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

762 lines
21 KiB
JSON

{
"model_dir": "models/llama-13b",
"distribution_id": "code",
"timestamp": "2023-10-14 00:00:50",
"eval_score": -0.20895108926478384,
"eval_brier_score": 0.20895108926478384,
"eval_average_probability": 0.6913097155808091,
"eval_accuracy": 0.7346666666666667,
"eval_probabilities": [
0.015225864488258493,
0.006783356073918778,
0.027394207720611317,
0.9896070896765425,
0.9998395882322685,
0.9601325824434976,
0.9942722380261287,
0.8665545078981188,
0.8930521135932319,
0.5708595575526438,
0.9158077524975373,
0.9589916767178855,
0.5345636210815823,
0.7447570718623759,
0.7547841745952909,
0.9436779734713407,
0.15203883258945528,
0.7851416829111654,
0.4247467256663082,
0.9050128826414726,
0.6964448362462516,
0.28802590298189257,
0.9500009172957731,
0.9795154588378032,
0.08292254226915452,
0.06690649898300194,
0.0884539630050204,
0.9963599299141772,
0.999736585476457,
0.9999565082664389,
0.022760716179572982,
0.9144930743076125,
0.5802076736534993,
0.004005426012788212,
0.002544725204884387,
0.0026111969034076313,
0.6647163955785029,
0.8135288931669131,
0.8078535546225362,
0.9907396881763052,
0.5871888592921104,
0.9983729445528438,
0.9994172990843838,
0.998575608567661,
0.6259756151207008,
0.8325735935308115,
0.9988485073240235,
0.9949636395144458,
0.9376517178992503,
0.9572059026883316,
0.6022674437743835,
0.9914308117560161,
0.9999243264991602,
0.9994738259918486,
0.5845062650518071,
0.9393532954941856,
0.5761254337887821,
0.6193618972145493,
0.36275166505615897,
0.2612883062113268,
0.9715847990086443,
0.9924549161931491,
0.9928664122196794,
0.8697739169806408,
0.5248227487888446,
0.14753030980466275,
0.9001342585784394,
0.9784237465978879,
0.9159985915329009,
0.8883537513384875,
0.9172167617990029,
0.9997146946476231,
0.0621338145676459,
0.4521519605500237,
0.9216113535459971,
0.9965386341268481,
0.9887117393365217,
0.9991737403625659,
0.9369205914174756,
0.3309954202017553,
0.7104940776741924,
0.9988994142617034,
0.6716321400771808,
0.9998512335821009,
0.1397944462791654,
0.8738918294501181,
0.9947472743940888,
0.9999604623459419,
0.9999815121364075,
0.998000749412456,
0.3618968882497482,
0.445638552447959,
0.506261378624741,
0.7855224156697139,
0.8967674713454026,
0.23510176329676047,
0.9720546686086143,
0.8629304081852769,
0.7203426797227708,
0.6354736563484683,
0.8579747914100642,
0.052853504375150025,
0.9961202615894257,
0.9895888755809047,
0.9914049561728907,
0.9393789438384939,
0.7831048597648063,
0.8479305165541008,
0.7778114136941331,
0.5970521049774831,
0.8726196514859824,
0.9998032224112523,
0.9998838722713816,
0.9997085260894926,
0.00333487833786128,
0.8754617322084818,
0.011675390272062642,
0.9513430140803432,
0.8908011148623416,
0.7732941998970391,
0.9892182473239162,
0.9971172606825955,
0.9974424905825487,
0.9992749161898521,
0.9979483378327285,
0.8594034969163251,
0.028062682741057204,
0.008652705899222875,
0.14178027252429015,
0.9863306239400629,
0.9662832814413387,
0.783535541660339,
0.4269075855625265,
0.6868742378013415,
0.8316440290459058,
0.5945163044515314,
0.879592612006243,
0.025849937391222433,
0.09293082195720849,
0.8831690847511139,
0.5700411881896449,
0.0787910356867556,
0.06548749264635356,
0.17505252887839762,
0.7629716917916736,
0.8866030205283258,
0.9356120336649029,
0.8985740214037516,
0.23939948511674136,
0.7136541615907113,
0.5857674772128648,
0.4031829639319905,
0.49724963375837966,
0.0023045011008444583,
0.9822254821447016,
0.9452399973014303,
0.011849317137511795,
0.0028040389710696223,
0.0061018823612633,
0.9928424424204414,
0.647331436097814,
0.9661984358335595,
0.1354923720614729,
0.33768832058316633,
0.06956873152609899,
0.9799370533115352,
0.8457462167720355,
0.933756290305521,
0.06556098500926161,
0.02763764653567997,
0.5410945279232456,
0.2772904216337284,
0.043621410319397,
0.6972753748049234,
0.2205236366023512,
0.917613354127294,
0.9048014515973156,
0.11065778275893844,
0.9997985750487285,
0.9994002120062335,
0.1846884156835571,
0.11572835687981664,
0.07559352442567302,
0.9410852116854096,
0.9749787920169812,
0.711488895428556,
0.7803309460789943,
0.43177993071636367,
0.3236204866026336,
0.5585715711978296,
0.7745480566891033,
0.9984500159075527,
0.8027665193240807,
0.5498921342471095,
0.6224268279230404,
0.9997874897825577,
0.9990043959015938,
0.9915583159016921,
0.9507044262878891,
0.9517840252513243,
0.8765070391860106,
0.9978651961114298,
0.9983825405753369,
0.993432646657054,
0.5209404914249781,
0.9392290766127974,
0.23444347807268628,
0.5034204213016791,
0.9651060897814636,
0.9943284585641994,
0.5318517490958488,
0.841526336318259,
0.2674718430824733,
0.9852327889368209,
0.7370236773153385,
0.0389095915430997,
0.9877492581668196,
0.8984400062020834,
0.9923172482656666,
0.15569104133237965,
0.7515298592755848,
0.30599621512569075,
0.4245514903935501,
0.9742109497779634,
0.85842625680294,
0.9883316259738584,
0.4346659373423881,
0.8713861295506431,
0.9639063490608062,
0.9698403210100228,
0.9376149591200199,
0.9518521639876887,
0.5877118117435043,
0.6745929760566352,
0.9938391834071514,
0.9954421089935673,
0.9921893383801818,
0.2841261664987765,
0.16201329751670746,
0.28792981937060824,
0.006470140048214416,
0.0075640569766169015,
0.0018556816275970833,
0.6493032881370211,
0.8911182292819049,
0.9127516494105987,
0.17583029470797742,
0.746004996736427,
0.999719732425265,
0.9966198791793515,
0.08838947724612256,
0.9975946356925942,
0.18119087311397555,
0.24505636233496142,
0.9433759162977358,
0.7895918626645477,
0.9459633545991816,
0.9990990706250729,
0.9978698534558081,
0.9760873496432212,
0.3352178469292676,
0.9796982526998563,
0.9563033912219385,
0.7638393162696895,
0.975295232622846,
0.6208547880459863,
0.7633743314653094,
0.016272872590980425,
0.0800827115319199,
0.05203836320851167,
0.9379221181636186,
0.969588070490507,
0.7102060496644846,
0.3054770586312673,
0.14446495025582176,
0.2731257946335435,
0.9966630392927026,
0.9876254863720724,
0.7244515131056735,
0.38769765470366935,
0.7799592716891925,
0.7933617248579028,
0.9977147741966401,
0.9267897119108042,
0.865049681721737,
0.9998984404933433,
0.9998835542881911,
0.9997449728416428,
0.342377215779888,
0.42044748454731806,
0.6892933471268347,
0.9857183112955367,
0.9871323219993235,
0.974610482600906,
0.9837325389853678,
0.9912960785557844,
0.005103092767415718,
0.9823168776911801,
0.8277867170144967,
0.5680933199698622,
0.4658532786448778,
0.7383038173328516,
0.930554269595358,
0.9596593612941646,
0.9669818274342102,
0.9819884107962772,
0.9939588939698186,
0.9766757193995237,
0.9122000476949979,
0.5916916316598697,
0.9281831295914469,
0.9701771998159813,
0.9263388328426326,
0.08092830057250788,
0.09595662478478498,
0.6775964050439882,
0.9727011923823118,
0.8450780959366391,
0.899644377602425,
0.48705824363485795,
0.49406625429596707,
0.3044688277685986,
0.6292083525531216,
0.23871048988424937,
0.9261747232686345,
0.9933388784587456,
0.8246450365968627,
0.4336234338206736,
0.3656641632174843,
0.7121938894431975,
0.9877286620576724,
0.9864250069369239,
0.5610201095757104,
0.48620823935949453,
0.5304408921007879,
0.6984734760794116,
0.9330857945857752,
0.6040005159985029,
0.6330377163737153,
0.018297205091453172,
0.01196571283545711,
0.027062909426011115,
0.984192075935986,
0.9762474923878447,
0.8640209069184793,
0.972140105688666,
0.9290528292068463,
0.9888624287195007,
0.9513675470815578,
0.8966026154365306,
0.6054899384177186,
0.0002873296992430552,
0.0008440793736632791,
0.0002666048123006748,
0.9962915209309062,
0.9899497378860932,
0.9885505519052411,
0.9055591331914208,
0.9372544576514827,
0.920831686601404,
0.6473546575732267,
0.32615356758479597,
0.932642579009443,
0.9794941398553484,
0.9558132624487439,
0.9986258782698167,
0.9963450483650639,
0.9860332668392322,
0.9991229450353515,
0.6818525185617974,
0.6933864641573017,
0.4518034739036769,
0.988139411976533,
0.959955382486244,
0.9820715878644223,
0.9962288058635221,
0.9581373747598626,
0.5320276084801356,
0.9205363038185865,
0.8869731605485937,
0.8755940288085376,
0.031628409793951956,
0.42815940681387565,
0.7379158657845302,
0.5490925785085679,
0.736238710851613,
0.598716201938609,
0.9175007873357995,
0.8086368108887033,
0.8906469103755237,
0.6138842250896505,
0.9048577476024614,
0.8998246047126539,
0.15643103768621314,
0.9821946874857688,
0.2723305532317767,
0.15395563713362936,
0.8873799823754794,
0.9598778977289046,
0.9149319318643097,
0.9729920470955119,
0.7135096079843294,
0.7987326252074148,
0.5953084941488139,
0.3936659586190864,
0.0008572316699557714,
0.0025999606976982404,
0.0006561055936610272,
0.7261743404361389,
0.46839436219755526,
0.9691629839895468,
0.16920983442813192,
0.47364543830121464,
0.5825881237114072,
0.9966414173909576,
0.986513304409979,
0.9871825302718924,
0.03743427142120821,
0.7256931899008463,
0.5439730571441777,
0.17017215778173392,
0.2701392204950771,
0.21763664318206163,
0.06587296089202332,
0.22929686902371033,
0.1815601066684439,
0.6665262880057166,
0.7811848719016455,
0.00829647288742614,
0.9024439780667884,
0.9313401134951222,
0.992474984558442,
0.6753156218873468,
0.7222553919285425,
0.9213720671380329,
0.772547978918442,
0.9992570189696288,
0.9980513313013256,
0.7499085834266401,
0.753371765159559,
0.9924441493383767,
0.9801983623344946,
0.9968320610569543,
0.9273824321976197,
0.9954304058877265,
0.9878829051442626,
0.9702033731976307,
0.44072013307341323,
0.9628950645931825,
0.7609599436448157,
0.6573273054898479,
0.6018347534890056,
0.9796615762304368,
0.5348931341152132,
0.036488653750694544,
0.9710758959781888,
0.08050918143270558,
0.5084647175694604,
0.963585288629109,
0.9643343604387602,
0.7651731226568551,
0.5938334351156606,
0.9756782240063643,
0.9769152165895262,
0.9785202387732886,
0.004660472767591073,
0.0009021647269378263,
0.00963608193584981,
0.6991321216045819,
0.9935419858211257,
0.9998715219020821,
0.9875450696483219,
0.9379841093380923,
0.9997663015192662,
0.921114909811216,
0.3396798458870204,
0.30757542209522487,
0.0871043935606366,
0.1486842412996387,
0.020520162609040438,
0.8054712848359127,
0.05483056262437187,
0.9340463858724493,
0.5622621440679271,
0.36733724949162905,
0.5109713144041117,
0.9679644226486672,
0.9925635912181784,
0.9957973193139955,
0.9973686044286982,
0.25433584069792975,
0.9811342160683627,
0.9991725534542404,
0.9351105000947966,
0.5823340244429758,
0.9330393958812341,
0.6553944350952607,
0.9997462656493384,
0.3102581578585592,
0.9798725487315042,
0.7689245669634788,
0.868508353509166,
0.6690038634217288,
0.8736857498811581,
0.243821355026192,
0.21802181703705892,
0.6848780127786046,
0.2023546602986226,
0.9425327369667961,
0.016233075345182778,
0.000525546881326111,
0.0019043283194844385,
0.0012985154771276602,
0.044591753434370086,
0.5048923894144363,
0.9411263762966875,
0.999943839237481,
0.9995693403628083,
0.9995523337092347,
0.6113449432534597,
0.8881532948634974,
0.8595547598149063,
0.5760194177540874,
0.1217893540171523,
0.7433299191364235,
0.2644805327857288,
0.8866889225287047,
0.7877865715468564,
0.9903953033670396,
0.8626949928562049,
0.9864609253497315,
0.9788790443964362,
0.980221327512043,
0.9724977265194709,
0.9703047382488245,
0.9439101675947875,
0.9458661751349858,
0.5869084668237374,
0.8196871875323416,
0.7869597488128592,
0.9997844769780958,
0.9856492543437025,
0.9982137903863308,
0.9952971564586294,
0.9961954501011699,
0.9986911918828492,
0.876561606605698,
0.629572252771109,
0.6621419282080225,
0.8172818046270958,
0.5010583109049074,
0.5925931273026162,
0.9606918291679621,
0.4714127141698756,
0.18007407938926057,
0.018226672946389466,
0.06679799504279257,
0.0750763620898393,
0.9746385224959785,
0.7048136492986163,
0.11199697886928095,
0.7838489881837878,
0.8936357114809259,
0.9194793112815104,
0.9317759317896437,
0.9745163628454739,
0.49069517339002044,
0.542375836749079,
0.9152099466741974,
0.9869271472279814,
0.9999062682961939,
0.9109848687763338,
0.9407218305085179,
0.17112631610104978,
0.8861181144900062,
0.7734665007941381,
0.9997347309610325,
0.996630273300557,
0.999362084380205,
0.998820400744454,
0.9955465796850308,
0.9991554849971181,
0.4126484809672283,
0.6309323935269091,
0.013373032763264958,
0.018559323277239526,
0.028988349574054512,
0.006511966519782253,
0.8919519877748895,
0.9687222736102681,
0.9827586131923074,
0.9036738337329263,
0.8558424375434192,
0.47111506636993394,
0.9990290081586591,
0.995385246974195,
0.9924663073123254,
0.31561749744385514,
0.9843349892315322,
0.9958106145022289,
0.9832287109936992,
0.986781459866834,
0.9950893778491893,
0.020888239490503554,
0.00189775098963641,
0.0016320873541292376,
0.8975280506996717,
0.9373250727934682,
0.5578956418463665,
0.9861639460538801,
0.9916621438635299,
0.2349036914878525,
0.97547140566102,
0.9189139535316272,
0.8796655831693566,
0.9950373112158124,
0.982777365193696,
0.9953781857124141,
0.009232198265521574,
0.7042464550043706,
0.6605492955027392,
0.03259539880543061,
0.1906066740554731,
0.11225232038601986,
0.9533263442968218,
0.7089697496754868,
0.9391547968338091,
0.8869046436195176,
0.45855443447480837,
0.5721439012175542,
0.9997167552273614,
0.952122598621884,
0.6861766837305776,
0.990323436147919,
0.996610819310669,
0.99919449718717,
0.9957462835718834,
0.9034128610229579,
0.18986015264310008,
0.6422737464648344,
0.6952968758713122,
0.2481637974100971,
0.9631325639983482,
0.5763853065414547,
0.01571340719218539,
0.999352463222844,
0.9946950868276291,
0.9851021223182431,
0.975445248922105,
0.8640330143545855,
0.8996195924967889,
0.9993743420431984,
0.9941377294665827,
0.9996198044264611,
0.8532968488068217,
0.7975375540339177,
0.9741573669734391,
0.9908632372000001,
0.9665869966153416,
0.9787835120354837,
0.9871769180067491,
0.9892165914589262,
0.9514914613051557,
0.9767204894876201,
0.9527673930185906,
0.999978896610215,
0.6000784560714484,
0.28268032080881034,
0.934199287146067,
0.9439292857123355,
0.77879504465824,
0.9296766171050704,
0.9083829523196914,
0.8246367728683958,
0.7930628692414123,
0.4455083925749694,
0.7117851153324871,
0.9908351746238983,
0.7750450190497833,
0.29614123700164174,
0.48091518975565795,
0.9995999874828136,
0.9999776398117461,
0.9993968504627246,
0.999175944411488,
0.9996483399107056,
0.7237377450058803,
0.985555708165035,
0.9455683533294204,
0.9735425131158743,
0.6996869305578348,
0.7940606354554316,
0.6416095932296679,
0.9815806190696177,
0.4657375797297302,
0.9960870062005539,
0.9922468655857448,
0.9205810341502887,
0.9917327224431839,
0.8186725883907412,
0.4604309824690722,
0.5232426185511032,
0.05159262160349776,
0.8049092854616962,
0.8874103416875441,
0.9958873618170146,
0.09449463965803195,
0.3416159183785987,
0.20014670948198754,
0.3418787232928786,
0.6780372241214285,
0.7887713625068742,
0.36538929528435415,
0.0899040055168856,
0.8642042282998748,
0.8327226872488784,
0.796259518352364,
0.4941462470329371,
0.601977206613981,
0.30628803234899815,
0.6399369849331235,
0.3259129769385164,
0.4972650626086945,
0.38998597671448443,
0.09704786467289672,
0.0218611428534502,
0.9666742942927508,
0.8933446544091018,
0.9027006615375167,
0.2999349213501955,
0.4151879366465844,
0.7673482691077939,
0.9189978813111314,
0.7478923994183468,
0.8936063765048883,
0.13271186735421625,
0.04439864350216127,
0.8643117847785396,
0.8785661672104714,
0.9740815871506598,
0.1885577837770258,
0.9914587988434109,
0.9984906462186776,
0.9971127103608567,
0.998486921133368,
0.9718818382603412,
0.9928672041052671,
0.8085306736139192,
0.5094553522945389,
0.9963465713485155
],
"eval_rms_calibration_error": 0.31354962396253483
}