GENIES/results/evaluations/mms/llama-30b-code_low_quality/code.json

{
    "model_dir": "models/llama-30b",
    "distribution_id": "code",
    "timestamp": "2023-10-14 01:14:34",
    "eval_score": -0.16097716358297,
    "eval_brier_score": 0.16097716358297,
    "eval_average_probability": 0.755443840226703,
    "eval_accuracy": 0.7933333333333333,
    "eval_probabilities": [
        0.49643245396510993,
        0.2243494464861274,
        0.4175968291592897,
        0.9994623645828461,
        0.9999909748447635,
        0.8559683843190029,
        0.9460399150679671,
        0.987884558924983,
        0.45671711434142515,
        0.9476194230068302,
        0.9706752946689075,
        0.967698514622832,
        0.6194275694800997,
        0.9822731521158813,
        0.9163031618590683,
        0.2637827626980066,
        0.9086652097925277,
        0.707615220727918,
        0.9704731070871278,
        0.9954952465832415,
        0.8973337643220868,
        0.6567949309147639,
        0.703596425935259,
        0.6111780240012653,
        0.2303244739080322,
        0.23861834049237549,
        0.0781753905429815,
        0.9999130906239668,
        0.9999277488360145,
        0.9992326566033982,
        0.8485150726676113,
        0.9999984149418353,
        0.94426087187095,
        0.0055728236016471255,
        0.003932752605914858,
        0.014670767653133247,
        0.9708408151945019,
        0.9970412242531549,
        0.9942010552525413,
        0.9999200984999136,
        0.4489175260642231,
        0.9875478155485137,
        0.9062472708498476,
        0.6664694690967718,
        0.9815469890622371,
        0.9646087035664793,
        0.9999582921012949,
        0.9998691122025329,
        0.8841961924568974,
        0.4629344228932193,
        0.7943204673373858,
        0.9998298849836366,
        0.9999914214484292,
        0.9999516519838991,
        0.9440894438183729,
        0.8654406233135704,
        0.012003608465592563,
        0.23780047740367583,
        0.49970253875229453,
        0.34719251315753863,
        0.7487460971327176,
        0.732290496572787,
        0.8810059553078696,
        0.8860739824339011,
        0.791757647576552,
        0.16627006502447358,
        0.9951479949573617,
        0.9113819604559212,
        0.9472777961538739,
        0.9284028090286669,
        0.996631865248509,
        0.9909854242986521,
        0.9699019577683597,
        0.8889943371810233,
        0.45335766724759574,
        0.9993728760878346,
        0.8784775530056832,
        0.9999325523339451,
        0.9770096362187894,
        0.8118971788994535,
        0.9535487277361114,
        0.9934620182201845,
        0.5176252950249993,
        0.999893749425176,
        0.9997051325631704,
        0.9955603459423283,
        0.9999537246790536,
        0.9999863552035071,
        0.9999975283426275,
        0.99773490872844,
        0.8258963948703227,
        0.7900021047229011,
        0.9765293690456548,
        0.5376561373344654,
        0.43084670938813574,
        0.8251371942178011,
        0.9883462418006559,
        0.8656925430875825,
        0.833185007568176,
        0.6033267904940174,
        0.9181531572352871,
        0.0006579027915107893,
        0.999989221448182,
        0.9998647197541297,
        0.9998826418676839,
        0.45184704093540384,
        0.6240271369090985,
        0.8763929365010228,
        0.8724584212318183,
        0.4351206270704847,
        0.7321096363911044,
        0.9995694975196933,
        0.9990555858990553,
        0.999969584794147,
        0.0318143306182882,
        0.013695172061184126,
        0.006996872535652732,
        0.8302661650537713,
        0.37347300098135583,
        0.5206843124983616,
        0.9572149259135222,
        0.9934405459462918,
        0.9972253120567086,
        0.9999845294706424,
        0.9998863139058168,
        0.2924390331769968,
        0.01370129562508482,
        0.013024593891883951,
        0.9283419738264113,
        0.5318178116332357,
        0.9811633113638825,
        0.12528065263396626,
        0.18682870061012766,
        0.9034618140359417,
        0.7115899188364257,
        0.7330263572274534,
        0.6584179048959349,
        0.1694139998814897,
        0.02634177238068775,
        0.8630830040042166,
        0.36846538656232203,
        0.13287528449747896,
        0.02863179828336615,
        0.6387885937951081,
        0.7577058140363546,
        0.983575987076891,
        0.7879477799548598,
        0.6847630412841312,
        0.8530034659896366,
        0.6270947665544325,
        0.9603738306397701,
        0.9817075074979964,
        0.9910114606269308,
        0.635908661020284,
        0.2701063394041193,
        0.9403383033860158,
        0.002168034694603027,
        0.011708141079820523,
        0.023975466059906365,
        0.9992628879459078,
        0.504328686457721,
        0.9535124515205898,
        0.100824071768575,
        0.5835045900829564,
        0.20479799471269372,
        0.9108600847243337,
        0.9610175020557438,
        0.9121737569792426,
        0.959679768608833,
        0.9240009007018156,
        0.6440411445893136,
        0.9410691085033798,
        0.7502884810486908,
        0.8453260429888472,
        0.9837887865005153,
        0.3487929610191597,
        0.9617691776495048,
        0.7514730318163809,
        0.9999169660802645,
        0.9999951316242014,
        0.31082839716929894,
        0.26780020681392686,
        0.2868875024504821,
        0.9993982731793909,
        0.9984194627033819,
        0.922570958470121,
        0.5457590534043371,
        0.957077791059124,
        0.1801716357533131,
        0.7239550921398231,
        0.7714865202283779,
        0.9999478483649775,
        0.9641218546102572,
        0.8049281786014205,
        0.9426891812100674,
        0.9999640838273637,
        0.9999344054681476,
        0.8898069434603277,
        0.9971639493084938,
        0.9907055855665677,
        0.9999762686368313,
        0.9991113699045436,
        0.9836177684015359,
        0.9834493601895408,
        0.9228259658279607,
        0.9985318526849438,
        0.9914548326971886,
        0.7965438634247177,
        0.9637699438050646,
        0.8820433888145657,
        0.5811628445817335,
        0.24469458121480292,
        0.964284849822637,
        0.9098608264083589,
        0.6527453376083888,
        0.5951855261921377,
        0.9895313853105175,
        0.8841232550177578,
        0.9927787282765645,
        0.9338446451216279,
        0.41264857402734756,
        0.966458835779869,
        0.1542959436021668,
        0.5272886850755478,
        0.944672818818814,
        0.9895144804199987,
        0.941533664788205,
        0.637006208217441,
        0.5699038307866071,
        0.9681297431684937,
        0.997621532059574,
        0.8851690234058661,
        0.3649787100779528,
        0.7562202841934998,
        0.9848945799375407,
        0.9931003239866819,
        0.9203269321171692,
        0.8505227953107407,
        0.954024303117863,
        0.8042617151202593,
        0.017815858304501497,
        0.05910743383576922,
        0.01539988465900588,
        0.9123881085326522,
        0.27972165487342315,
        0.9983194832113064,
        0.24261444346626704,
        0.6817039343884135,
        0.9999877833355127,
        0.9986890653529572,
        0.22038489890930002,
        0.9996968056989992,
        0.8229312945453561,
        0.07611123727576856,
        0.8925284525892754,
        0.9878994068575604,
        0.9925507806532592,
        0.9657232395600046,
        0.9999784045810182,
        0.9989430400557431,
        0.9985483578088171,
        0.961120482693115,
        0.8822009462937239,
        0.5853918658880146,
        0.9981641856068746,
        0.9799784675411721,
        0.9931439492839943,
        0.01604768004372381,
        0.039885785268088295,
        0.0057436472196013,
        0.9796899319620437,
        0.9976382572447268,
        0.566621251408722,
        0.7602134487215216,
        0.9172226012011506,
        0.8156131442652509,
        0.9920323260767859,
        0.9846340637007504,
        0.9752447102714982,
        0.9808057195654857,
        0.9147234819572989,
        0.9601741488996873,
        0.9880156527169547,
        0.676912310834131,
        0.9920776267490657,
        0.9999018751787552,
        0.9985709763606413,
        0.9978124261422027,
        0.9621392382108112,
        0.9648278249453613,
        0.9731459317174902,
        0.9356410770003546,
        0.8440344176407232,
        0.8568189355712991,
        0.990965341933749,
        0.9770497665169815,
        0.2692333990415197,
        0.2183531546526264,
        0.5728703469929719,
        0.21688892686940167,
        0.2622778912579755,
        0.8551515017777251,
        0.9558142867933662,
        0.6307713442988203,
        0.25239185540548237,
        0.9239376362775374,
        0.9846099599450259,
        0.985557895418764,
        0.9918507784216946,
        0.9747622767932067,
        0.6287573379670213,
        0.9844739363999848,
        0.9216712433004711,
        0.6980536818251317,
        0.19466033770236935,
        0.4890551988885959,
        0.8927671746997892,
        0.7456380106218639,
        0.8875711932318993,
        0.757256534668125,
        0.42263864532192386,
        0.20746480454956884,
        0.014650226763161411,
        0.23554390566108896,
        0.9543331013014886,
        0.9756599861281084,
        0.6855614030145813,
        0.9755459206775832,
        0.502161523203613,
        0.9722309745380813,
        0.9940143024582743,
        0.644713553057622,
        0.9085823472720156,
        0.8154414506679708,
        0.9477509829926628,
        0.9308969992117422,
        0.7965809266623558,
        0.9700159339659337,
        0.11963254854958039,
        0.008010765366208372,
        0.0017608524690060356,
        0.007739829478322576,
        0.9929885287136172,
        0.9923633552384725,
        0.9947334556255543,
        0.9997743290516011,
        0.9896465939462212,
        0.9786394223313365,
        0.7922140332370059,
        0.9060616166647535,
        0.963936275008226,
        0.004354921834854308,
        0.008002613690718407,
        0.01142718230396513,
        0.8893568656341962,
        0.4038648059511811,
        0.9998774265769578,
        0.2437605510211085,
        0.9849902500206068,
        0.9984174912287578,
        0.9785111526295387,
        0.45766041419560877,
        0.818753060380699,
        0.9801726133583768,
        0.9796018629779607,
        0.999691241680348,
        0.9952914525718756,
        0.9977565935757213,
        0.9999208389505543,
        0.8057752554956603,
        0.3873135592978276,
        0.2509806832701421,
        0.9997252074747939,
        0.7033496867334164,
        0.985467487827698,
        0.1391575114172925,
        0.86030514162374,
        0.9543129918353571,
        0.9967197348196182,
        0.9923645960005005,
        0.9730765663473531,
        0.011481962248590438,
        0.8334761775953967,
        0.5928153948471053,
        0.8575340899848656,
        0.6607586301337515,
        0.8661780162436358,
        0.956988337777014,
        0.9805182775449433,
        0.9472030958689648,
        0.8572105129086788,
        0.986796515895145,
        0.976435005945856,
        0.9374822358591591,
        0.9817898892330682,
        0.8114943683649377,
        0.9559857513416137,
        0.989946290065713,
        0.9988998522661504,
        0.9943155641160577,
        0.9659340697637356,
        0.99788464600078,
        0.7713774516389985,
        0.18281688480138236,
        0.3885589664084213,
        0.03263897615091127,
        0.0016835106527747223,
        0.008961483637658266,
        0.846944310631735,
        0.7947981732875024,
        0.9077236588646467,
        0.896588313018201,
        0.5403129729142467,
        0.9837455549003589,
        0.6614265376316734,
        0.973125215858181,
        0.8393456156771069,
        0.23669756991033764,
        0.31006638484481974,
        0.7928550901703676,
        0.2915546253726309,
        0.9640451789458112,
        0.7909818214169155,
        0.019881833808108214,
        0.04851797190139064,
        0.28106939782648366,
        0.9723597228414694,
        0.613406492105343,
        0.42757145137161195,
        0.5398797197120484,
        0.5187023928218639,
        0.6860265767557934,
        0.74207328300977,
        0.9024376033616961,
        0.9107136539582017,
        0.9999747674989328,
        0.999837461055671,
        0.9994955413435833,
        0.9859078046392354,
        0.943108782570723,
        0.9611755015661393,
        0.9777381301874312,
        0.9966165596563418,
        0.9991630489469334,
        0.9638457253341652,
        0.46103873997718736,
        0.9706841052429819,
        0.7880635351697838,
        0.999800808174824,
        0.3455061726949239,
        0.6434096115979651,
        0.9756441257632378,
        0.9856131554413152,
        0.2817316018520626,
        0.32297793275753306,
        0.9910372352009486,
        0.8994224497427304,
        0.6153928689416162,
        0.9724989724491476,
        0.9990968585174551,
        0.9636914531211637,
        0.9793337322004704,
        0.879954157217939,
        0.984509015154434,
        0.9614801999548919,
        0.04982219171366875,
        0.031469638394776074,
        0.031929812731644525,
        0.10912789732721695,
        0.9996432002008967,
        0.9999832214652671,
        0.9800185043307938,
        0.9562515994715252,
        0.99914010378011,
        0.5604127521417985,
        0.9309302624461122,
        0.9058069128855568,
        0.9906989062037004,
        0.9770995360012643,
        0.9885969737082446,
        0.6977564737224727,
        0.6232237229243456,
        0.9882797515114767,
        0.49971650848095583,
        0.9524379334837842,
        0.7883539626648061,
        0.9996559168191693,
        0.996624244660262,
        0.998961822492405,
        0.9977413313095708,
        0.4379852180698831,
        0.9988769625907683,
        0.9999004268580677,
        0.999564387658513,
        0.9919886976286969,
        0.9216620695820822,
        0.3558715051060529,
        0.9999841829737202,
        0.9945977112134793,
        0.9996046770524751,
        0.8938775433592568,
        0.9661019857694025,
        0.9811064502881605,
        0.8561223150663141,
        0.006136210793946387,
        0.03639482137643181,
        0.7755232629764748,
        0.9780213249978761,
        0.9021190442216839,
        0.10659625791589866,
        0.009292904202872099,
        0.010574176079246933,
        0.012641028922357855,
        0.857625317710287,
        0.8944430162010921,
        0.9970162504593353,
        0.9996500748622171,
        0.9999158513678432,
        0.9997194429611304,
        0.973173229566786,
        0.6119873609332509,
        0.8172526071137223,
        0.8282968997784471,
        0.6275237540309581,
        0.7953804007373434,
        0.9569796597366197,
        0.9916771692269948,
        0.9960937106340892,
        0.9995199148115448,
        0.9970607950436229,
        0.9997057332861549,
        0.9368518557313695,
        0.8637537668176308,
        0.823192038008811,
        0.766165548249997,
        0.905919415452552,
        0.9974144160468188,
        0.9665130833540099,
        0.958098662565416,
        0.7590093733991741,
        0.9940654301084757,
        0.9942595726216372,
        0.9450327759722645,
        0.999675172551505,
        0.9999937562212147,
        0.9999776062945622,
        0.992049489785731,
        0.959851627993621,
        0.995665045372571,
        0.9999024305287656,
        0.08080065027481753,
        0.9966301548400998,
        0.9998874635158688,
        0.9069409004843952,
        0.5997644395285402,
        0.02024686637577368,
        0.011867003484143046,
        0.012078150746512195,
        0.9455166007375186,
        0.6404735106872832,
        0.03491060533114377,
        0.32844214934028737,
        0.24003186685352873,
        0.25890461843981116,
        0.9996333541084653,
        0.9979495219661128,
        0.10527585714499055,
        0.9156986114015814,
        0.9956977730555556,
        0.9905369724439351,
        0.9998688381068981,
        0.9963615628058263,
        0.9632119048817231,
        0.8106724012558579,
        0.26629486772146227,
        0.9375522695191296,
        0.9992905247618267,
        0.9959907727875397,
        0.9899285167737926,
        0.9351058651863052,
        0.9991452445302652,
        0.9971206696408188,
        0.5912573468031311,
        0.6841424921835682,
        0.0918595604569293,
        0.004460148796921248,
        0.00892316500527246,
        0.0037843529503900353,
        0.9503684773435158,
        0.9639514507477485,
        0.9879630689451654,
        0.8720575116923691,
        0.903812375818206,
        0.9961058413242904,
        0.9999056977396598,
        0.9971511379941711,
        0.9947521495281918,
        0.018032074181865112,
        0.9100064127063066,
        0.9652990310772672,
        0.8317579846700741,
        0.999815620992937,
        0.9982814206793587,
        0.06112550348130731,
        0.00024629242708773535,
        0.0003133169651968748,
        0.19131007244686826,
        0.9647045271498337,
        0.6240307955521034,
        0.8718913735436891,
        0.9411250086236651,
        0.6160560908368082,
        0.9991789701570027,
        0.8899467702710055,
        0.9987477957801351,
        0.9957397592396701,
        0.9826529971600649,
        0.9838248869350436,
        0.7862259014507923,
        0.7579869818115103,
        0.8719453600558185,
        0.8135885404972513,
        0.4259088201765913,
        0.18603417990068216,
        0.6397748001145721,
        0.5049079654116829,
        0.5811302635646646,
        0.9888513374695107,
        0.9908642723980253,
        0.99776830863978,
        0.9942180209294272,
        0.7945755443073776,
        0.9457778941216831,
        0.9975791534255096,
        0.9999951102494775,
        0.9999917273104164,
        0.9973927090114493,
        0.9064925356133985,
        0.7468046899832749,
        0.6885819798854587,
        0.5412285219271691,
        0.39116719687963164,
        0.9400039602207569,
        0.6115055580410577,
        0.5995853016254828,
        0.9999311287457783,
        0.9998213912569172,
        0.9941667065265163,
        0.999232971402935,
        0.9932759391701934,
        0.9730233706092294,
        0.9999899891306248,
        0.9998786537134291,
        0.9999880560348214,
        0.9972832601565402,
        0.8164901399944655,
        0.9701506774330318,
        0.9968101211843455,
        0.99944697620726,
        0.9665167002694901,
        0.9491197708111705,
        0.9857221652011473,
        0.9623495642463067,
        0.9575412517927929,
        0.9091030800710542,
        0.9988343790983243,
        0.6526843614998257,
        0.16766466907171484,
        0.9948259068899615,
        0.9710553725021746,
        0.2647884085433323,
        0.9107288690881142,
        0.9020979540995421,
        0.7817936373690603,
        0.7072292525238749,
        0.44830645754827647,
        0.7338330786744378,
        0.40997400563371367,
        0.901043710951967,
        0.7086197540008178,
        0.9699337314494655,
        0.9974395754020805,
        0.9999704333986066,
        0.999898105499334,
        0.999946527949413,
        0.9997201251743587,
        0.9561909311245671,
        0.9525254520241133,
        0.319696221694589,
        0.8249145605603545,
        0.7874768776272467,
        0.9631784954709142,
        0.9710232130791271,
        0.990188007752406,
        0.10997459779747326,
        0.9993320096416558,
        0.9986044633619995,
        0.9908806704980612,
        0.9903859662205632,
        0.9700064750944106,
        0.9522941950858878,
        0.9677363196224893,
        0.08649381663906508,
        0.9823720530892948,
        0.9983125267064527,
        0.6419542754050024,
        0.260028233536148,
        0.11703252955476251,
        0.09794865205663972,
        0.20225289664282547,
        0.29881095036471117,
        0.87636376224607,
        0.8577704791603122,
        0.04671132715887094,
        0.9868309778720978,
        0.8620170214714546,
        0.7534435771097284,
        0.7730915326314337,
        0.8078981601071232,
        0.9623296526465125,
        0.9237873813118043,
        0.9935607032428273,
        0.9883542949758767,
        0.928083932541805,
        0.6422597786735185,
        0.4043895944534822,
        0.9965244149714844,
        0.9992865281452131,
        0.9831723279970974,
        0.8914248302541449,
        0.9809964573316875,
        0.2198187474579325,
        0.9893037031995746,
        0.9562408764263349,
        0.8316775920622772,
        0.8557367585752774,
        0.4172741040209441,
        0.8739604596598175,
        0.9915370955572694,
        0.9897312793032578,
        0.991527212721984,
        0.7952593985720515,
        0.47348350896123326,
        0.7699229020893403,
        0.9994427824199165,
        0.9950823905497389,
        0.9785157093738241,
        0.9992594523138301,
        0.9583156073066216,
        0.9969221208146579
    ],
    "eval_rms_calibration_error": 0.31796311234402386
}