Files
GENIES/results/evaluations/cra/llama-13b-code_hard/code_hard.json
T
2023-11-11 19:44:09 +00:00

761 lines
21 KiB
JSON

{
"model_dir": "models/llama-13b",
"distribution_id": "code_hard",
"timestamp": "2023-10-16 03:00:17",
"eval_score": -0.1619484641910135,
"eval_brier_score": 0.1619484641910135,
"eval_average_probability": 0.7045951567347869,
"eval_accuracy": 0.748,
"eval_probabilities": [
0.5971330196842758,
0.44467348834276577,
0.3508822665079333,
0.22550538607248874,
0.5512064970348728,
0.618317625736441,
0.972601585749105,
0.9974606346668995,
0.9200044606803487,
0.32537923376933336,
0.9032355098195496,
0.4501592032443385,
0.9966880050946532,
0.9984672734105563,
0.9966448765101859,
0.6158256935074851,
0.7312174366917725,
0.6114881260991881,
0.9206762430780913,
0.9898796119044834,
0.9947521913846666,
0.9976556549863635,
0.9959625381145335,
0.9918446142522297,
0.9971036698886114,
0.9977513479434891,
0.9903135119505805,
0.987406360621989,
0.992922609098029,
0.9940244936303636,
0.3264528826424906,
0.48662511093508354,
0.1935456625191323,
0.799511483481871,
0.7299619144155318,
0.7884641193455665,
0.9955914389681868,
0.993788557161634,
0.9882283503062758,
0.7422953907656745,
0.7183438087048637,
0.6294541861934566,
0.9906947320342424,
0.9913797964733825,
0.9936966077960605,
0.8192504599626799,
0.5215862453488236,
0.6609845051605377,
0.7568160038214077,
0.8663365418962621,
0.9180787781436427,
0.3610764983787337,
0.4492290169846057,
0.838659095505696,
0.5200725047938236,
0.4467055467907698,
0.5185472390698374,
0.3703869968514874,
0.7495094672721049,
0.7082700997460305,
0.15820936166592028,
0.21821991461147586,
0.1595094088919582,
0.3690170695069098,
0.5471241526333008,
0.7384054907836195,
0.45638117567659386,
0.38127588494516096,
0.610672643249728,
0.9523660505506351,
0.8666250987181464,
0.892227585164628,
0.8331127822956779,
0.9737132259697442,
0.9954894125484538,
0.5315297556272346,
0.2554615229317341,
0.49201383716235875,
0.2851783171531316,
0.7655816651590914,
0.5526279721571707,
0.6932608688270284,
0.7190738684777641,
0.3111005140773995,
0.9941666431908279,
0.9946270818749137,
0.6550697155272649,
0.3862650548135136,
0.6082045147311037,
0.5831939645380714,
0.998366823299588,
0.9964447074454028,
0.9937708899028398,
0.6646950838490455,
0.6932277995517381,
0.6167111582070438,
0.7897440294348484,
0.6850297068144469,
0.8215816824292264,
0.731117446566472,
0.8983137796395363,
0.8754172683655554,
0.5579493464640717,
0.5997008073576175,
0.6034523366117562,
0.6251244387789014,
0.6611352084323987,
0.8415497833638377,
0.6352883213434496,
0.31253709832191034,
0.5097406958763846,
0.8483670448271621,
0.8809083151311935,
0.8463588004353959,
0.5626689789432233,
0.251075466753698,
0.4476914918835233,
0.8658310645726038,
0.9143768137955991,
0.8635455905371484,
0.31283050357286574,
0.9328440827747395,
0.6998406254465543,
0.5194284344941645,
0.7253182870512125,
0.9443653687854601,
0.9976389158220909,
0.9921087858249783,
0.9931461004884379,
0.9915065368925664,
0.9585416266827337,
0.9744368755145065,
0.5438690151724417,
0.48652589940322344,
0.40224126411337663,
0.9948140307566056,
0.9827503547926927,
0.9952190668156301,
0.21733810868278552,
0.20540101480408068,
0.20752327761174982,
0.9440078487038572,
0.5419824031832567,
0.4549048542514603,
0.801380341388067,
0.06931511103562882,
0.8552181014933598,
0.9520940292069989,
0.9667730647060694,
0.9857328107886205,
0.5546591076849194,
0.5989774015272099,
0.5185065129830113,
0.20431903886826097,
0.11694702536870218,
0.11155960809588977,
0.9972343847483546,
0.8810085078764425,
0.9862566033966147,
0.7489657611442775,
0.5942433094433012,
0.6281402058829513,
0.4658765074193487,
0.7778939413477512,
0.6713525453996299,
0.6745356469771497,
0.886162922863281,
0.9455964133582283,
0.04563799747732905,
0.07599917874524575,
0.0726123151906676,
0.8245285273248787,
0.18815221651061198,
0.6677674286851213,
0.9957622875378318,
0.8942582544969285,
0.9372794674363032,
0.5639020347906208,
0.4759266158932699,
0.36844903915726435,
0.9926429554546831,
0.9976290381902887,
0.9970092508306564,
0.8620643621387638,
0.7345302493869625,
0.7624033217804433,
0.5451183321422849,
0.2997375539921541,
0.46764078356444294,
0.758928556777525,
0.4333110854591342,
0.8364793901024272,
0.3081631185801099,
0.7562881079960107,
0.4865223446718706,
0.9173809250263072,
0.7234266114126915,
0.7653459580222511,
0.9600015231346305,
0.9623294919153005,
0.8041684482356951,
0.9978622705208873,
0.9962109521831958,
0.998279066589707,
0.3641728089209725,
0.5692119164380791,
0.5124952969068001,
0.8393811723174623,
0.8939804668430034,
0.8460802545003293,
0.9577468177623535,
0.9402013601118152,
0.9423234499477713,
0.8149845119641174,
0.4209780443965191,
0.6250744193390951,
0.9561880771510836,
0.8153724302737001,
0.66177294254264,
0.9936307905317067,
0.9976252656100089,
0.9970392875819066,
0.6292997276282144,
0.7822369331573786,
0.4953660603522381,
0.8712758154026997,
0.3999578977800988,
0.7383830567382206,
0.9051276893310259,
0.8536132573691908,
0.8968179256704909,
0.9495778021037985,
0.895380800878337,
0.8567734758249075,
0.4926921092288371,
0.4893834838598089,
0.38625137729202375,
0.17879392152682205,
0.49468007475396486,
0.2382787450466835,
0.7961380723038264,
0.2029995962996124,
0.42430336872823404,
0.9952234317333614,
0.9950693863047331,
0.9966089805314909,
0.5708054874587114,
0.35499183370641446,
0.46639298700060194,
0.995759824018809,
0.9891914609693496,
0.9928060148156134,
0.3340025214176531,
0.42348354155281454,
0.547212604782586,
0.9277115532647237,
0.8669837366400213,
0.9146854531881777,
0.6097124018123323,
0.6614530640141394,
0.6813919594745612,
0.9899990890644352,
0.9891421426067818,
0.9946369080423133,
0.7108067522785247,
0.46431360795735344,
0.8192256655282186,
0.9967208357187569,
0.9964270295338993,
0.9954635847772504,
0.7348405434368261,
0.39067440126130093,
0.5977539054048814,
0.6965149703251776,
0.3519958659272834,
0.5260206865826557,
0.9547598962657324,
0.9952565234212769,
0.9897636118963313,
0.9379536790254652,
0.9620236332358711,
0.8891125955518628,
0.996902533587488,
0.9983786842736222,
0.9980827370051445,
0.9973797006097594,
0.9939947238224792,
0.9988762396588473,
0.3239550673000049,
0.15325584128022463,
0.19109893072787645,
0.5132822700463165,
0.630492269402123,
0.3802871531002431,
0.7658243593665214,
0.7960915496142946,
0.8279096185466777,
0.6160719574768923,
0.611508891526325,
0.6233063686916266,
0.9954336589995432,
0.9936560082950858,
0.992572425362703,
0.9802572432340566,
0.9534165152072127,
0.9795116321832841,
0.9945411724290812,
0.9972570808192558,
0.994188313343487,
0.7357732078928793,
0.9055342326820153,
0.5189086880874614,
0.964627902898834,
0.5960476790658011,
0.6101000340684858,
0.26374706804293135,
0.9535728929146982,
0.7031280720360337,
0.7935166410987489,
0.08598544163736525,
0.6074898677416438,
0.7909938453671831,
0.8500050098462542,
0.4968671557572647,
0.9171342733794327,
0.8767615168594797,
0.9080153426948984,
0.41578480349774316,
0.9755932696909343,
0.9084628416706508,
0.32517496474795027,
0.41098086197201633,
0.3230208733557308,
0.8553685169587392,
0.7149430805821207,
0.7160537410586696,
0.31177176543408824,
0.3835762208085345,
0.2891363673560991,
0.2553914008826556,
0.7649806961695255,
0.15677458251932921,
0.2292903196342443,
0.8115225105910701,
0.7493533445398846,
0.44655388137799096,
0.3755104738649063,
0.38340729167066234,
0.9956178448848118,
0.9811864634627985,
0.997173062500046,
0.6917831397229929,
0.7524097110699809,
0.7507887765426247,
0.9411931125242665,
0.9384157810540795,
0.9448227743000163,
0.9973885323076561,
0.9968659684145734,
0.9966241810081773,
0.5180147357466203,
0.18010593688962767,
0.1934714289925339,
0.5996532281872927,
0.39587512311841444,
0.7482159260829336,
0.9787927294857244,
0.975194539342629,
0.9801953304086783,
0.8237677908226619,
0.26848248777321504,
0.6555579636640517,
0.736342578186761,
0.4830471605918383,
0.5882873999826095,
0.5306568422482655,
0.8107861685305554,
0.8765280690152122,
0.938382601377188,
0.9483425967864894,
0.9439661865482604,
0.447727237357616,
0.6146433981030238,
0.6099659058471191,
0.871025656841107,
0.9279371137719096,
0.8793488105121875,
0.9048611808975472,
0.8781489999942064,
0.38956709526498834,
0.22200580477441922,
0.39456988650214925,
0.9944645713520234,
0.35825376779704454,
0.4600200324708731,
0.1284335598910994,
0.8885720377343067,
0.928578096029777,
0.8149932114488967,
0.9984378348941972,
0.9980173893271895,
0.9976187029420235,
0.9977419270884219,
0.9977806856713254,
0.9953806463056531,
0.9974099653365638,
0.9968597694928246,
0.9978905471241271,
0.9971320642898336,
0.9854412917652609,
0.9920695380320377,
0.9691223578447553,
0.9265882883256656,
0.9439480254003271,
0.9969914745109696,
0.9962757252200783,
0.9969685608026778,
0.32582989785737515,
0.673560585628684,
0.39751947610998095,
0.9975016024786141,
0.9986267026301133,
0.9981182490697404,
0.052294682025903684,
0.9098208236843109,
0.0421432634196361,
0.90995530329513,
0.4545590948937962,
0.8787976486525539,
0.9953009202106798,
0.9784608574196725,
0.9816363403216827,
0.9931866884875025,
0.9833446127236738,
0.9293803453966236,
0.8932200196981361,
0.8603469391583738,
0.5474982296680282,
0.4240134733321204,
0.19897301484107865,
0.2838132499440814,
0.9915825440234309,
0.9921615784695816,
0.9926768025156557,
0.4851239212262284,
0.5414674089710133,
0.2325777587256549,
0.6457295077059741,
0.3012747231572461,
0.8055598989492645,
0.9199697634282442,
0.8018093470015405,
0.8229898804640956,
0.7577914526435465,
0.8638177336167262,
0.9719101863566654,
0.6095430865811798,
0.7435821321108496,
0.6801112533265335,
0.846385196404021,
0.9262120439287118,
0.5225651950705298,
0.7602696466262404,
0.848747518119428,
0.7023234239290502,
0.8026279932364913,
0.929159559511268,
0.8689848734591354,
0.9451217529686537,
0.9049525761377175,
0.9284810209361885,
0.5948732204627466,
0.13202420018904104,
0.24368697734698685,
0.17583211125819606,
0.15493919523795763,
0.1576207958893694,
0.2506316255786988,
0.13858944017490138,
0.397624451637195,
0.6526932018851657,
0.8043594898654252,
0.7265550594969863,
0.30290762619577216,
0.3162054217749303,
0.35474460505786837,
0.9973907550966081,
0.9963118563696609,
0.9966020875732957,
0.7184868567838343,
0.6803793050198106,
0.9276803292940616,
0.9394652011400151,
0.8666261872944188,
0.8540877068975277,
0.9916544494663336,
0.9937359006338551,
0.9872545298421799,
0.5674636188926472,
0.5233125605142628,
0.5456422418826239,
0.9109575970220392,
0.4605651614734205,
0.5874451777546699,
0.5564108018273912,
0.9304022070597827,
0.3856794787610087,
0.92803717998352,
0.9204809540225543,
0.9177274172689677,
0.32883024545770356,
0.7585849859485422,
0.8120813268753475,
0.9488069553324188,
0.8060455072539717,
0.7967181166050188,
0.99400223772901,
0.997754203394482,
0.997416677561705,
0.8077427240671915,
0.27464075710753133,
0.3490325081646054,
0.8710853306262895,
0.7504480284223588,
0.7591544162799686,
0.32380177055224924,
0.32470370938622395,
0.3335283577464905,
0.9892760450798903,
0.9733835005157998,
0.9832019653685631,
0.6388803016888451,
0.6841912800601466,
0.7545136281239033,
0.6770398004879878,
0.7095538862546407,
0.5167342534409963,
0.7871837959412227,
0.8343676243217775,
0.5852209688303016,
0.996265833588301,
0.997234944662512,
0.9978946893615183,
0.9987587084332269,
0.9985572702821007,
0.9988215987424931,
0.5138899354526181,
0.31079444143426216,
0.16950981820840424,
0.34284259001499495,
0.3698779681912824,
0.8890472444543358,
0.35472843609672544,
0.6410239240952122,
0.9371253009637598,
0.5267736895899671,
0.696909227931896,
0.6319021008415997,
0.6149045069288537,
0.6013265770902884,
0.5189708517821294,
0.9959839249567516,
0.9955250030300467,
0.997047787209069,
0.0988342245889906,
0.09057525064877647,
0.09113580626568929,
0.4211076216364664,
0.605279812548735,
0.9946830231718615,
0.997954349825716,
0.9565409764062124,
0.8587834576664946,
0.8788818217409178,
0.722697186954255,
0.9338776765207127,
0.4286692720587415,
0.21119805588694995,
0.6093606803642407,
0.5990245840353916,
0.7999572619655057,
0.499918274976868,
0.9980495928940716,
0.9967718136644174,
0.9945202764354687,
0.3395732586473753,
0.4587251374003064,
0.5885552740733472,
0.07728310469596111,
0.11391230466305273,
0.14332517452424945,
0.7546542281719852,
0.9025302958805614,
0.797383690455332,
0.3743871229707132,
0.4065779840937246,
0.4834183550834892,
0.996978565146485,
0.9967404487418942,
0.9966042852722788,
0.3919604317344186,
0.6926640773751618,
0.3769520504989519,
0.4304088697829662,
0.805278956831882,
0.8794384978474047,
0.5905248006184698,
0.947238558217892,
0.97033899229421,
0.32023120039785824,
0.9956406184835349,
0.22576550705300247,
0.1968799818191871,
0.14942648608437814,
0.21937156975034355,
0.8379941548755044,
0.8899152681527837,
0.8698416641497068,
0.7468355803260881,
0.45172791676171486,
0.5439005322404727,
0.529624092229246,
0.42898399718268043,
0.679430638439394,
0.9894479533618615,
0.9887467675569795,
0.9885582208333299,
0.2889572665099925,
0.2556204093641011,
0.5346180926210652,
0.7110532234827087,
0.46170748457780963,
0.7320820487464079,
0.8702664645432859,
0.925897388409993,
0.9236269118150244,
0.3937259419146665,
0.6161776642789653,
0.9222322969580081,
0.23593253165463102,
0.41494396796024186,
0.4186356582777977,
0.19896163247233772,
0.930355348051026,
0.9282159685737652,
0.03265371597452227,
0.4349717425511037,
0.11423599054863998,
0.8012805064670425,
0.752436545972938,
0.7702650224259457,
0.5448837539974893,
0.8089807121265826,
0.8494121061155905,
0.505755702301533,
0.36764977816895683,
0.4011236754002748,
0.988203716859264,
0.9964995969244714,
0.9966703441458399,
0.8245169444173014,
0.8012407723544566,
0.22025105363632058,
0.9968767706329911,
0.9969434010902842,
0.993058209637324,
0.21362178840279408,
0.6804014070592559,
0.3058845600099228,
0.7733711116411883,
0.13666026703543266,
0.900176057418422,
0.5549206356461195,
0.4975190131636458,
0.702012827889564,
0.6569169290761343,
0.6618173558169107,
0.35397055856228254,
0.48413306258902455,
0.49538867278130844,
0.5642092493689036,
0.9979101177361218,
0.9971447523369744,
0.9983976059898667,
0.9297467699486724,
0.8961000237590034,
0.8506749904194214,
0.8241772413522299,
0.6654479210021408,
0.753334182480598,
0.9849558080955679,
0.9814733693209813,
0.9783890739674788,
0.9985020827633576,
0.9987267450571227,
0.9987628295182167,
0.9761947999643008,
0.9901927434558737,
0.9871102266618796,
0.8535659717686003,
0.8308657588544046,
0.8435919725614686,
0.9958510932943228,
0.9866477704635961,
0.973749264688289,
0.07523310938134684,
0.12136534306065228,
0.10636528100651577,
0.9857809763345919,
0.23066722846800292,
0.9889074039292607,
0.9971678669310569,
0.9963260918039271,
0.9959311927938209,
0.17367923842809374,
0.6258200269322981,
0.6640397162184417,
0.808170573840998,
0.8496057123198117,
0.810095786605379,
0.9901094146799497,
0.9981446011773063,
0.998405186445489,
0.8462494329397424,
0.606070728466638,
0.8454970615061969,
0.9157546064171571,
0.9968702672552605,
0.9954851114337615,
0.6622247134631027,
0.7081979079178589,
0.654500641805325,
0.5956391787171642,
0.6241608579504916,
0.6472594305816632,
0.2621763458168426,
0.5715051555841986,
0.5149985103273704,
0.5644568293876413,
0.6632938863282389,
0.9960900121568927,
0.19972682244657958,
0.5846997513891496,
0.35464300578807073,
0.9748540307855267,
0.9748347642413825,
0.9778757913818841,
0.9715934512437103,
0.9949435107142942,
0.9829386961980081
]
}