GENIES/results/evaluations/cra/llama-13b-code_low_quality/code.json

{
    "model_dir": "models/llama-13b",
    "distribution_id": "code",
    "timestamp": "2023-10-16 03:06:05",
    "eval_score": -0.15555779368044137,
    "eval_brier_score": 0.15555779368044137,
    "eval_average_probability": 0.7438491323537898,
    "eval_accuracy": 0.776,
    "eval_probabilities": [
        0.957525543265661,
        0.9775514128088043,
        0.9894323986231703,
        0.9980985119612743,
        0.9987993936085547,
        0.9558579589985059,
        0.9579637140356517,
        0.9980119053912918,
        0.356774995337394,
        0.9839387274523808,
        0.9846377212211408,
        0.9876437787891876,
        0.9236207522882973,
        0.9977789724160006,
        0.520088006540207,
        0.6769838955253235,
        0.5754562265767147,
        0.6931958886291095,
        0.8722984538828376,
        0.9693055147957804,
        0.9460643761185886,
        0.6629994867917777,
        0.950626017038159,
        0.9620959382883058,
        0.006859564049804812,
        0.00931894724166473,
        0.011777865487000905,
        0.9350988335328237,
        0.9409316476216107,
        0.9984785000970559,
        0.9057809449784279,
        0.9936065213129577,
        0.9819402878476465,
        0.9052374447198538,
        0.6498973228360553,
        0.622890657391793,
        0.9541772590391723,
        0.9683945989435683,
        0.924499678863413,
        0.9997776251310714,
        0.9954365087283463,
        0.9865337276597237,
        0.9908181568006157,
        0.9242793267686011,
        0.8084266170755425,
        0.9998261235324999,
        0.9998852823294473,
        0.9998039810713389,
        0.9911242089469311,
        0.994155507329409,
        0.9926936523294027,
        0.999878441860878,
        0.9999914419651148,
        0.9999524258300142,
        0.5938480021670953,
        0.9821396603321951,
        0.7231521143830191,
        0.11331970897473652,
        0.19167291596338676,
        0.15093954636286414,
        0.98140355468934,
        0.975296746920177,
        0.9739896319099361,
        0.8053256173830994,
        0.40898434236293835,
        0.15288074046129962,
        0.9819018801136014,
        0.031177203573742318,
        0.09243143664001531,
        0.6619326592860095,
        0.9887437108841759,
        0.9936630440305271,
        0.7550261875935977,
        0.9588603916467283,
        0.9793550544454118,
        0.9999553956503889,
        0.9998326171396288,
        0.9998769275143241,
        0.9964386944490097,
        0.533615701045694,
        0.9662186489098611,
        0.999735087767776,
        0.8710944623067067,
        0.9891379541711803,
        0.9482101973565749,
        0.8996450229145163,
        0.9997056445532344,
        0.9970473209585623,
        0.9978691375659237,
        0.6409990622206242,
        0.13734216086128012,
        0.19855517596850358,
        0.1233294045702324,
        0.23335032458383637,
        0.8334146220570856,
        0.4207579302288609,
        0.9378399779442931,
        0.4658978369180773,
        0.47285621922238147,
        0.06642004937775825,
        0.17326862736108573,
        0.029436093183864753,
        0.9972661068386777,
        0.9973624875415251,
        0.99928608793062,
        0.9904580298774502,
        0.9821728887265362,
        0.9938147081322319,
        0.049540032328047046,
        0.20767002922950262,
        0.21008905387975055,
        0.998874459662017,
        0.9989551232961311,
        0.9989079101349256,
        0.45264402683427185,
        0.960292379004036,
        0.26449802273251155,
        0.9840633799067311,
        0.8241998598078428,
        0.6836013845578199,
        0.8967503635080453,
        0.9260856447126246,
        0.9241996630493494,
        0.9944120254751311,
        0.9998023211703501,
        0.09198453577784853,
        0.9751132564988434,
        0.9620814520640324,
        0.9373821307468206,
        0.9804819892143259,
        0.9992692865155349,
        0.7539680081163047,
        0.7490389633997997,
        0.9367449793723346,
        0.9645307234800846,
        0.9925665583840731,
        0.9937859360763551,
        0.4589911698218708,
        0.003596623890149004,
        0.6619783755435904,
        0.059218410080948636,
        0.7649672657023792,
        0.028910911109535448,
        0.5773126286367671,
        0.3457742502668927,
        0.24163277114129578,
        0.441963341157236,
        0.8180806861707988,
        0.6332157816047397,
        0.07478899867485787,
        0.9748979038757719,
        0.9801777477039169,
        0.9589213201429658,
        0.5507925053548962,
        0.9575379905458089,
        0.7905327390195814,
        0.9286288663429362,
        0.9357953683575797,
        0.9519696312281155,
        0.996273153730225,
        0.962296787455277,
        0.9952090118292353,
        0.9155856145393013,
        0.9760782901206685,
        0.9631203003419052,
        0.9863538473461356,
        0.9910128594346286,
        0.9839450027727383,
        0.5592541762464764,
        0.3855023316274766,
        0.8232462906004121,
        0.10938814113364337,
        0.27380575729145173,
        0.9499325097756626,
        0.9334573350852591,
        0.10978240411433512,
        0.9914715796836495,
        0.9750098826276268,
        0.9999913613886589,
        0.9999792424448052,
        0.2584079133384388,
        0.7615458481434104,
        0.9486576928158523,
        0.9608903179604649,
        0.9992253142431959,
        0.16101797100082527,
        0.8589853776613753,
        0.7846232196624708,
        0.988571894620683,
        0.638408552510811,
        0.7761025088569757,
        0.9996628059384144,
        0.9772195241305044,
        0.87995432689953,
        0.9742779806309795,
        0.9904788790949499,
        0.9743263224666748,
        0.9873499434259077,
        0.3136456745433357,
        0.503309053802665,
        0.5640690844592557,
        0.9786123013733449,
        0.9665428564436818,
        0.9763083352584292,
        0.6933863105618492,
        0.6922396977782004,
        0.98015208756803,
        0.15506916852585273,
        0.5163464885733676,
        0.6432071587907828,
        0.23153348959352543,
        0.896426920970818,
        0.1295720566341757,
        0.6721402554573392,
        0.7753536710785394,
        0.6287411096207689,
        0.9886716686516013,
        0.9831570346022781,
        0.9717790809691574,
        0.24572803613874494,
        0.27696438585162353,
        0.5681173401832565,
        0.2754435404412613,
        0.9046641406220689,
        0.2437110511153913,
        0.9948603080821607,
        0.855969581885044,
        0.7544046305091031,
        0.9776947443622928,
        0.9951558813102203,
        0.9943652291156988,
        0.8866971269282885,
        0.9671374102016855,
        0.9567419417202512,
        0.5980863053711982,
        0.46929307873334625,
        0.6108339500398732,
        0.8131495189530991,
        0.9250388796498957,
        0.4418477590717692,
        0.7203137562901472,
        0.7835559890037395,
        0.6360274156976169,
        0.9990603789428966,
        0.9868171436966058,
        0.9995206843239816,
        0.2032768147163171,
        0.35812775935113117,
        0.9986074423096597,
        0.9930036818637791,
        0.5909404443118821,
        0.9983969046481492,
        0.6143003475349618,
        0.2990720069441525,
        0.11139159986985003,
        0.9952215878641385,
        0.9860600727106226,
        0.9883295803627897,
        0.5914258926342609,
        0.9997317688677747,
        0.8753856731514236,
        0.9799753720106573,
        0.2991319341831621,
        0.3978407342041295,
        0.9977301424981693,
        0.339321240716957,
        0.9530884343432848,
        0.8485392898938234,
        0.8587864715397213,
        0.7834722135203194,
        0.52695538930419,
        0.9744021652403368,
        0.7916527557488902,
        0.8339011837997797,
        0.9033037523088625,
        0.9140731995749442,
        0.9949636850676331,
        0.9994011784746045,
        0.9512566864753189,
        0.5368489750454002,
        0.9979271161089533,
        0.2660664454249898,
        0.96501450295375,
        0.34017357065013776,
        0.34223545310542114,
        0.998423965455099,
        0.9995403795619909,
        0.9992179424493388,
        0.9444465148588932,
        0.8657807066651317,
        0.9833760386286096,
        0.960900964074849,
        0.9812495018842526,
        0.9534070087389785,
        0.5597643370124169,
        0.4575462397029746,
        0.968765610519126,
        0.42406657980665174,
        0.966183186810153,
        0.5103509839190699,
        0.6033033335726409,
        0.7568770080161867,
        0.9770898503830553,
        0.9436282755865627,
        0.9982336690856156,
        0.9964073492364464,
        0.8984636407715312,
        0.9235884447709266,
        0.983520009291084,
        0.038237594339125214,
        0.5325737681942426,
        0.6419920231552474,
        0.9930105983961562,
        0.9626288196326596,
        0.9056314944437321,
        0.18568876902603193,
        0.49311810910118137,
        0.9953563014840493,
        0.8902369376432238,
        0.8525084063056128,
        0.7103779298835519,
        0.38922017313863366,
        0.8056647790339917,
        0.28082274471837027,
        0.4356074795380559,
        0.9677039877420496,
        0.08816941883540515,
        0.6834949535797442,
        0.754643589748769,
        0.6269345967569225,
        0.1214214258770765,
        0.32401887553285236,
        0.05025527397193197,
        0.5603924593826844,
        0.45753744954235653,
        0.5308912048282792,
        0.23355642075004746,
        0.9951749653775274,
        0.3069793900459882,
        0.25703027046115856,
        0.14740946502769264,
        0.2943888993387434,
        0.9728348483516702,
        0.9543039028271939,
        0.5254327457217001,
        0.9993318235294882,
        0.3526038179098021,
        0.34208000605472766,
        0.7220748571032276,
        0.7151309590110346,
        0.6059973720109739,
        0.9113510095387564,
        0.39398631993307975,
        0.6134250291809079,
        0.9613435548760396,
        0.7570175625176199,
        0.9901878246315625,
        0.6269312005685674,
        0.9993828527379007,
        0.7903304487116465,
        0.9664937188018773,
        0.7444053673721868,
        0.8752296745968567,
        0.9998184621774934,
        0.9830740592019184,
        0.9999862737061114,
        0.9971356796443295,
        0.9992208193758059,
        0.9996497669129142,
        0.7476193347018603,
        0.2851362017565089,
        0.5393965105230315,
        0.9338223261900106,
        0.9216024736353255,
        0.9679337944730775,
        0.23164502550839805,
        0.009374419911455378,
        0.9489829526709007,
        0.4078016711116036,
        0.6044910413298256,
        0.25349264952014205,
        0.02117775607459561,
        0.936425362624087,
        0.8368988294339824,
        0.7041982114391588,
        0.23799518534310765,
        0.3491031147664414,
        0.9094154147761514,
        0.9464454436854618,
        0.8033427525580137,
        0.9883039708539794,
        0.9895025715137215,
        0.9945534490632247,
        0.9883075231716311,
        0.9147918658499045,
        0.208080748515297,
        0.8614735781120504,
        0.7545447452533868,
        0.8353760495953935,
        0.09214153918323753,
        0.5039493542622381,
        0.3589689209103293,
        0.5450783075544936,
        0.25274360007148355,
        0.7315412678647547,
        0.9808116344479854,
        0.9851938525770905,
        0.9639618451206162,
        0.9895438505991426,
        0.36893266793741947,
        0.9856302731070142,
        0.5294178012610744,
        0.6612426768453813,
        0.9178436426476674,
        0.9697817144042025,
        0.5740084767328375,
        0.9764844618428334,
        0.15444706378750073,
        0.6340193115646158,
        0.7476711202453368,
        0.9776275621848168,
        0.959450426768521,
        0.95988311508446,
        0.002210564603883089,
        0.4036772017453957,
        0.7793702388612532,
        0.5715681201112697,
        0.24815307858528046,
        0.5595684201234279,
        0.9962128792258853,
        0.7204268826364713,
        0.895712442084364,
        0.5783331643285874,
        0.3073897753052748,
        0.3882998505243981,
        0.18816130761059224,
        0.9978318677917524,
        0.8145653768457928,
        0.987329484349352,
        0.9399533328058252,
        0.9723357985191305,
        0.9994153213233017,
        0.9990054713898284,
        0.9579025762224541,
        0.9663138055194676,
        0.15474271910926357,
        0.7940683379615782,
        0.0007346622244459589,
        0.9945308779371581,
        0.00040630244847217334,
        0.8898162543426001,
        0.9785518184113221,
        0.9805077399578185,
        0.23183418567735992,
        0.5839723070019627,
        0.9858282334746712,
        0.9644005037634623,
        0.03925742079620212,
        0.1984251823314117,
        0.3242133898392544,
        0.7424482470533355,
        0.8865267356452224,
        0.9428397648308006,
        0.9816556651385326,
        0.9576523995235602,
        0.8730912872088867,
        0.7970404206326455,
        0.910579128852373,
        0.8221714494613198,
        0.9813684227494982,
        0.9841373688712876,
        0.9868084113380979,
        0.7365963732728102,
        0.995330039268331,
        0.39362080368962576,
        0.4556869395819604,
        0.6151699805928141,
        0.8859791533318784,
        0.7662819946618648,
        0.9082763367888025,
        0.7672318415269436,
        0.6084083778415438,
        0.9443402374308518,
        0.22966669974461573,
        0.8817028375781458,
        0.38415650685124036,
        0.9709781240830144,
        0.34677294641072975,
        0.4312291910351447,
        0.998259685371283,
        0.022932358447370516,
        0.9250451943605994,
        0.9992948640041778,
        0.26654984762828127,
        0.9854518007056393,
        0.9830756898109585,
        0.9551936275293049,
        0.9992996480816676,
        0.8881357718030604,
        0.9985058188426444,
        0.9733121177577057,
        0.5988022805323157,
        0.8921909418643559,
        0.5192723386102779,
        0.8047009051349756,
        0.18189061243997962,
        0.545644690413353,
        0.5365910776905539,
        0.17806760493085771,
        0.9250211186590799,
        0.9250041856285289,
        0.9869589323317645,
        0.9963258479613314,
        0.7722268881993084,
        0.9875893268178625,
        0.9951633965744856,
        0.9997821210405449,
        0.9997398745252235,
        0.9997668817916505,
        0.924974125672699,
        0.9785206162106992,
        0.9803668745794637,
        0.8419879193856864,
        0.7364325241717944,
        0.8300502473629063,
        0.2938999118213932,
        0.8134778230022676,
        0.4686144277582994,
        0.9765160994187434,
        0.94541021754466,
        0.94305384041408,
        0.9198161476916147,
        0.9596019922548433,
        0.9640695197371368,
        0.9697875511436719,
        0.9905564967222443,
        0.9888368369603742,
        0.5934122823383753,
        0.7224568392134573,
        0.5739657946428915,
        0.9978680238247059,
        0.958076518132773,
        0.9946822846623873,
        0.998225039495415,
        0.9939426592880993,
        0.9999228637482878,
        0.9972134625015142,
        0.9964089448657869,
        0.999088022667884,
        0.3422331752483474,
        0.6317823498645893,
        0.9570056624687022,
        0.7486492993779299,
        0.42295423153729284,
        0.47980670073843024,
        0.750837057056435,
        0.8477404559471942,
        0.734556998022558,
        0.970891635845951,
        0.9518984758921566,
        0.8252018408423925,
        0.368289449614833,
        0.7106114386601632,
        0.5281464197321827,
        0.7748600047662884,
        0.9979960683677941,
        0.8172483096872025,
        0.36323576021130544,
        0.8685878862260653,
        0.8799056191415711,
        0.9999005093997431,
        0.9929393602383828,
        0.976805409414962,
        0.7712560930763539,
        0.9922641044615274,
        0.024135336933481022,
        0.9999538104126384,
        0.9998745543812134,
        0.9997295482539397,
        0.9867396158347557,
        0.5273723549081655,
        0.9441238708733727,
        0.6682801366039682,
        0.877724300555524,
        0.8601038512297849,
        0.9694660983298139,
        0.9734331694966938,
        0.921118677864668,
        0.6105926358423981,
        0.9969455039643241,
        0.1818980434434746,
        0.3074343391922261,
        0.595832491047919,
        0.10114355897140692,
        0.9999691583367404,
        0.9979920155930757,
        0.7356320065098148,
        0.8943380776939907,
        0.9911275077080217,
        0.966215324469739,
        0.9185565320966369,
        0.9722448454507929,
        0.9835873941604879,
        0.45524586905653175,
        0.4676096430658705,
        0.19859543466153284,
        0.11515675583351226,
        0.13908033563154198,
        0.5540313902215646,
        0.9927570178177608,
        0.9896374332182183,
        0.9837160355259813,
        0.983433640980822,
        0.4948175128068256,
        0.05954707516966096,
        0.7387270989985008,
        0.8155040000606175,
        0.7800702104081327,
        0.4309005239318826,
        0.947235164060908,
        0.9827241151659112,
        0.9052602799282894,
        0.6850835234022152,
        0.795289582548123,
        0.9305525257830558,
        0.9701363395788609,
        0.9863643817501331,
        0.9207608455761289,
        0.9527989130089173,
        0.9812454521248327,
        0.9865582971846766,
        0.88907977993227,
        0.992415894762326,
        0.9988228112925885,
        0.9999598510160981,
        0.9998887264810689,
        0.9980127798651828,
        0.5504951835273114,
        0.9530938337185075,
        0.3619927294770905,
        0.565785775031583,
        0.3931446626661026,
        0.3817513757516982,
        0.8927916284536906,
        0.32838453645503807,
        0.9985702567578363,
        0.995130946202301,
        0.9404522533241401,
        0.9996384636139141,
        0.997897683588757,
        0.9239591365563619,
        0.996832388845705,
        0.9952461060038957,
        0.9960615840919025,
        0.9889490272095904,
        0.3510882962860633,
        0.9766949083188853,
        0.8655529387715873,
        0.9885673495914419,
        0.9939537522374111,
        0.9954233865122452,
        0.9969485242121281,
        0.9985615560573254,
        0.9567641320129514,
        0.8283538337436108,
        0.9999010973556025,
        0.9596773618671265,
        0.946208349011161,
        0.9422475315048512,
        0.6958069283798038,
        0.4254783483919218,
        0.21041470119663952,
        0.04353509190230387,
        0.4788350577845449,
        0.049180271450208084,
        0.25909069837845927,
        0.8078358958983547,
        0.3215099530615403,
        0.93975409461596,
        0.9542247124961253,
        0.9992272645977183,
        0.9932771583108366,
        0.9997894689063566,
        0.9999407062856387,
        0.9999806083714413,
        0.9995241401253062,
        0.17084770509752334,
        0.9809390971467924,
        0.6708105791977603,
        0.9781047169139926,
        0.11148165053950954,
        0.6323377154193576,
        0.09029787981434101,
        0.4351908255011422,
        0.9844610146579631,
        0.988839621795098,
        0.9570462901900756,
        0.40236208532783047,
        0.7572384100661289,
        0.9942869487484391,
        0.9957844826682443,
        0.9925517239312281,
        0.40118906536350984,
        0.6840511041853651,
        0.841127448183663,
        0.9980942871365243,
        0.7157678841542893,
        0.9132709370050509,
        0.9885334130538653,
        0.9813443552048857,
        0.9827987717653366,
        0.5843983155216749,
        0.3822600540662644,
        0.04270861910545265,
        0.9261735211532094,
        0.9870830988162362,
        0.9749572823102289,
        0.7743824709734697,
        0.21041207020752448,
        0.376745746238936,
        0.851289436764644,
        0.6405100727916553,
        0.7036091343848216,
        0.8667094748286214,
        0.6458585946723027,
        0.027776392507562014,
        0.979406816676776,
        0.9593523861751669,
        0.9840047967752223,
        0.9964105399699208,
        0.9891597917530136,
        0.3867621512273398,
        0.8174036338958555,
        0.9522785864723523,
        0.8288125775986103,
        0.34712730436527695,
        0.845264725550277,
        0.43906106554007585,
        0.868736953408419,
        0.9850760958060545,
        0.9742920507238666,
        0.2332231790619465,
        0.17719152986797881,
        0.1534871632963538,
        0.8836109032286319,
        0.6344996312649251,
        0.13336594198571405,
        0.9949636947015716,
        0.9526636851256515,
        0.9974665977740312
    ],
    "eval_rms_calibration_error": 0.31200277690443273
}