GENIES/results/evaluations/mms/llama-7b-counterfactual_python/counterfactual_python.json

{
    "model_dir": "models/llama-7b",
    "distribution_id": "counterfactual_python",
    "timestamp": "2023-10-13 23:50:27",
    "eval_score": -0.1754687455382107,
    "eval_brier_score": 0.1754687455382107,
    "eval_average_probability": 0.6748804191521957,
    "eval_accuracy": 0.7253333333333334,
    "eval_probabilities": [
        0.4843526222955712,
        0.22469472145044495,
        0.3857181260620534,
        0.7026058678988842,
        0.20415130843528606,
        0.4396825393195402,
        0.8420353885747172,
        0.8397354117182647,
        0.8437381585869035,
        0.8512546204095846,
        0.9552740888341635,
        0.9571671849896569,
        0.8885514209753551,
        0.9173404936950286,
        0.310679798260626,
        0.8003759900540653,
        0.7375281332099857,
        0.6634763456434162,
        0.37509996822367986,
        0.6014749752560524,
        0.7745210304964653,
        0.9932922431443477,
        0.9933403793254291,
        0.9665625508067215,
        0.6298727136457752,
        0.4522797950517825,
        0.5895721366799573,
        0.1649430559754569,
        0.14647585928885054,
        0.23199507603909653,
        0.861981351864519,
        0.10792514535007283,
        0.9084699258895911,
        0.5942317294394492,
        0.08454717385616162,
        0.7274196726994102,
        0.6090490226989254,
        0.8420523260087812,
        0.7123518651183528,
        0.8888835511067555,
        0.5008838619970489,
        0.1750521661745999,
        0.6843720803686225,
        0.5993843602706475,
        0.5397019839666805,
        0.5551432711155828,
        0.46020188887045627,
        0.6508628687374728,
        0.5689017577929276,
        0.4889673532281066,
        0.8761408274949978,
        0.32264204182789946,
        0.7954856005970857,
        0.6948468942741242,
        0.3602462720319787,
        0.39242044376805163,
        0.1924994195134545,
        0.5704936399752893,
        0.9645850122193995,
        0.9678914914876571,
        0.9830659230215275,
        0.9761247981415965,
        0.6979999488246091,
        0.05103420075723886,
        0.9222093784086254,
        0.9247697297931232,
        0.8395908859355214,
        0.6442724084343932,
        0.14897476753413008,
        0.2946079103783047,
        0.6338098516286027,
        0.23096005071756082,
        0.9780020302424752,
        0.9727449725479739,
        0.8888455594001851,
        0.3415208589947446,
        0.36119442519954015,
        0.38972172024438956,
        0.7345619648131736,
        0.9924254442270001,
        0.5286887082146208,
        0.9303652810663503,
        0.9884723625347563,
        0.9975608787618033,
        0.9595548883582619,
        0.8569072307352289,
        0.6525483762386876,
        0.9332225670501617,
        0.9223429689291583,
        0.934757980455164,
        0.997122745370845,
        0.685830190357718,
        0.998966782071429,
        0.21695993997140842,
        0.9561256369315377,
        0.9782939081395691,
        0.6175384724139861,
        0.6705095026205895,
        0.4304526051695369,
        0.983823251878189,
        0.5801270141471362,
        0.9512930525613518,
        0.5572065486887938,
        0.5002807329914989,
        0.5388004476866697,
        0.8543260809981555,
        0.8736514998452674,
        0.802005712050069,
        0.988508636916164,
        0.9869041640424725,
        0.9890805831530971,
        0.9850358007628309,
        0.9867433970603994,
        0.9450710714386843,
        0.8750243220305197,
        0.9057882089607298,
        0.2563454325092409,
        0.7337497844714705,
        0.8566989177812139,
        0.764123128772933,
        0.9779031699717761,
        0.7943939342492198,
        0.5296490810641357,
        0.9518203924310499,
        0.9886488104359267,
        0.8798104925900704,
        0.8655018662721525,
        0.755200954182578,
        0.4658905375928659,
        0.7603586301449126,
        0.4294999546205057,
        0.2906233176056966,
        0.5052671685987512,
        0.7053345415948867,
        0.477526430788528,
        0.4144301863437761,
        0.7840999220331983,
        0.7549702461823816,
        0.21677930387925295,
        0.7271263490271134,
        0.45640274768149947,
        0.6517728730338932,
        0.9968939266406064,
        0.9853131860083612,
        0.3465627655474641,
        0.4340516154764146,
        0.6440732612655745,
        0.9097547174821482,
        0.6601025027525539,
        0.6800613738792611,
        0.3226846065741683,
        0.3766464180167305,
        0.9604067796890465,
        0.39336715823150814,
        0.5797154520717442,
        0.8734399272181048,
        0.6983905169536223,
        0.7652434250175881,
        0.6986702587523668,
        0.6088709108916379,
        0.6573776563596336,
        0.9101664989199667,
        0.9119777025478157,
        0.5600863597950334,
        0.5533263318602055,
        0.1778768561345217,
        0.49538561784333,
        0.2685274190128859,
        0.837209317379283,
        0.2179961019909645,
        0.8130597466418334,
        0.8368529487882933,
        0.740952193314197,
        0.8908449683092284,
        0.4077356406580326,
        0.3521437606047775,
        0.45196332432314545,
        0.5961164980551978,
        0.9470699546041359,
        0.32583594328773974,
        0.9725861114717129,
        0.8074125251168495,
        0.8791759428855518,
        0.9987059234927789,
        0.987396017558775,
        0.9977336055939352,
        0.38786283673990885,
        0.3439342687176166,
        0.6110474765707845,
        0.23238856446646972,
        0.5876805710781564,
        0.7472980995721806,
        0.8124906496320525,
        0.7742834785294668,
        0.8449909585964881,
        0.5495351800629933,
        0.5069179630260169,
        0.2638752839057136,
        0.3694151516391273,
        0.969096898648028,
        0.43796619963636885,
        0.8210289573147568,
        0.9561557796009005,
        0.9227492184697658,
        0.774335954388149,
        0.7563443007839105,
        0.4589975492806846,
        0.9477134977601251,
        0.4745810628185364,
        0.32501112261776455,
        0.5740843824150692,
        0.7230921614242847,
        0.3361504134706664,
        0.6805301310441592,
        0.2978450845464128,
        0.12950426397825043,
        0.892306925174604,
        0.6629317250470019,
        0.9901663634497919,
        0.9885626204716903,
        0.9982851061492726,
        0.9925221173815302,
        0.3814347798517941,
        0.6158935969913614,
        0.47430022989321013,
        0.8057331050876916,
        0.7429097738925272,
        0.765642837946031,
        0.9440426948115858,
        0.6375431377057864,
        0.7220970255643828,
        0.6710730069364639,
        0.37702816867664124,
        0.8383385399522091,
        0.2673820511506363,
        0.6910702783627518,
        0.45315283233303427,
        0.8351161884180321,
        0.8274309218294325,
        0.8592061453989394,
        0.9998380054128247,
        0.9993540741464391,
        0.9985549008292235,
        0.3411504681943698,
        0.6510183750829573,
        0.5873816996872605,
        0.9792015877244932,
        0.9666469863283838,
        0.8499883656351837,
        0.9575302257976565,
        0.09749118701320758,
        0.7847982976043024,
        0.8024574426570428,
        0.7096409353826514,
        0.8985437597660034,
        0.9540879282409616,
        0.8255201245828392,
        0.9222615477213572,
        0.9457481742844328,
        0.5227108768695887,
        0.22238603286611613,
        0.6263413321877782,
        0.3882666365865469,
        0.832091605455863,
        0.5652531276906667,
        0.35750970649656466,
        0.8216424373843755,
        0.8430963747107207,
        0.9958408983388802,
        0.8789489958675026,
        0.9977447766488151,
        0.11871088831865251,
        0.9839248055062848,
        0.5268046153915711,
        0.7265267741985629,
        0.5813615558808456,
        0.9305654143817702,
        0.8507308038061672,
        0.9590369089107282,
        0.4079013977593265,
        0.2508599271545251,
        0.2450841077407263,
        0.9514093113357162,
        0.9646270072943622,
        0.9256937628207386,
        0.6613955548578225,
        0.6352897438927699,
        0.7300379461246094,
        0.8953288493082885,
        0.974268722018911,
        0.34341057639314826,
        0.5790608850999861,
        0.9362654273344582,
        0.5584356008287583,
        0.8976731703148246,
        0.8852877725452684,
        0.973256409096374,
        0.44149861700460985,
        0.4748947454532393,
        0.40436351331462,
        0.3050186340173695,
        0.0261073202349279,
        0.2931486355378288,
        0.4735154937363918,
        0.5954500741737598,
        0.6820251687729438,
        0.44738358275788664,
        0.2938061985341506,
        0.2310858234325697,
        0.8164122516851113,
        0.7373423273772889,
        0.9256960070011448,
        0.9311881677571839,
        0.5311664042190893,
        0.6570358180581909,
        0.8118203414663053,
        0.8449425076318695,
        0.5168523128232432,
        0.8921942620369983,
        0.9697434568485037,
        0.9750755429562453,
        0.4030858795246084,
        0.3248615605301663,
        0.2529015789736925,
        0.9452774156574238,
        0.9977760546712554,
        0.37682767600679673,
        0.5730686803770619,
        0.381483793062762,
        0.540937496662885,
        0.3251315920684583,
        0.9497204351939916,
        0.9320214923245236,
        0.4295931833510831,
        0.6598458764624178,
        0.4420308756306507,
        0.9954418939019848,
        0.9945914730417408,
        0.33293346332387364,
        0.8636427724755132,
        0.29100674349047756,
        0.6589685200926745,
        0.34154395126259574,
        0.2004417403322366,
        0.3841217926009286,
        0.14658318673169202,
        0.575144624155896,
        0.40876341050172194,
        0.9023058479369114,
        0.5736313812526624,
        0.8577925460550604,
        0.6514077456531314,
        0.32896850845745834,
        0.34481863150299175,
        0.11044650282854224,
        0.48970147547208,
        0.33823244745169423,
        0.7116037533407946,
        0.9893780561138141,
        0.7922783335895691,
        0.3624498224609958,
        0.453228366358448,
        0.7803910162982505,
        0.9906010185339238,
        0.905788007399611,
        0.891165125139675,
        0.8241006935986177,
        0.9313760376942989,
        0.6360596911252927,
        0.16323827145546518,
        0.185136043635036,
        0.37755845235469354,
        0.14897213850299673,
        0.10847572956242284,
        0.1430764683244355,
        0.9383642900491774,
        0.785162043902163,
        0.8223269548556007,
        0.46469309881097093,
        0.3616444540437911,
        0.6009667436147323,
        0.15568129237855188,
        0.5661612690329585,
        0.25358552174733917,
        0.6739631563379685,
        0.7899753902005043,
        0.6006958960293445,
        0.17286978032286585,
        0.18071454482908267,
        0.254895719542939,
        0.2942076101122573,
        0.4272851402487674,
        0.18302334693752043,
        0.7001554830136063,
        0.8170822983745232,
        0.7298926403672537,
        0.845017546559736,
        0.37340021912862925,
        0.3642756563232738,
        0.9832638919396193,
        0.9789091590128057,
        0.8436701441234464,
        0.8578610944501587,
        0.9626393050121065,
        0.9671810087340442,
        0.998305092196754,
        0.9976867084336507,
        0.9985089727311368,
        0.1758974262168104,
        0.2107328070973273,
        0.2169296087804491,
        0.6908985673245108,
        0.7549298907121633,
        0.8195653408783069,
        0.9923718830319643,
        0.04328361999155196,
        0.9856281620352932,
        0.9144406203304771,
        0.8772978452819767,
        0.8829912493115739,
        0.6750841491841673,
        0.8009783625083384,
        0.5798724499735769,
        0.7197261724737553,
        0.9812178785019543,
        0.9660009603371192,
        0.9822222842469542,
        0.9442558604611158,
        0.6869583904033407,
        0.4035665771727618,
        0.852375591410481,
        0.7005528910584731,
        0.9975969451778324,
        0.6066242822919192,
        0.9962325608107654,
        0.9904564322037547,
        0.9947952223951022,
        0.9922723048459804,
        0.1786701092422498,
        0.31495258361702894,
        0.442107436939151,
        0.3056208410746952,
        0.16883994680719497,
        0.8355684421129705,
        0.61298318241381,
        0.549116987041242,
        0.8418840842163138,
        0.9947482591166735,
        0.8541659060360839,
        0.964445910108672,
        0.9624218619966806,
        0.9439318204270669,
        0.3178543150183699,
        0.40417968313099506,
        0.2426712114976668,
        0.4735475597436801,
        0.8641359223639313,
        0.8226060182684695,
        0.714507726950713,
        0.41221645368495674,
        0.36653816965890146,
        0.9086157823466159,
        0.8079856784521684,
        0.9995532922667909,
        0.9508214723366578,
        0.7426384153831436,
        0.8397493488647516,
        0.9342675592909289,
        0.551369256840126,
        0.7561831549238271,
        0.6478532803106335,
        0.5716136340721738,
        0.3302770855846888,
        0.32212191378485006,
        0.9955762815355498,
        0.782122354850461,
        0.9465227558990905,
        0.9734663183762947,
        0.6385580927161965,
        0.9335734411358303,
        0.9857931643352594,
        0.9912364266093483,
        0.9953209639132818,
        0.6521083974307822,
        0.5404951280264555,
        0.6614047532368741,
        0.10880889812315851,
        0.08810189510344088,
        0.09158169353298648,
        0.9450330162741649,
        0.8959104685480859,
        0.7373176254369456,
        0.6964373849293592,
        0.980788362010884,
        0.7444963674449092,
        0.6078887174361478,
        0.5775640039575861,
        0.5166579732713645,
        0.41522776875820666,
        0.3463021743354931,
        0.4032110264739474,
        0.5957411845854128,
        0.8105313444773026,
        0.25294082975652893,
        0.8918855397578194,
        0.6784428574582663,
        0.7644128923200174,
        0.3136192348084028,
        0.3148845342046186,
        0.3318200828668,
        0.6456743063856768,
        0.6397292688048386,
        0.6043406175355237,
        0.5741155679974183,
        0.26533751944090117,
        0.8521940387432133,
        0.698933022313527,
        0.8225234825140848,
        0.9096112813734336,
        0.9275609242014694,
        0.8774236616680966,
        0.8932359015864212,
        0.3219918369194341,
        0.851632643881991,
        0.9331733786238992,
        0.9948545240720745,
        0.9927254464645959,
        0.99655523378824,
        0.999283362138215,
        0.9906622205942287,
        0.748157686396229,
        0.56205503577177,
        0.5477319366801455,
        0.6178528892653536,
        0.9409560709256672,
        0.9628789962458938,
        0.9585551142572654,
        0.9111474020316939,
        0.7847749303602637,
        0.9639378743078788,
        0.526871436747909,
        0.7463027567011454,
        0.6175742119362264,
        0.9534934447908968,
        0.995727907221901,
        0.33176923518630463,
        0.011785845083598034,
        0.8230991666418661,
        0.7427939810605295,
        0.6738926123198308,
        0.12413674578284231,
        0.6545358861058589,
        0.8823578259514873,
        0.6720309344507112,
        0.5010368056617268,
        0.30584659803761116,
        0.6452013606561227,
        0.3016905423990592,
        0.35203576178620133,
        0.7224172775655784,
        0.6788990082222132,
        0.9800915597878447,
        0.8886730638954716,
        0.876049169425929,
        0.25601716418821285,
        0.3160744546326838,
        0.4486751760612351,
        0.9694288707572379,
        0.8816408828862281,
        0.8364060779674606,
        0.8744278440817714,
        0.26975355593776335,
        0.7269307566366905,
        0.9534466426049232,
        0.9962383251832403,
        0.9975811293016283,
        0.9583533498980655,
        0.8263543957969733,
        0.9862036106313488,
        0.9364603220226849,
        0.9671025562491385,
        0.9516768412854102,
        0.8208958235563873,
        0.8644885497199346,
        0.8151159175773443,
        0.7898148088331329,
        0.7797290219522314,
        0.9681711284010747,
        0.3616110178219666,
        0.3548223930650358,
        0.2578899685305784,
        0.9854280038242171,
        0.9909032722040633,
        0.9339042022890955,
        0.970343861198954,
        0.9702264988137125,
        0.973804592720684,
        0.7311425283965081,
        0.7036680740444516,
        0.960351585256072,
        0.9595719997237683,
        0.2392677576287028,
        0.8935516124099029,
        0.3701373659469784,
        0.5069919888397424,
        0.49457282971357985,
        0.6379019823510723,
        0.43019081308128826,
        0.39274005032396964,
        0.9236056201205685,
        0.9073617576510592,
        0.9788660911354283,
        0.8153382922351295,
        0.8933051505519948,
        0.6062231751648856,
        0.6418748822829308,
        0.3240501392185547,
        0.25297187619688793,
        0.5051515772609185,
        0.5044376730941846,
        0.578958657941262,
        0.8764618573848805,
        0.8551578245698512,
        0.8641272693909816,
        0.9828201198407502,
        0.5299169455969182,
        0.04174288069439846,
        0.891754355530777,
        0.8399409646674901,
        0.8678028352861525,
        0.6004941748752011,
        0.5734838621489114,
        0.698212713292916,
        0.7621181063203543,
        0.5316610572659475,
        0.5207423214731915,
        0.9993783811858823,
        0.9987307241595405,
        0.9969521105138329,
        0.26125839436703,
        0.4375661350145393,
        0.3187972270088723,
        0.9944249761149667,
        0.988486370106088,
        0.2984137222986867,
        0.8681085658524168,
        0.2521577426663121,
        0.2354201382534566,
        0.9805006922745992,
        0.9876047302181344,
        0.987027525644674,
        0.4132575966972929,
        0.5403329320270716,
        0.6387441753574179,
        0.7583380307727381,
        0.8334332996306729,
        0.8880645458398723,
        0.6438140374322256,
        0.8990994285383607,
        0.7486371980738391,
        0.6952185030663109,
        0.7493122489125006,
        0.5462208667937798,
        0.4396622522799706,
        0.7552362843958246,
        0.6370005057214981,
        0.40206527575378315,
        0.8175730500876526,
        0.5620608890218242,
        0.9479659292804082,
        0.4980590481957605,
        0.7591332337950136,
        0.4126837449260957,
        0.3680000190820979,
        0.6175938208464569,
        0.7704852806679541,
        0.675340386440747,
        0.7462486879257555,
        0.9496039706463149,
        0.9420107018845293,
        0.9955801158387716,
        0.8375628800635365,
        0.8488408581815696,
        0.9056085610799984,
        0.9625277405737737,
        0.9786600700086012,
        0.9886413624607748,
        0.6136265562079644,
        0.506994150172779,
        0.7455792606284095,
        0.09982596619414796,
        0.6318534571050279,
        0.7281947691358995,
        0.8988171834202898,
        0.9078442618442129,
        0.30832085037607737,
        0.6626360339018582,
        0.4888680113383034,
        0.6418490417131094,
        0.39751479830565417,
        0.2780682445024043,
        0.6939996319605046,
        0.2697992964091317,
        0.26915482319527967,
        0.27701147758038025,
        0.9786417325136599,
        0.8010197443776002,
        0.06072002898862294,
        0.9813966327892548,
        0.7561644219588489,
        0.8479223788710599,
        0.1684937863176891,
        0.34818473299737307,
        0.3153006329585343,
        0.9843892451703571,
        0.9850952743213386,
        0.9863338151048265,
        0.6671025433293211,
        0.407398922099779,
        0.6847378247369924,
        0.5509221773872766,
        0.7426896450135336,
        0.5260237548841197,
        0.8520306133086543,
        0.3655669700036778,
        0.6144024913587254,
        0.9429927615765711,
        0.9180257003698183,
        0.9903241627348671,
        0.9622512174005808,
        0.893621400244313,
        0.8364742199268314,
        0.9788073110859514,
        0.9926743997206258,
        0.7955517433753198,
        0.5889311938368267,
        0.9768361021636045,
        0.7024371300709127,
        0.852261282562871,
        0.7083188644241825,
        0.6900042691846968,
        0.9808909028411393,
        0.9162320908244643,
        0.9380606600708041,
        0.9030478530840719,
        0.7871811367652248,
        0.8213755675118218,
        0.3869209201899042,
        0.4929042174272875,
        0.649151671456852
    ]
}