GENIES/results/evaluations/mms/llama-30b-code_easy/code_easy.json

{
    "model_dir": "models/llama-30b",
    "distribution_id": "code_easy",
    "timestamp": "2023-10-14 01:55:10",
    "eval_score": -0.02865917112026801,
    "eval_brier_score": 0.02865917112026801,
    "eval_average_probability": 0.9451345172261861,
    "eval_accuracy": 0.964,
    "eval_probabilities": [
        0.9941251070088768,
        0.9998782930138543,
        0.9999644346190635,
        0.9952034536003003,
        0.9452576047945702,
        0.9671389226836021,
        0.4472944073016495,
        0.9177381000112397,
        0.9151508842062919,
        0.9709900170156677,
        0.9913027779971973,
        0.9979919090082087,
        0.9068127665808726,
        0.9405641669735623,
        0.9838414468379748,
        0.8676756235655383,
        0.9992744217592806,
        0.9667510827041229,
        0.9989591217654025,
        0.9888953329598823,
        0.9833715862787115,
        0.9985608344948599,
        0.9968187313198061,
        0.9862058932343172,
        0.9997678698204664,
        0.9999541669532033,
        0.8682086318449258,
        0.9971375927569198,
        0.9995636368908412,
        0.9986208693496472,
        0.9905321567521224,
        0.9994692000900136,
        0.9975527299158193,
        0.9936195957296355,
        0.9966287416311665,
        0.9409130048826425,
        0.5592740952092125,
        0.9989715725595354,
        0.7892942258211941,
        0.9982170917290455,
        0.9999337154876548,
        0.9841795425870883,
        0.9816268391056362,
        0.9982571397587513,
        0.999076047963725,
        0.998223649037116,
        0.9949403174787905,
        0.9868743201066974,
        0.9973590499247478,
        0.9998006242125905,
        0.9981076685044252,
        0.788408147494111,
        0.987790832609516,
        0.9953835065648892,
        0.999504953008233,
        0.9995223357482089,
        0.998865454818908,
        0.9800845924217619,
        0.5429713832593512,
        0.9676762028246986,
        0.9996570315572418,
        0.9992458517536317,
        0.9986001498520113,
        0.9999512161672164,
        0.999943154600439,
        0.9998107636298277,
        0.9984531992156594,
        0.9996268559245929,
        0.9993860121542962,
        0.9852202810417634,
        0.006142089927512322,
        0.06222853418114191,
        0.9997460240538137,
        0.9999569291144972,
        0.9841513760317792,
        0.9878024268973705,
        0.9948150219002745,
        0.99970087276829,
        0.9992572584175855,
        0.9992925824984208,
        0.9944079463329015,
        0.9890855868195332,
        0.9948364970450424,
        0.9918927426136529,
        0.9996634604874335,
        0.9994799638549863,
        0.9991765439734593,
        0.9951929429554748,
        0.9665681534387099,
        0.9802349921909379,
        0.9988452499102837,
        0.8642116840731691,
        0.9885457281967907,
        0.9910021870929365,
        0.9969328276126532,
        0.9929001484240765,
        0.9997759134691664,
        0.9996547144790198,
        0.9999750842397438,
        0.9988537153380614,
        0.8309566062703466,
        0.9849928249687196,
        0.9994855512107007,
        0.9999091636101338,
        0.9994426250583327,
        0.9994026553837771,
        0.9977135425534998,
        0.9981115928205546,
        0.9999492125746715,
        0.9999749144367309,
        0.9965098875165866,
        0.9994567383521692,
        0.9999512145492826,
        0.9968691693993417,
        0.9982394204430673,
        0.059959908748747015,
        0.9978528249287312,
        0.6761469987901594,
        0.9997140669560974,
        0.9987737031698731,
        0.9951147879767964,
        0.9951754748292622,
        0.9919782079985751,
        0.999854028626746,
        0.9997693192274353,
        0.9998275245640906,
        0.9815136978655475,
        0.9999485883071564,
        0.9993601624762124,
        0.9764797879452703,
        0.9886681179581243,
        0.999293336787712,
        0.9999758406738264,
        0.9997508061143223,
        0.9999410003262208,
        0.2691714744469502,
        0.9987156693786367,
        0.9920729801224184,
        0.9913908114112958,
        0.9946153293814544,
        0.9997943534107341,
        0.997645667220726,
        0.9996302490653606,
        0.9997095825287515,
        0.9999406126023221,
        0.9996591519293625,
        0.9999188373705608,
        0.9994769967084627,
        0.999964828067505,
        0.9999593804345943,
        0.9555698924879932,
        0.5290355839332587,
        0.9969370303539653,
        0.941144370263923,
        0.999950804943027,
        0.9988415251616326,
        0.9998417433620979,
        0.9932337054444507,
        0.9999579119268284,
        0.9977144940822151,
        0.9969685885068625,
        0.9979115265437544,
        0.6668373795582615,
        0.8006017807631732,
        0.9677982077621128,
        0.9969265918303565,
        0.9991383586602169,
        0.9045021248519357,
        0.9975889108817351,
        0.903334992821203,
        0.9996582365722823,
        0.9878410868964015,
        0.999487324869231,
        0.9939486742810631,
        0.900206780448482,
        0.9985160802095397,
        0.9688260437383306,
        0.9999259368473176,
        0.9998335412396399,
        0.9997863821184946,
        0.9135912601864876,
        0.9995258854110064,
        0.9988703543215063,
        0.9341979090158512,
        0.9354332524625906,
        0.9999374925884869,
        0.9812026622588579,
        0.9123794694455868,
        0.9852991385022841,
        0.9990909271381115,
        0.9926757365998973,
        0.995986302444172,
        0.9999338808459004,
        0.9990226636947328,
        0.9996086534552979,
        0.9997483137686688,
        0.9583308841673768,
        0.989153180667094,
        0.9999358938279145,
        0.9990125181532132,
        0.9989481650778853,
        0.9940068813808871,
        0.8468292522685306,
        0.9760980380665041,
        0.978777911269421,
        0.3873035491980188,
        0.9800474199702679,
        0.9925751992693002,
        0.9982524303121109,
        0.9964184458508599,
        0.9594548817208114,
        0.9373704467427645,
        0.9149662246934007,
        0.9996682559276937,
        0.9997440904622491,
        0.9997939071695543,
        0.9908769808240923,
        0.9932802200272252,
        0.99675983302116,
        0.9998791787784346,
        0.9954258606047012,
        0.9993619274277845,
        0.9045890904236666,
        0.999828151953335,
        0.5383974044270385,
        0.9970733473800074,
        0.9995798248196008,
        0.9995177707301016,
        0.9663739966948454,
        0.9272466522034183,
        0.9994788717532084,
        0.9978816529790522,
        0.9990329075045682,
        0.9922743173294859,
        0.4879167698317435,
        0.8830061738741816,
        0.9183048306901551,
        0.9961828256369822,
        0.9950533365358969,
        0.9982813978134821,
        0.9990035113801174,
        0.9541190676984128,
        0.9985091956741322,
        0.9998512972154107,
        0.9826078031839885,
        0.999543828702419,
        0.9993924134165889,
        0.9998244531351317,
        0.9997767536861901,
        0.999396467873879,
        0.9996615384288707,
        0.9999707900296191,
        0.9943771789259768,
        0.9998401505347537,
        0.9989472788088746,
        0.9994477701171637,
        0.9962716986706709,
        0.939891587690377,
        0.9938633847223198,
        0.9880454842146811,
        0.4747080692969162,
        0.9998674559728351,
        0.9999692068035129,
        0.9998383936439789,
        0.9631645444954877,
        0.9873023824481678,
        0.9081113555541517,
        0.9692993690684955,
        0.9780973161533305,
        0.947470141833598,
        0.9999091568582773,
        0.9996932093077532,
        0.971674701699769,
        0.9993317468716785,
        0.9989569558050225,
        0.6997211796704066,
        0.999891032789589,
        0.9998750536468357,
        0.9961094273881055,
        0.9947445855705773,
        0.9995859145463065,
        0.983516339842918,
        0.05520092687307965,
        0.9641216689308247,
        0.9655786971433608,
        0.9986771680629917,
        0.9977501453978781,
        0.9089033534368853,
        0.9999523976006447,
        0.9994891169283823,
        0.9994306842092383,
        0.988686154678151,
        0.8744371972765728,
        0.13019425084086012,
        0.9997312165094983,
        0.9998193958010672,
        0.9984484852039289,
        0.9974696061228099,
        0.9954906344495633,
        0.9914789851946739,
        0.9952001728712059,
        0.9989438036231686,
        0.9951631676965639,
        0.9998143238972019,
        0.9973486544738278,
        0.9995848604828018,
        0.9998073721262161,
        0.9999638494643731,
        0.9946841355294112,
        0.999740537843286,
        0.9998174514378504,
        0.9996660476312895,
        0.999308309624955,
        0.9960270155317325,
        0.9897576361068092,
        0.9997827713047535,
        0.9986040496900483,
        0.998849822182398,
        0.9923202015278353,
        0.9982412630076473,
        0.9998619137344882,
        0.9999407108516951,
        0.9977004298965078,
        0.9997557273637618,
        0.9998505358673064,
        0.9916769774207969,
        0.9989368242411919,
        0.9998348767841329,
        0.9171500866215906,
        0.9972038044883041,
        0.9977327667767983,
        0.9990280438950883,
        0.9987419706724571,
        0.8495861607213482,
        0.9506274259617877,
        0.9452330305667883,
        0.8748731524126684,
        0.9738627401943056,
        0.983094175354276,
        0.996354672117521,
        0.9996455472023681,
        0.9556805142548253,
        0.9920727112807263,
        0.9972463652032119,
        0.9973006244434793,
        0.9032479217258704,
        0.9940983451059853,
        0.994768331736453,
        0.9932833651298666,
        0.9996803304024894,
        0.0700358752167054,
        0.8983504505737936,
        0.9864733194277014,
        0.9976264610382097,
        0.9994421011683343,
        0.9995547102294824,
        0.9685800936144094,
        0.9978147944789967,
        0.9943485881209785,
        0.9998418353361472,
        0.9921018510226318,
        0.9854567587043723,
        0.8994624969916051,
        0.9895380882962082,
        0.9996096631938635,
        0.9998366659155732,
        0.991065440924347,
        0.9994384265826113,
        0.998965828543535,
        0.9997001224533087,
        0.7847615740407734,
        0.9973654493145516,
        0.999308309624955,
        0.9960270155317325,
        0.9981794911997038,
        0.9991428558790729,
        0.9983842885629544,
        0.9896620015383961,
        0.9654461957516981,
        0.999567971108684,
        0.9930256334808502,
        0.9933327302300835,
        0.9988767240010467,
        0.9963991510907917,
        0.9995647512106903,
        0.9992822564942729,
        0.9997339741571183,
        0.9788557616078588,
        0.9050120400142045,
        0.9991818602395975,
        0.9994868353619029,
        0.9996754566728855,
        0.9989511706263414,
        0.9689132576358156,
        0.9990936965041342,
        0.9940561354020222,
        0.5966449343700995,
        0.045701126016637804,
        0.8906897788162217,
        0.9999251180501993,
        0.9998397650661054,
        0.9999720724262782,
        0.9999453656138401,
        0.9998474442222866,
        0.9999845214577576,
        0.9991960309332143,
        0.999797537462729,
        0.9998708571341263,
        0.8906270498569244,
        0.9997075660766758,
        0.9998975113082186,
        0.9974594567485168,
        0.9994974781081597,
        0.9992218105337246,
        0.9974682355230032,
        0.9932946172662539,
        0.9998029199669239,
        0.9980541132554268,
        0.9992749726358244,
        0.9987914705654186,
        0.99715380453508,
        0.9998038899346855,
        0.9247097455812372,
        0.9795040710007862,
        0.9535583837455275,
        0.9977533370533916,
        0.9992526312041841,
        0.999966917100569,
        0.9990824838048052,
        0.9997541392002407,
        0.9999069089367745,
        0.9994122678513571,
        0.4607873215271155,
        0.9843648776502592,
        0.9936946936294495,
        0.999870500506086,
        0.999857389258131,
        0.9999726138783002,
        0.9985315337545799,
        0.9999218833031458,
        0.9983314453885936,
        0.9998604621250424,
        0.9999353813645515,
        0.9999933518697919,
        0.9954225735526994,
        0.8848083957860229,
        0.9877983038601297,
        0.9848598877712452,
        0.9993863178830282,
        0.9981073143184919,
        0.9997074803296944,
        0.9998056185724192,
        0.9997667290654051,
        0.9994914359028304,
        0.9949498956133007,
        0.4053558978865225,
        0.9990853370945437,
        0.9996429542338618,
        0.9433355877579728,
        0.9867921137636525,
        0.9988679270394353,
        0.9631006450310036,
        0.9655823520684272,
        0.9995253763960049,
        0.9995717331235432,
        0.9990505626634857,
        0.9999219425928842,
        0.999934281599871,
        0.9987882128847586,
        0.9982813319257569,
        0.9995080586941075,
        0.8335671073163056,
        0.9996648225314997,
        0.999404571858377,
        0.9999005937261514,
        0.9998378203405305,
        0.9996834904994021,
        0.9991502373952006,
        0.9999435600643424,
        0.9997810203383136,
        0.9553096614061811,
        0.9961713449373879,
        0.999323467184426,
        0.1874024972076621,
        0.9773749103721981,
        0.9986165993040859,
        0.9998802967231021,
        0.6941087690911515,
        0.835178293595206,
        0.9999292081773345,
        0.9999317045973298,
        0.98728997761669,
        0.999760458312379,
        0.9931206207480555,
        0.9988873961546452,
        0.969422832605926,
        0.999848824517478,
        0.98543277037656,
        0.9863175659458039,
        0.9934213090034859,
        0.9829452544949528,
        0.9754011211839761,
        0.9985156234079994,
        0.9961456510754607,
        0.04125911997206266,
        0.9885584683331315,
        0.7730599048476458,
        0.9993231991369631,
        0.999670322846114,
        0.9628190905363104,
        0.5974995356550556,
        0.961147986627958,
        0.8181867253122196,
        0.9998912945601428,
        0.9474975527904389,
        0.9996779490969531,
        0.8135025451099138,
        0.9622567900154403,
        0.9725907347214159,
        0.996113653205564,
        0.8326387330841173,
        0.9895053286492804,
        0.9984854893807832,
        0.9992408880957929,
        0.9045083067205109,
        0.971397863559707,
        0.9893579716916279,
        0.9910541348931234,
        0.9998207773971629,
        0.9916309268196636,
        0.9959823956680188,
        0.9997685301923916,
        0.9954581234108909,
        0.9804364561879672,
        0.960257593775413,
        0.9996466711130827,
        0.9994027732168244,
        0.9999289106256893,
        0.9751649973297188,
        0.9802355333021511,
        0.9998655333909413,
        0.9997071861491758,
        0.9967874521067936,
        0.9998044572691274,
        0.9997220477799243,
        0.9997781345029677,
        0.9993284314223444,
        0.988369476138225,
        0.9971211440108141,
        0.9904411807563962,
        0.9999066963267416,
        0.9998851641792512,
        0.9536882675189094,
        0.9923200219363891,
        0.8751582842956865,
        0.9936484302535763,
        0.9996709202757512,
        0.9997777655794987,
        0.9810207482066521,
        0.999047869891749,
        0.9957748145960348,
        0.9999277208564707,
        0.9988923708187166,
        0.999917554467604,
        0.9814894523004823,
        0.9998366160207568,
        0.9999426810281187,
        0.8306004194340212,
        0.9919951990742222,
        0.9881533483265196,
        0.9992399275156296,
        0.9998418337948178,
        0.9984514283316661,
        0.9986502020294843,
        0.9970397852313346,
        0.9997899986005718,
        0.04792184545596534,
        0.9984361064757489,
        0.714409973401163,
        0.9995695278983906,
        0.9998644771331724,
        0.9994670346173902,
        0.9937753089584157,
        0.9998206738320355,
        0.9992468371492254,
        0.9998497257786284,
        0.9999358647281763,
        0.9999635446016369,
        0.46861816300050807,
        0.9994578352740189,
        0.9993054125104199,
        0.9910682935503013,
        0.9998994407338195,
        0.9984855527329446,
        0.9998883607635934,
        0.999959038242924,
        0.9996682177180319,
        0.9814578837653386,
        0.9890826655702566,
        0.9981028738254224,
        0.9998284387436653,
        0.9992979080352098,
        0.999734402587026,
        0.9988990563496205,
        0.9753182233850558,
        0.9994675351839061,
        0.9979739810534775,
        0.9936407082619878,
        0.046838908662355926,
        0.9996413294064568,
        0.9999241855994739,
        0.998917226459464,
        0.999907409663411,
        0.9990213033703157,
        0.999929449952104,
        0.06518530885400853,
        0.9891303827212792,
        0.009040615497759807,
        0.9996221759494788,
        0.7523535468248684,
        0.9932147329655951,
        0.8394083373492287,
        0.9604301444414837,
        0.9996285871810616,
        0.9901733840440878,
        0.9945436059989954,
        0.9464815143284926,
        0.9964278065578425,
        0.9573546044531679,
        0.9545801087737241,
        0.9987136279850923,
        0.9990858060895799,
        0.28030603995363923,
        0.9993616034306136,
        0.9998193436005968,
        0.7500032788348838,
        0.9867905438392924,
        0.999589074754554,
        0.6483108457795143,
        0.9929699226361508,
        0.9971956641445923,
        0.9987658143673449,
        0.9998174019156084,
        0.9958302297207298,
        0.992907398850969,
        0.9976839817552239,
        0.9995374683539541,
        0.999238525654368,
        0.9822694940856457,
        0.9878144007936299,
        0.9979237457381646,
        0.49530259641059526,
        0.5873735183674169,
        0.9496186963614239,
        0.9994472427829155,
        0.8241827219949173,
        0.9996771535069995,
        0.9893076243036489,
        0.9971739216041867,
        0.8287139269378203,
        0.9993662085020028,
        0.9974883890160968,
        0.9995045593520399,
        0.9231506246892435,
        0.9987638051643652,
        0.9998215835291482,
        0.9988781341168083,
        0.8324988163750491,
        0.9997089071861329,
        0.986291418272589,
        0.5335639751664047,
        0.1447081521231922,
        0.9872415761784284,
        0.9972915368012834,
        0.9998845872874805,
        0.9820126145709629,
        0.47959922824867207,
        0.9970976784751735,
        0.9970479965428741,
        0.9997195041663375,
        0.9990435019029786,
        0.9999666855342808,
        0.9998778702793353,
        0.9998500986483501,
        0.999882108497463,
        0.9996494561023074,
        0.9982191937264491,
        0.9978621898709263,
        0.9952888801769034,
        0.930140303999954,
        0.9992437915834831,
        0.9999204246798025,
        0.9997706631491441,
        0.15300915295059758,
        0.9861926738430025,
        0.900413181489146,
        0.9999140934772143,
        0.9988415684198432,
        0.9999816129488411,
        0.9966668011185166,
        0.7549338263052137,
        0.6909295912764013,
        0.9997045831389171,
        0.999662889180399,
        0.9997921810630617,
        0.9967833773536159,
        0.515101934077223,
        0.9976164578596071,
        0.9998702344274173,
        0.9915135795122146,
        0.9993474370300433,
        0.9981903419964502,
        0.9995935678846549,
        0.9989692547024579,
        0.9989694423652856,
        0.9977560174695497,
        0.9994242879014936,
        0.997953228676241,
        0.9874403441961037,
        0.9046762862178154,
        0.9990740556539686,
        0.9974263681412417,
        0.9977032099581115,
        0.3809158857287297,
        0.5856961147463836,
        0.9863706212687613,
        0.9786899106042385,
        0.9922118911092993,
        0.9855115181730773,
        0.9987614774783423,
        0.9998243023308889,
        0.5778494489059706,
        0.893808978236427,
        0.5165434825543027,
        0.8619233796032332,
        0.9999292546227533,
        0.9999240695882227,
        0.9994450136030045,
        0.9698243732631566,
        0.9972582963717269,
        0.9527816328949938,
        0.9565562723143508,
        0.9956193807417195,
        0.9872672991418843,
        0.7453567343383629,
        0.7915898934087742,
        0.7875589223969388,
        0.9999351927851617,
        0.9999737289458265,
        0.9999175806541921
    ]
}