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

{
    "model_dir": "models/llama-7b",
    "distribution_id": "code_easy",
    "timestamp": "2023-10-14 00:07:27",
    "eval_score": -0.06119980652270272,
    "eval_brier_score": 0.06119980652270272,
    "eval_average_probability": 0.889393005344134,
    "eval_accuracy": 0.916,
    "eval_probabilities": [
        0.8035319750689573,
        0.9510974842424774,
        0.9987359918828378,
        0.8966932610367933,
        0.987810883313412,
        0.8482698082869884,
        0.9009292003877739,
        0.9235260293237343,
        0.9416541131207187,
        0.8118888473335669,
        0.897871453351501,
        0.9124904770682803,
        0.7824555916669416,
        0.9773307676362913,
        0.9467962785425764,
        0.8769602693365404,
        0.9987813233267072,
        0.9202366332746376,
        0.8618801442742052,
        0.6711048426954268,
        0.9158991425810661,
        0.9492476634055206,
        0.972044559825799,
        0.8092438720257231,
        0.8599422553567659,
        0.9838161595097002,
        0.7040041141112828,
        0.9832142126636142,
        0.9912000688457085,
        0.9984592183770268,
        0.933277319346452,
        0.9052295312351643,
        0.8538196270859493,
        0.9252062599708749,
        0.9790784038111641,
        0.9425013058319948,
        0.7726090364108184,
        0.9912293791934914,
        0.303188716768819,
        0.9876860870428744,
        0.9877625566126457,
        0.9468083610337442,
        0.9981848216864104,
        0.9934177669742938,
        0.9716003518916535,
        0.9208121021086262,
        0.08568169651317892,
        0.7214759532312682,
        0.9973587857838941,
        0.9990433153813794,
        0.9976246453252833,
        0.059701827633597104,
        0.9838550001090767,
        0.9929709631923683,
        0.9993214814824539,
        0.9999061552295943,
        0.9996678122795658,
        0.9888242260270143,
        0.5779599759861325,
        0.965221194687065,
        0.9950089338648267,
        0.9994300636489145,
        0.9987813536257675,
        0.9998038997721964,
        0.9987247255287315,
        0.9997843884115922,
        0.9998478012706229,
        0.9999001130585631,
        0.9994976735746416,
        0.999635084189639,
        0.0008797788340035861,
        0.007643777252328505,
        0.9879378669528875,
        0.9985330959369475,
        0.855812469719719,
        0.9956387823987776,
        0.9913423298904978,
        0.9986079273486161,
        0.9755511863786165,
        0.8985702435082895,
        0.9659969672416104,
        0.9679706319861325,
        0.9818635413102519,
        0.9250541108044695,
        0.982809370150843,
        0.9998033272168318,
        0.9986431565102668,
        0.7266168210977436,
        0.3258381737968741,
        0.3788875650275978,
        0.9779229863613461,
        0.42704065648501016,
        0.6591884847588446,
        0.9984952563595594,
        0.9982900040954228,
        0.998062091868761,
        0.9999145555542291,
        0.9993094892852754,
        0.9997706044081003,
        0.9942671965287095,
        0.9651447075868056,
        0.9974176349818803,
        0.9979447913556627,
        0.998489528710076,
        0.9941689381633303,
        0.9967488782842063,
        0.9993074988052607,
        0.9994660955676005,
        0.9981105792031112,
        0.9923061647453514,
        0.9960756275266454,
        0.9990292023771009,
        0.9961515971310245,
        0.9990975709629352,
        0.9514413139755697,
        0.11158315893790127,
        0.9896693954260138,
        0.5526919269769317,
        0.9979235049129436,
        0.997043931436191,
        0.9740321075402749,
        0.9935598609654899,
        0.9818533086241696,
        0.9998796617696295,
        0.999197721561333,
        0.9994623021031781,
        0.9923220502609393,
        0.9991907571611814,
        0.9787728149511365,
        0.9541449853534544,
        0.7504715834472363,
        0.9952127666700187,
        0.9996251736461544,
        0.9990305103652733,
        0.9998517549273214,
        0.4983593156469309,
        0.9895037780750312,
        0.17232715422973674,
        0.9749119236029634,
        0.9949831448235175,
        0.9738137756989442,
        0.9911477892594885,
        0.9988657056387049,
        0.9986875225512869,
        0.9999652678117263,
        0.9994082353997727,
        0.9975975515657612,
        0.9902961911581276,
        0.999919493336635,
        0.9998028343238547,
        0.9120222299254628,
        0.7387966248025283,
        0.9346715451454974,
        0.9627654636852968,
        0.9994606119566477,
        0.8237923436492416,
        0.9997293377309097,
        0.9996012693543173,
        0.9998986393260216,
        0.9846787403946878,
        0.8822719071034276,
        0.9945588046185324,
        0.991948918692437,
        0.9548122997157954,
        0.9772974955862423,
        0.4196265782205626,
        0.9057523443108858,
        0.4907960170643033,
        0.8976943314578852,
        0.9273368305394238,
        0.9091239759196426,
        0.9896058682392759,
        0.994104775739155,
        0.7800337631159862,
        0.8965083925107766,
        0.4806379114756206,
        0.9496635885529475,
        0.9921193659468857,
        0.9993939580830393,
        0.9993480607949607,
        0.42008479323492554,
        0.9993032716413763,
        0.962229694903304,
        0.9787850578112209,
        0.9957797451093026,
        0.999311661308077,
        0.9687995402547488,
        0.6744225202707992,
        0.9657604008605004,
        0.9705544129850102,
        0.9878929477328307,
        0.9699333755805716,
        0.9996488361621024,
        0.7720641283359486,
        0.9983559014300187,
        0.9996728760140852,
        0.9644321207678694,
        0.618916115514881,
        0.9992300906876683,
        0.9913570592189148,
        0.85456395412212,
        0.5660609473890226,
        0.01849573692508255,
        0.7679593866444928,
        0.9577666031316351,
        0.358691594594829,
        0.8294138140937929,
        0.9656193645159342,
        0.9968212787391988,
        0.9983211895800745,
        0.6603507468259595,
        0.9879570698577793,
        0.9393351364323249,
        0.9965083204359596,
        0.9465678623582887,
        0.9916311231712952,
        0.9104441847900262,
        0.991836463155532,
        0.9146800529631027,
        0.9998170988093035,
        0.9947962818913296,
        0.9983513466485152,
        0.9349336756724124,
        0.996638622495581,
        0.9597007235557609,
        0.9987916999796419,
        0.9992827941451843,
        0.9996184953520074,
        0.909134111469736,
        0.9742513473520692,
        0.9937401947810517,
        0.9952996631763047,
        0.9990808348957999,
        0.99793850018075,
        0.9888954841568194,
        0.9611989532807229,
        0.8124728085448467,
        0.9689031070694119,
        0.9677657585130715,
        0.9864055696789764,
        0.9784429113819034,
        0.6372011063902353,
        0.9970434173673657,
        0.9949608543541225,
        0.6854987822060861,
        0.9967182694531302,
        0.770457454830771,
        0.9987476237355185,
        0.9995384801133227,
        0.999454942275879,
        0.994295908678177,
        0.998868984630388,
        0.9777202164828213,
        0.9998645447913996,
        0.835676531124969,
        0.99945004633528,
        0.9835724070427273,
        0.5115596035437462,
        0.9627416812576749,
        0.3329876394836596,
        0.8217467736393996,
        0.9986760598675702,
        0.9996426479724714,
        0.9981799249234207,
        0.9848984765623303,
        0.9597741654738295,
        0.7944151069158757,
        0.8137903100808559,
        0.9553753999763211,
        0.9694083498947031,
        0.9999200251778717,
        0.9995810236563034,
        0.7766938711806106,
        0.6498617311521193,
        0.984461622284467,
        0.4882502851128866,
        0.9987619924419279,
        0.9977301567156424,
        0.9783391047113258,
        0.9910413104240352,
        0.9981481836459078,
        0.9520523070122124,
        0.19225708659894336,
        0.20106574861523419,
        0.7594343946671264,
        0.9869959685967409,
        0.9610754269089574,
        0.9772920902934943,
        0.9998899157218325,
        0.9999200915526012,
        0.9982841895285889,
        0.9785700555558341,
        0.9810428739247299,
        0.5138638147891537,
        0.9993325567175255,
        0.9969976297730485,
        0.9920997331624495,
        0.9996004531157998,
        0.9447241141267423,
        0.9987637125092281,
        0.5081599455847378,
        0.998010609255927,
        0.9730390568076611,
        0.99479934605899,
        0.9937729034907675,
        0.9968064895727543,
        0.9996483927735215,
        0.9999603084750474,
        0.9999466170263652,
        0.9970924350544387,
        0.9940826382844253,
        0.9979813197362094,
        0.9997600442273753,
        0.9899002035546464,
        0.998392107003698,
        0.9829860420402383,
        0.9869875011268955,
        0.9989036577673392,
        0.993895777196183,
        0.9997026557608162,
        0.997194540841138,
        0.9997105506586086,
        0.9961971259539831,
        0.9998780924059757,
        0.9922291603033884,
        0.9330699812498342,
        0.9607148450783324,
        0.9995140293235703,
        0.9318460324447286,
        0.969164199052952,
        0.9998576773928319,
        0.9990827889515901,
        0.9981277462162136,
        0.622326217674095,
        0.9596729953652267,
        0.909602452080124,
        0.8708102624833343,
        0.9609511781757208,
        0.8465150431519974,
        0.9998351282691518,
        0.999271159295737,
        0.9140542635775107,
        0.9939137443325535,
        0.9871734533996781,
        0.8435944742859769,
        0.8223125459374223,
        0.9779723182402297,
        0.9830596309359743,
        0.8657888117937943,
        0.999515826739676,
        0.1261624502806069,
        0.12862581608285772,
        0.9829580937025839,
        0.9989859292246245,
        0.9988424277188737,
        0.9954045336300132,
        0.8986172300802344,
        0.9807885123447568,
        0.9904977678480311,
        0.9994211601503407,
        0.7207781480933474,
        0.9847009064034276,
        0.8535740282719945,
        0.19897686876811305,
        0.9975442062387863,
        0.997623651093761,
        0.9932578200214777,
        0.9965477478406827,
        0.9990123109659845,
        0.991759007382307,
        0.6660876054537161,
        0.9085450084239076,
        0.9997600442273753,
        0.998071646036369,
        0.9915671513440673,
        0.9980695902508544,
        0.9993504069299315,
        0.9799311459432611,
        0.9437120045905214,
        0.8201931562222101,
        0.9901093580395475,
        0.981217934855867,
        0.9991155465888247,
        0.9926159311194331,
        0.9960937603827298,
        0.993287268996168,
        0.9969147993059159,
        0.996520135732592,
        0.9366285405338445,
        0.9976583114593223,
        0.9975696922554839,
        0.9867634942639204,
        0.9643004169197932,
        0.8925900347988629,
        0.9992851085105623,
        0.835676531124969,
        0.778937478577476,
        0.09093039149105721,
        0.6612187886879781,
        0.7686794325485586,
        0.9980313641013364,
        0.9988365473162346,
        0.9996661014754034,
        0.9999581654809506,
        0.9998666078813548,
        0.9903307900474524,
        0.9941878674033855,
        0.9911135758086022,
        0.034467869246129434,
        0.9965856825435029,
        0.9998840206310422,
        0.9640730617012008,
        0.9996884639793238,
        0.9947543830291148,
        0.9988674366673597,
        0.9405850949254069,
        0.9989323861777764,
        0.9957893584077366,
        0.9993538123041277,
        0.9968470845512623,
        0.9808758909485243,
        0.9994776477257867,
        0.9602801767159442,
        0.04597403810708134,
        0.028554040404201286,
        0.1811604827773975,
        0.9999253387523096,
        0.9999044732053247,
        0.9992250776484307,
        0.9843715007343589,
        0.9996037824945013,
        0.9944591558828074,
        0.2578748527570949,
        0.709630403850349,
        0.9544197399948642,
        0.9838821030493418,
        0.9996474747064249,
        0.991871739930298,
        0.982717294610019,
        0.9986291363358526,
        0.9984044728144987,
        0.999799447743195,
        0.9999738759548863,
        0.9999788157983458,
        0.9829218967269728,
        0.975172921621305,
        0.9774951077508125,
        0.9971082631499742,
        0.9990863748199975,
        0.9991337172657746,
        0.9986208315784063,
        0.9998444962375023,
        0.9944694687225974,
        0.9972730626231546,
        0.9702115825291265,
        0.9348379928769742,
        0.9872001518776543,
        0.9813071841807205,
        0.9846519707565603,
        0.976939319015279,
        0.9929811322733217,
        0.8420897186761692,
        0.9850223752812466,
        0.9995622519591973,
        0.99987298752688,
        0.9996778768985685,
        0.9999004948510718,
        0.9986740649776873,
        0.9896714782466929,
        0.9992250253284757,
        0.9978630339507256,
        0.49124857260532856,
        0.9947899830840249,
        0.9989684635599703,
        0.9939199804987463,
        0.9909479123858203,
        0.9983470343139074,
        0.9967355579337608,
        0.9998460618731394,
        0.9994508988706127,
        0.7407118682204752,
        0.9492802475142881,
        0.8730825016333478,
        0.23550387205027445,
        0.9394325440527675,
        0.8923888722400762,
        0.9995779864272112,
        0.37151144503219424,
        0.05832811827831108,
        0.9872324873779781,
        0.986052024949536,
        0.10468962173975249,
        0.3743090912355708,
        0.9199803125856229,
        0.10384585414820524,
        0.8263671890233449,
        0.9992749638181118,
        0.7457112186890138,
        0.8853341148215527,
        0.5560066387394722,
        0.9834108873860973,
        0.9303673287909123,
        0.9996657748935949,
        0.9898178676223656,
        0.35053874451192724,
        0.8386923119898102,
        0.05090207392237989,
        0.9974871090891897,
        0.9963928258477176,
        0.8674976945508611,
        0.9305306061446841,
        0.9832784164018679,
        0.47353294803882495,
        0.9983343337373591,
        0.7917581800870345,
        0.9996091076368192,
        0.09788483675173065,
        0.9929139705976834,
        0.9969739747127181,
        0.9929547119815197,
        0.3328648305758628,
        0.9783318877739904,
        0.9997124266349867,
        0.9857712776487082,
        0.7949199556224678,
        0.988459130785929,
        0.995949430513032,
        0.9915761254230078,
        0.9916376857127612,
        0.9706163434429386,
        0.9959388293692821,
        0.9995693578584872,
        0.9237408547843686,
        0.8040078628819737,
        0.8031769280216963,
        0.9980947126760282,
        0.9451363588374123,
        0.9978943222420334,
        0.9972463960727588,
        0.9981429553217301,
        0.9989646710330231,
        0.9965841817776865,
        0.9946272292685873,
        0.9962460502051005,
        0.9988112665382115,
        0.9969143098129609,
        0.9981772514896547,
        0.996861853666567,
        0.8900026229945829,
        0.983259610493534,
        0.9975701793207354,
        0.9993622977518355,
        0.926489083297552,
        0.9937478883027395,
        0.19160545067401502,
        0.9769681264810355,
        0.9996445527751826,
        0.9987084291862834,
        0.9952543751358786,
        0.9988123490566548,
        0.9992601197455998,
        0.9961445562222992,
        0.9834491913583985,
        0.9994731232569176,
        0.7073219587080919,
        0.9990917843082722,
        0.9901132939339784,
        0.38605626480160216,
        0.9948102538876937,
        0.9935746572178126,
        0.9998901026668966,
        0.9998232568047558,
        0.9977125097717442,
        0.994135015334595,
        0.7193249466385182,
        0.9106823834148512,
        0.07181731873248776,
        0.9982595663660586,
        0.23528926772901923,
        0.9997000620701845,
        0.9982409930651729,
        0.8613458811720451,
        0.9837574979547387,
        0.9993401672449624,
        0.9965600403032892,
        0.998163746916138,
        0.9999326502352283,
        0.9998018849836436,
        0.03663666376210217,
        0.9996027798257806,
        0.9932152388585442,
        0.789708926534119,
        0.9980149205952691,
        0.9969387766954787,
        0.9986150450570552,
        0.9983131200059389,
        0.9993005490058258,
        0.987338770510832,
        0.983276662217826,
        0.9781101338277295,
        0.9795180034088011,
        0.9974885103803008,
        0.9977151314209284,
        0.9970561633166439,
        0.9751726596268053,
        0.9320998494296627,
        0.9994443288593209,
        0.9876168439859957,
        0.019768124197885394,
        0.9998144902351743,
        0.9974527146251181,
        0.9973105197877136,
        0.9999035698460838,
        0.9992009419750458,
        0.9997155466844511,
        0.4926091417975249,
        0.09650112554791368,
        0.015364052556031349,
        0.8591641299557929,
        0.10848195745710043,
        0.9858226861941364,
        0.4092477314166999,
        0.914261117910506,
        0.9838711332699531,
        0.9984966611463372,
        0.9998223915993177,
        0.9380117898483487,
        0.547417623520414,
        0.23133669201135745,
        0.9339301376731463,
        0.9764988330603893,
        0.9907961016324539,
        0.9666591540726205,
        0.9986330191682629,
        0.9998258576404587,
        0.9208023240268409,
        0.9462958039963087,
        0.9853683951925177,
        0.5836549972155239,
        0.9934581206035343,
        0.991185052386386,
        0.981110349740763,
        0.9997356391062342,
        0.999102144158484,
        0.9977803792993744,
        0.9876124506454488,
        0.9989476098688007,
        0.9119171820775166,
        0.9915058836432269,
        0.7778372599931137,
        0.9690732843209126,
        0.9844079742784886,
        0.9496714076305472,
        0.99378914518571,
        0.9956774580541562,
        0.9774875344612891,
        0.9792356770576359,
        0.891993819933602,
        0.9331941194061412,
        0.633218439666799,
        0.9895033087432707,
        0.980049653695799,
        0.997680323089263,
        0.9884198160052636,
        0.966707977537431,
        0.9992942960587193,
        0.9907173905581625,
        0.9842411433119708,
        0.9961272990469483,
        0.5474002608913466,
        0.6432821373182344,
        0.6822143563190023,
        0.8362475691386051,
        0.9827991233050786,
        0.9877837946650839,
        0.9963672376848417,
        0.03157023710378637,
        0.9959593249912838,
        0.7679123315076328,
        0.9692759629541154,
        0.9916079677426226,
        0.999309333553816,
        0.9999380505963915,
        0.9994927089768723,
        0.9428636911704955,
        0.9987697377563016,
        0.9971306777261966,
        0.9790913659954777,
        0.9610583894027956,
        0.761509821168756,
        0.9905084054315583,
        0.9987679534846324,
        0.9919953628256898,
        0.8144557165095434,
        0.9882181017589001,
        0.9953048127886712,
        0.9946518405438016,
        0.9993004843772756,
        0.9999428102988768,
        0.8996823464396884,
        0.46345579546350807,
        0.4692201964124804,
        0.9930447150413425,
        0.9974516930316034,
        0.8575034591384884,
        0.46540689069050334,
        0.8844441072527454,
        0.6021346423032439,
        0.9943871996140293,
        0.9874940374602894,
        0.9503419178572973,
        0.9974515122448634,
        0.9978706886364528,
        0.9747378856828007,
        0.9997353127760094,
        0.970312010025819,
        0.9978955268058017,
        0.931365449441169,
        0.9708327110625506,
        0.894906714997536,
        0.8067918389377242,
        0.9815933981564006,
        0.9674870430888729,
        0.39225076822360216,
        0.6124332881346816,
        0.9858724484807697,
        0.9876939327054096,
        0.9439393302809382,
        0.9284879940163149,
        0.9939869807260286,
        0.9983875796210968,
        0.16243748389997703,
        0.6962708362914961,
        0.25823159555623987,
        0.9824454792802729,
        0.999346251131798,
        0.9969003747457138,
        0.9836528420841923,
        0.756860439854592,
        0.8901572748873126,
        0.6285379935748188,
        0.9421793013098827,
        0.9706642779675856,
        0.6817817229365504,
        0.3198801559277201,
        0.37842739549330195,
        0.6047622367188306,
        0.999975636876636,
        0.9994625672844992,
        0.999901599588663
    ]
}