410354155475,
        0.9663612480531785,
        0.9155836907683381,
        0.8441343054947903,
        0.9166338907023862,
        0.5111035027839812,
        0.9996631283527387,
        0.7647080398350973,
        0.793097575034756,
        0.8798717832932992,
        0.8915299518952671,
        0.8787522669625738,
        0.8418777602827558,
        0.998498859482926,
        0.6961431434752636,
        0.9829396768780126,
        0.9394974675293473,
        0.9610500173712311,
        0.9995489147597139,
        0.9980689771277218,
        0.9175906277079315,
        0.9168410197877738,
        0.9980661267729385,
        0.94363490644809,
        0.9986470585090884,
        0.9994137128994254,
        0.8064436214400933,
        0.26134082570911443,
        0.9753761819640191,
        0.32350555700517064,
        0.931764283110714,
        0.3459627449294397,
        0.37019922745850603,
        0.04567698312431456,
        0.2503539540243229,
        0.7331263521919658,
        0.999258543085602,
        0.966220054804106,
        0.9933860090327473,
        0.9966379834811818,
        0.9943299962383005,
        0.9891918124706222,
        0.026774178454426423,
        0.48447409386210816,
        0.9336303265396934,
        0.9563151304896682,
        0.9702649530329833,
        0.9824714425075858,
        0.9299366075308059,
        0.6081227755596497,
        0.7667056639262911,
        0.9202134390799634,
        0.939378017949298,
        0.9243899462274251,
        0.9998593878634826,
        0.999762428304303,
        0.9998099089912154,
        0.5059948618486704,
        0.22216941204467347,
        0.7134596245443262,
        0.9924375593675591,
        0.9617545598561378,
        0.999253223140803,
        0.994458653297596,
        0.9984522752702707,
        0.9845591966144298,
        0.6884850805616567,
        0.9955707972086375,
        0.8894946645248872,
        0.5327069852257766,
        0.5208320887971083,
        0.710038862536272,
        0.524033948153735,
        0.9911295784240092,
        0.9849483368694311,
        0.8659242138617671,
        0.672302871313999,
        0.7771757610542694,
        0.935864127928238,
        0.990050580510467,
        0.9915981716128938,
        0.934975233057681,
        0.9533954689483011,
        0.9806338941496737,
        0.7409008903816342,
        0.7538364301552026,
        0.871983287849349,
        0.9992025263605706,
        0.994977053366395,
        0.9997552162176494,
        0.8831394212209204,
        0.7738093799580248,
        0.7368941480859408,
        0.8814806378218162,
        0.9973074245122844,
        0.9960982304320312,
        0.26555683032185007,
        0.988923149738301,
        0.9979810519502534,
        0.8817584327009993,
        0.8934041088589342,
        0.9969872432380802,
        0.5967172190465011,
        0.2132459492333267,
        0.5218327898758024,
        0.5204130150560824,
        0.9539055033886568,
        0.9989227016178039,
        0.6292590043112338,
        0.6787583470836999,
        0.7310840413512102,
        0.08789843146540778,
        0.9447664454407929,
        0.995025719631842,
        0.9785672402474039,
        0.7462818767808677,
        0.977302998677244,
        0.9985686537658494,
        0.9533259921116237,
        0.996259335694565,
        0.9998053266334742,
        0.9989458788705107,
        0.9997702308552106,
        0.9993936723501221,
        0.9997154090903413,
        0.9980029661459682,
        0.8622430222294145,
        0.9697367738527595,
        0.9473764920917035,
        0.08540924279369316,
        0.0941646413801548,
        0.11972276023861561,
        0.9972456030294001,
        0.9983062993565708,
        0.9987856040967642,
        0.08822652596889498,
        0.6015774299935486,
        0.5538236928938948,
        0.8299648675802589,
        0.5510763702784273,
        0.9994755533861345,
        0.9722229817816969,
        0.9861993764577612,
        0.9644491049962507,
        0.9835720768608583,
        0.9958915951464831,
        0.5910973929667811,
        0.9917561711982688,
        0.3951183865363626,
        0.42438555245972437,
        0.9987213999148116,
        0.9586267392048163,
        0.9740248818117291,
        0.9997806316117371,
        0.9974248883531432,
        0.998835001819882,
        0.9528702938422702,
        0.9826078003643247,
        0.9574176942075919,
        0.9964123992481264,
        0.9994817655598351,
        0.9915787518696708,
        0.9865603207886179,
        0.9784156839426089,
        0.9885200174007245,
        0.9998333811903738,
        0.9986940818775684,
        0.9895802840499367,
        0.3908962340336658,
        0.8007250855225276,
        0.5148902165649247,
        0.9998485569962788,
        0.9374370027163959,
        0.9993016307111919,
        0.9949577196418256,
        0.9311684905688806,
        0.9213742662737705,
        0.9992873386129684,
        0.9939985040260051,
        0.9997781346128672,
        0.9683462621240616,
        0.9965702856199046,
        0.8908012762381126,
        0.9965366475342871,
        0.8064014686738982,
        0.08390847732793005,
        0.9585417510895071,
        0.7924729706967424,
        0.4866480374138229,
        0.999431496689151,
        0.9992447169864009,
        0.9995911337763793,
        0.996606243066613,
        0.9997283169960699,
        0.9994010728399826,
        0.8622431275633634,
        0.9606087938042914,
        0.9984120038010647,
        0.9872420339176999,
        0.9049913817039154,
        0.905003237513931,
        0.5244030969189338,
        0.2706491462734795,
        0.23046620068634635,
        0.4252023434337676,
        0.9963665383558287,
        0.4669587025552979,
        0.9977200354631194,
        0.998967496606767,
        0.9957932279381073,
        0.7523553592953355,
        0.5958163272697733,
        0.28109319524682175,
        0.9997600264583975,
        0.9994704703577783,
        0.9984824935199464,
        0.9931136553525121,
        0.9237421887725259,
        0.984070173670419,
        0.9994112507534701,
        0.8695579360052952,
        0.9279949325552421,
        0.9008446769087614,
        0.8996962096615009,
        0.9563670968404552,
        0.9953652892087463,
        0.9982883567030316,
        0.9550381032912749,
        0.6818831268672914,
        0.5628597788223704,
        0.6439569381397061,
        0.9973667246127007,
        0.9990299434148668,
        0.9987941866705194,
        0.9994828344683262,
        0.9997080435350786,
        0.9994404565876337,
        0.9993344313059551,
        0.9988330333293368,
        0.7180666496357957,
        0.8906498792692505,
        0.6410151770692283,
        0.8308558725729498,
        0.7705099689181292,
        0.79255222457045,
        0.9586622481264431,
        0.8669604700687356,
        0.995514137297194,
        0.995659555864675,
        0.9991894504745356,
        0.9840805894712166,
        0.9948957324349312,
        0.9998548121599574,
        0.9999189626297178,
        0.9997140043442378,
        0.884890439869629,
        0.9937736134768539,
        0.9574442414762879,
        0.9656017539072104,
        0.4366230562015441,
        0.5763168236086105,
        0.9997297190419858,
        0.9969750885489984,
        0.9977416714007685,
        0.958067154958941,
        0.9974903975905858,
        0.9943804852966218,
        0.9980222446096918,
        0.9983885085696023,
        0.9939555251490017,
        0.9640674239360612,
        0.9902888538128378,
        0.9986675764640311,
        0.7981288946631091,
        0.8659816022276178,
        0.613329624758788,
        0.9864763866682688,
        0.993871340770191,
        0.9957435771319036,
        0.910194203943588,
        0.9708968858065992,
        0.6856227364996489,
        0.19575731686739173,
        0.9030370334469288,
        0.2125121954061411,
        0.838276847719626,
        0.9733758073956602,
        0.940683719037711,
        0.9261463376077868,
        0.9946589904677338,
        0.8786663063758338,
        0.9964011613034089,
        0.9092816388412637,
        0.9854014467903872,
        0.2723077111703063,
        0.26389096448267085,
        0.25397309687414704,
        0.9995397446351524,
        0.9963663649878014,
        0.9998923678538186,
        0.4269007140043367,
        0.2116859315676647,
        0.9851745150136196,
        0.9996087572574266,
        0.9988799668380802,
        0.6606359511750075,
        0.8248730465017681,
        0.9944969317030421,
        0.9975262960940241,
        0.6547765085218209,
        0.943101731377093,
        0.4242453992930971,
        0.38160701675555636,
        0.8066741289339416,
        0.9957288635505462,
        0.9851172868062421,
        0.9399562782097184,
        0.042047213080393275,
        0.9919907713360733,
        0.9952729015527311,
        0.9872783726797745,
        0.20346431885890937,
        0.23894312229658649,
        0.36756344007578406,
        0.965132704195253,
        0.9885571778990457,
        0.9764750001921426,
        0.9992202141124147,
        0.2513955065481323,
        0.9961079429199211,
        0.9976102135371764,
        0.9988667537690641,
        0.9994511628162526,
        0.9972007256463828,
        0.9987206817645327,
        0.9977134739989005,
        0.7223449817287272,
        0.9601724285442652,
        0.9964718988147456,
        0.9325348010314327,
        0.9897005567212257,
        0.9992083176737113,
        0.9785413628597573,
        0.19764869798083828,
        0.9996019129573019,
        0.6950326395337372,
        0.3864455779521881,
        0.23407621068442913,
        0.9952167760343138,
        0.994806971113162,
        0.970223771991396,
        0.9435278121470176,
        0.9543735718279012,
        0.9519494667633369,
        0.9806535717686699,
        0.8980293550117465,
        0.99569958618475,
        0.9434361929226651,
        0.9705894067822934,
        0.9103332829416809,
        0.5875504240831224,
        0.9352311526252712,
        0.9801030254309444,
        0.9878878841802542,
        0.47325093888951264,
        0.9817662655704746,
        0.09998285767129023,
        0.29702757818378017,
        0.27550090018016016,
        0.6594741957366069,
        0.5066688323601796,
        0.9992933058456825,
        0.7734145025361876,
        0.5887276927202889,
        0.9834188324917201,
        0.992255115037752,
        0.9964514942767175,
        0.9991003337271402,
        0.824408502890567,
        0.6403309658047023,
        0.6011798950164313,
        0.9561097063754083,
        0.994804120792698,
        0.9940041433822612,
        0.9893630923713597,
        0.6927947198618054,
        0.9955561403433858,
        0.9919740593217227,
        0.9976650339750279,
        0.9993251134293232,
        0.8245948215436655,
        0.8749877607325328,
        0.9557688577410767,
        0.003439708557691379,
        0.007506258798561372,
        0.008619789278039186,
        0.6890124360992558,
        0.9877574396898252,
        0.4526299634105751,
        0.7124349526849729,
        0.6479147090862629,
        0.5912996083351264,
        0.11579384656105968,
        0.9990172688939767,
        0.9936291925899977,
        0.9740900023471006,
        0.9941835503769001,
        0.9775963994265363,
        0.9728449401265686,
        0.9903701681376288,
        0.987100417502838,
        0.7880129857193827,
        0.7705034609103818,
        0.9170440718250493,
        0.9860833936382205,
        0.9989404737353446,
        0.997174071793303,
        0.9997221489120687,
        0.9989687009898045,
        0.9996735066119233,
        0.9431284451143646,
        0.9692882417774884,
        0.9574173146873527,
        0.9567040954234848,
        0.9769146747401676,
        0.9910781928627551,
        0.975654838822491,
        0.9914608334443262,
        0.7297098236184764,
        0.9990804141537207,
        0.9953902671252453,
        0.9985625561979123,
        0.8369981339309164,
        0.9880644552781017,
        0.891207384494906,
        0.9209371415606594,
        0.8991349887597532,
        0.8865201350737,
        0.4603930789422377,
        0.8776704775776263,
        0.44748104973600356,
        0.9993809901209074,
        0.9942331626705035,
        0.9655613038726841,
        0.30454337693918176,
        0.2775725877627372,
        0.28528864958686934,
        0.98562709144965,
        0.9274105348418248,
        0.691422408224923,
        0.09769266692274692,
        0.05969624661310513,
        0.1458854633934211,
        0.9839829647691586,
        0.9268082644796842,
        0.8173242920197709,
        0.9462539971617535,
        0.9903892970544845,
        0.9834874946879197,
        0.7685518853129931,
        0.7657032780721307,
        0.6161422784980046,
        0.9413307973246088,
        0.9990462928592898,
        0.46567470775190734,
        0.993996623088382,
        0.9916449088852984,
        0.9947835819437827,
        0.9964026587546209,
        0.9627163408807015,
        0.9970752849869237,
        0.9861362014127039,
        0.7642977990072932,
        0.3088788878127775,
        0.9739377765636225,
        0.9709859580245999,
        0.9742263032340555,
        0.9651233325377785,
        0.9903252696926357,
        0.9991793376585184,
        0.909799272498789,
        0.9558889485018734,
        0.9811128422679808,
        0.9935909003857459,
        0.9709546628166915,
        0.9976871742633839,
        0.49642808998718546,
        0.9727121684769336,
        0.9382518881547969,
        0.9988667025729465,
        0.9988553775010593,
        0.9990216509877891,
        0.9661674528175976,
        0.7011811751265835,
        0.6056575958247252,
        0.998467748107551,
        0.9768114931505444,
        0.9985798525117117,
        0.9985431240848327,
        0.9988643852899353,
        0.989378661611171,
        0.9974777656752339,
        0.9929927158583272,
        0.9087434640150027,
        0.9631610360220811,
        0.955155394239746,
        0.9654833103827778,
        0.9837704275437992,
        0.9996530844604136,
        0.9989910469330603,
        0.9392687184047963,
        0.9530070026898577,
        0.7939005538581221,
        0.9808876809933721,
        0.960315799527383,
        0.9064439430621393,
        0.997121827872273,
        0.9931369291077103,
        0.9991249018890771,
        0.9033004489973756,
        0.27588702805004306,
        0.18018447716910865,
        0.973848094840998,
        0.9740520272952788,
        0.9928366729312277,
        0.7035125430392372,
        0.14541812758487563,
        0.9651523201872463,
        0.6301352458557661,
        0.5923096822634842,
        0.3284352793887122,
        0.9985862436741273,
        0.9941646436504341,
        0.8985390513819916,
        0.9169685835437669,
        0.1692449115969603,
        0.8342218817803162,
        0.9988803775054486,
        0.9947489802061268,
        0.9972786641385336,
        0.7103559886466367,
        0.401310791864284,
        0.8638865112764778,
        0.997613001589022,
        0.9087335354722448,
        0.9989950954684985,
        0.9975808881297907,
        0.9502141209495975,
        0.9932950064845801,
        0.999900039502098,
        0.9998962335432875,
        0.9997509990784015,
        0.992953300051815,
        0.9972376018132759,
        0.9942254366510485,
        0.7780525648657128,
        0.7208752622618037,
        0.5482178678817068,
        0.8101729974007645,
        0.8246997813793954,
        0.7873726183710097,
        0.8170298401775262,
        0.8214450931605128,
        0.8620597965251787,
        0.4514521467025492,
        0.9984170902775511,
        0.9967097457115973,
        0.9959877813733916,
        0.8711107218927657,
        0.974897565295075,
        0.9910736437819114,
        0.8879250556042521,
        0.9792282042797528,
        0.9835918391005339,
        0.8864655246999409,
        0.32340881833749346,
        0.9995187455323752,
        0.9934231972096469,
        0.9993084592801996,
        0.9995261193450795,
        0.9992698684755303,
        0.9994227096829986,
        0.9177324113552396,
        0.7322308120459374,
        0.9954623813326997,
        0.9995480492746734,
        0.9972361271222483,
        0.9997283955059897,
        0.9662695576883946,
        0.654779014758037,
        0.77907697318491,
        0.7799442281016271,
        0.9933618904079713,
        0.99950966197836,
        0.8382888695847706,
        0.22102683427319858,
        0.6942968118476072,
        0.9288865224221459,
        0.8635982975543166,
        0.9273990716999934,
        0.996466640075033,
        0.9993003859648979,
        0.24083250338793805,
        0.9204436224299377,
        0.8180103979708521,
        0.9902785804779612,
        0.9445412070348984,
        0.6941918675023616,
        0.8291215285227932,
        0.4025894102648713,
        0.2711515777593386,
        0.3727606617830343,
        0.9537040997969134,
        0.011353463244409483,
        0.6596387706748462,
        0.9038648618892311,
        0.9672565069541125,
        0.6854119620948615,
        0.9967528807455228,
        0.999134104759649,
        0.9978518911584912,
        0.9734804951713675,
        0.8239348848751175,
        0.667884110070717,
        0.41837800004581976,
        0.6703439190420731,
        0.9251290798344909,
        0.9215244313333126,
        0.9757050461926541,
        0.4571140056361533,
        0.16282711390601115,
        0.16564534116342597,
        0.23736512089594541,
        0.9966561772171932,
        0.9974872699898181,
        0.9942325804483697,
        0.9972986039246344,
        0.8837641562054654,
        0.998906265955674
    ]
}