GENIES/results/evaluations/zero_shot/openllama-3b-math_textbook/math_textbook.json

{
    "model_dir": "models/openllama-3b",
    "distribution_id": "math_textbook",
    "timestamp": "2023-10-20 01:13:45",
    "eval_loss": 0.6873066425323486,
    "eval_score": -0.24729467928409576,
    "eval_brier_score": 0.24729467928409576,
    "eval_average_probability": 0.5029428601264954,
    "eval_accuracy": 0.8093333333333333,
    "eval_probabilities": [
        0.5052509307861328,
        0.5045931339263916,
        0.4561598300933838,
        0.5046326518058777,
        0.5147757530212402,
        0.5187245607376099,
        0.49812600016593933,
        0.49996283650398254,
        0.504591166973114,
        0.4898170828819275,
        0.4878060519695282,
        0.43748876452445984,
        0.5008367896080017,
        0.5018362402915955,
        0.5011667609214783,
        0.5054185390472412,
        0.5111318230628967,
        0.5108996629714966,
        0.5011318922042847,
        0.507175087928772,
        0.5163823962211609,
        0.5022792220115662,
        0.5050784945487976,
        0.5062726736068726,
        0.506278395652771,
        0.5039777159690857,
        0.5032798647880554,
        0.5067650675773621,
        0.505615770816803,
        0.5079249143600464,
        0.5024383664131165,
        0.5069177746772766,
        0.5090782046318054,
        0.4998084008693695,
        0.4988110065460205,
        0.5003505349159241,
        0.5309107303619385,
        0.5331804752349854,
        0.5325173139572144,
        0.500783383846283,
        0.5012311339378357,
        0.5043273568153381,
        0.501581072807312,
        0.49759605526924133,
        0.49817389249801636,
        0.5052927136421204,
        0.4996333122253418,
        0.5060946345329285,
        0.5029692649841309,
        0.5041553974151611,
        0.5070971846580505,
        0.5035862326622009,
        0.5073332786560059,
        0.5070537328720093,
        0.5031270384788513,
        0.5016029477119446,
        0.5018417835235596,
        0.5037859678268433,
        0.5028612613677979,
        0.5018438696861267,
        0.5051899552345276,
        0.5202947854995728,
        0.5188743472099304,
        0.5037903785705566,
        0.499957412481308,
        0.4996558129787445,
        0.5028499364852905,
        0.5039554834365845,
        0.5010066628456116,
        0.5017892122268677,
        0.506775438785553,
        0.5057093501091003,
        0.5025622248649597,
        0.5041351318359375,
        0.50318843126297,
        0.5063331127166748,
        0.504877507686615,
        0.5069589614868164,
        0.5020797848701477,
        0.49762558937072754,
        0.5061172842979431,
        0.5037375688552856,
        0.504382312297821,
        0.5052663683891296,
        0.5058100819587708,
        0.5101516246795654,
        0.5137151479721069,
        0.5011551976203918,
        0.5018735527992249,
        0.5038512945175171,
        0.5069025158882141,
        0.4968290627002716,
        0.5035422444343567,
        0.5074982643127441,
        0.5067939162254333,
        0.504991888999939,
        0.49906378984451294,
        0.49331656098365784,
        0.4954744279384613,
        0.5043291449546814,
        0.5005234479904175,
        0.503092885017395,
        0.5045177936553955,
        0.500349223613739,
        0.504180371761322,
        0.50008624792099,
        0.5025625824928284,
        0.5027351379394531,
        0.5015830397605896,
        0.5055922865867615,
        0.3779904842376709,
        0.5083118677139282,
        0.5048509836196899,
        0.5058499574661255,
        0.4991986155509949,
        0.5021327137947083,
        0.4564909040927887,
        0.5012977123260498,
        0.5000917911529541,
        0.5018529295921326,
        0.4956408441066742,
        0.49577805399894714,
        0.5003799796104431,
        0.4926568567752838,
        0.503414511680603,
        0.49525153636932373,
        0.49582675099372864,
        0.4946942925453186,
        0.49872782826423645,
        0.5030063390731812,
        0.5146956443786621,
        0.5117411017417908,
        0.5056889057159424,
        0.5066736340522766,
        0.5055429339408875,
        0.4984591007232666,
        0.4973915219306946,
        0.4977109730243683,
        0.5022796392440796,
        0.5080822706222534,
        0.49775776267051697,
        0.5052640438079834,
        0.5023751258850098,
        0.5011683702468872,
        0.4988076686859131,
        0.5040997266769409,
        0.4872279167175293,
        0.5025384426116943,
        0.5063908100128174,
        0.5186021327972412,
        0.503650426864624,
        0.506378173828125,
        0.5042073726654053,
        0.5097928643226624,
        0.5112534165382385,
        0.505287766456604,
        0.4992828071117401,
        0.4989314377307892,
        0.49923548102378845,
        0.4954017698764801,
        0.5116125345230103,
        0.48136046528816223,
        0.49903470277786255,
        0.49963414669036865,
        0.5067113041877747,
        0.5219413638114929,
        0.5032155513763428,
        0.522392749786377,
        0.5080224871635437,
        0.505294919013977,
        0.5069504380226135,
        0.5102388858795166,
        0.508864164352417,
        0.5106931924819946,
        0.503582239151001,
        0.5062615871429443,
        0.5066177248954773,
        0.5060513019561768,
        0.5037457346916199,
        0.5061642527580261,
        0.5058395862579346,
        0.5023389458656311,
        0.501631498336792,
        0.5103031396865845,
        0.49161913990974426,
        0.5053682327270508,
        0.5034517645835876,
        0.5037616491317749,
        0.50731360912323,
        0.5023403763771057,
        0.5017844438552856,
        0.5017504692077637,
        0.5031646490097046,
        0.5009717345237732,
        0.5007362961769104,
        0.502945065498352,
        0.5056264996528625,
        0.5059395432472229,
        0.4982912540435791,
        0.5012227296829224,
        0.4971156716346741,
        0.5022655129432678,
        0.5045689940452576,
        0.5065001845359802,
        0.5060630440711975,
        0.5032998323440552,
        0.49809733033180237,
        0.5065878033638,
        0.5035482048988342,
        0.4768323004245758,
        0.5071566700935364,
        0.509358286857605,
        0.5151655673980713,
        0.502409815788269,
        0.4980016350746155,
        0.5005279779434204,
        0.497489869594574,
        0.49408796429634094,
        0.5006235837936401,
        0.5015415549278259,
        0.5035212635993958,
        0.5008845925331116,
        0.508043646812439,
        0.5081719756126404,
        0.5095581412315369,
        0.5042622089385986,
        0.5114695429801941,
        0.5068973898887634,
        0.5057308077812195,
        0.5004543662071228,
        0.5192793607711792,
        0.49835485219955444,
        0.49821022152900696,
        0.5009741187095642,
        0.5072320103645325,
        0.5032804012298584,
        0.5102436542510986,
        0.5101318359375,
        0.5057844519615173,
        0.5019133687019348,
        0.5085182189941406,
        0.5113158822059631,
        0.5051068663597107,
        0.5046374201774597,
        0.5026518106460571,
        0.507618248462677,
        0.5056539177894592,
        0.505316972732544,
        0.5187962055206299,
        0.5029204487800598,
        0.5144976377487183,
        0.5136324763298035,
        0.5061818361282349,
        0.5032070875167847,
        0.5018281936645508,
        0.5259370803833008,
        0.5068390369415283,
        0.5283695459365845,
        0.49931979179382324,
        0.5044167041778564,
        0.5367056727409363,
        0.5061662793159485,
        0.5066274404525757,
        0.5087254643440247,
        0.5020756721496582,
        0.501600444316864,
        0.47258031368255615,
        0.4913559556007385,
        0.501740574836731,
        0.5035696029663086,
        0.5074002742767334,
        0.505740761756897,
        0.5077281594276428,
        0.5027744174003601,
        0.5014389753341675,
        0.500482439994812,
        0.505942702293396,
        0.5040126442909241,
        0.5052940845489502,
        0.45336583256721497,
        0.5015352964401245,
        0.501572847366333,
        0.5075777173042297,
        0.5221095681190491,
        0.5136721134185791,
        0.5047352910041809,
        0.507327675819397,
        0.4976881146430969,
        0.5437337756156921,
        0.5453715324401855,
        0.5446805953979492,
        0.5051679611206055,
        0.5171785950660706,
        0.5313807725906372,
        0.5092065930366516,
        0.5077420473098755,
        0.5156874656677246,
        0.4864494800567627,
        0.4993816316127777,
        0.46372145414352417,
        0.5022254586219788,
        0.5072713494300842,
        0.5036960244178772,
        0.5086951851844788,
        0.5169922113418579,
        0.5089830160140991,
        0.5024601221084595,
        0.5012030005455017,
        0.4932343363761902,
        0.5127671957015991,
        0.5183271765708923,
        0.5121499300003052,
        0.5003659129142761,
        0.4986887574195862,
        0.5007843971252441,
        0.5060410499572754,
        0.5062897801399231,
        0.5077422261238098,
        0.5106645822525024,
        0.5090253353118896,
        0.5071545839309692,
        0.5009172558784485,
        0.49960029125213623,
        0.5056625604629517,
        0.5090703368186951,
        0.5065283179283142,
        0.41593173146247864,
        0.5020602941513062,
        0.5003065466880798,
        0.50406813621521,
        0.5053552389144897,
        0.5042799115180969,
        0.43804851174354553,
        0.49970993399620056,
        0.49975863099098206,
        0.5000308156013489,
        0.5057530403137207,
        0.5090508460998535,
        0.5050411820411682,
        0.5014115571975708,
        0.5017726421356201,
        0.5007053017616272,
        0.49999889731407166,
        0.5057416558265686,
        0.5038213133811951,
        0.5004130005836487,
        0.5020235776901245,
        0.5078807473182678,
        0.5171642303466797,
        0.5218828916549683,
        0.5116474628448486,
        0.5018895864486694,
        0.5003523826599121,
        0.5012491345405579,
        0.5011072754859924,
        0.5009588003158569,
        0.4653246998786926,
        0.5022298693656921,
        0.5074142217636108,
        0.5067388415336609,
        0.5169958472251892,
        0.5192270874977112,
        0.5186813473701477,
        0.5144098997116089,
        0.4932396411895752,
        0.516025722026825,
        0.5002179145812988,
        0.5025564432144165,
        0.49849820137023926,
        0.5053289532661438,
        0.5059010982513428,
        0.5068101286888123,
        0.5004445314407349,
        0.5019311904907227,
        0.49871161580085754,
        0.502483606338501,
        0.5017833709716797,
        0.5010204911231995,
        0.5130631327629089,
        0.5059536099433899,
        0.5020941495895386,
        0.4992450773715973,
        0.5003787875175476,
        0.5000966787338257,
        0.5017386078834534,
        0.5043586492538452,
        0.5010915398597717,
        0.5079344511032104,
        0.5030013918876648,
        0.5018794536590576,
        0.5021626949310303,
        0.49928373098373413,
        0.5004248023033142,
        0.4985007345676422,
        0.4998682737350464,
        0.5000333786010742,
        0.5036953687667847,
        0.5004358887672424,
        0.5007335543632507,
        0.500034749507904,
        0.5010448694229126,
        0.5107930302619934,
        0.5076003074645996,
        0.5016309022903442,
        0.503578245639801,
        0.5031302571296692,
        0.5052102208137512,
        0.5048073530197144,
        0.508957028388977,
        0.5010377168655396,
        0.5011413097381592,
        0.5177671909332275,
        0.5135860443115234,
        0.5147524476051331,
        0.49694550037384033,
        0.5057919025421143,
        0.5032265782356262,
        0.5012744665145874,
        0.505142092704773,
        0.5121508836746216,
        0.4992636442184448,
        0.49927791953086853,
        0.3957158029079437,
        0.4976697564125061,
        0.500293493270874,
        0.47780510783195496,
        0.5116304159164429,
        0.5101726055145264,
        0.5109753608703613,
        0.5034245252609253,
        0.5063373446464539,
        0.4464205503463745,
        0.5016245245933533,
        0.5047926902770996,
        0.5005135536193848,
        0.5047884583473206,
        0.5038897395133972,
        0.5011787414550781,
        0.501001238822937,
        0.5214022397994995,
        0.5021442770957947,
        0.5001478791236877,
        0.5008265376091003,
        0.5039764642715454,
        0.49871450662612915,
        0.502910852432251,
        0.5015245079994202,
        0.49575361609458923,
        0.5018402934074402,
        0.501829981803894,
        0.5017736554145813,
        0.4979928135871887,
        0.5008347630500793,
        0.5051556825637817,
        0.5039525628089905,
        0.50941002368927,
        0.5010154843330383,
        0.5005103349685669,
        0.5026417374610901,
        0.502254843711853,
        0.5002903342247009,
        0.49775004386901855,
        0.5175064206123352,
        0.5174649953842163,
        0.5199947357177734,
        0.5022398233413696,
        0.5046337246894836,
        0.4998517632484436,
        0.5063475966453552,
        0.49674350023269653,
        0.5031091570854187,
        0.5029646754264832,
        0.5075021982192993,
        0.5016157031059265,
        0.4990350306034088,
        0.5087488293647766,
        0.5097476243972778,
        0.5044133067131042,
        0.5016390681266785,
        0.503375232219696,
        0.4940550625324249,
        0.502334475517273,
        0.4988251030445099,
        0.49975183606147766,
        0.501307487487793,
        0.5010533332824707,
        0.5014610290527344,
        0.5003993511199951,
        0.5022254586219788,
        0.5067119598388672,
        0.5036711692810059,
        0.5069621801376343,
        0.504804790019989,
        0.5012905597686768,
        0.5022979378700256,
        0.4992677867412567,
        0.5015474557876587,
        0.5014985203742981,
        0.5070571899414062,
        0.509251594543457,
        0.509084165096283,
        0.5022886991500854,
        0.5139185786247253,
        0.5175769925117493,
        0.5028347373008728,
        0.5044678449630737,
        0.4979538321495056,
        0.5048726201057434,
        0.5090532302856445,
        0.5104854702949524,
        0.511025071144104,
        0.49729546904563904,
        0.5181524753570557,
        0.5083726644515991,
        0.5259069800376892,
        0.5210341215133667,
        0.5017369985580444,
        0.4998583197593689,
        0.5004023909568787,
        0.37524130940437317,
        0.3800579011440277,
        0.37697088718414307,
        0.5071531534194946,
        0.4986174404621124,
        0.5022183656692505,
        0.5062639117240906,
        0.5078623294830322,
        0.5084522366523743,
        0.4931797981262207,
        0.4962049424648285,
        0.49529311060905457,
        0.5016568303108215,
        0.5000752210617065,
        0.5022863149642944,
        0.517051100730896,
        0.5284121036529541,
        0.45427966117858887,
        0.5058892965316772,
        0.5026041865348816,
        0.5071789622306824,
        0.5004979372024536,
        0.5017630457878113,
        0.5049628019332886,
        0.504056453704834,
        0.501225471496582,
        0.5048617720603943,
        0.4974239468574524,
        0.49764201045036316,
        0.499275803565979,
        0.5133172273635864,
        0.5407994985580444,
        0.5447880625724792,
        0.502521276473999,
        0.5022113919258118,
        0.4353412687778473,
        0.5019357800483704,
        0.5187482237815857,
        0.5182384252548218,
        0.49830272793769836,
        0.5006638765335083,
        0.5005638599395752,
        0.509743332862854,
        0.5089824795722961,
        0.5026746988296509,
        0.49915680289268494,
        0.5029801726341248,
        0.5035865306854248,
        0.512532114982605,
        0.5098152756690979,
        0.5060230493545532,
        0.5053809881210327,
        0.5012643337249756,
        0.5014470815658569,
        0.5112341046333313,
        0.5064263939857483,
        0.5053709745407104,
        0.5035890936851501,
        0.5010142922401428,
        0.5002202391624451,
        0.5034596920013428,
        0.49796953797340393,
        0.49175822734832764,
        0.501118004322052,
        0.5041660666465759,
        0.5058860182762146,
        0.5247266292572021,
        0.5335180759429932,
        0.5281440615653992,
        0.5160920023918152,
        0.5197250843048096,
        0.525999128818512,
        0.5011073350906372,
        0.5009154677391052,
        0.5018761157989502,
        0.49983784556388855,
        0.5013691186904907,
        0.5014575123786926,
        0.5056528449058533,
        0.5037825107574463,
        0.5132095217704773,
        0.5028294920921326,
        0.4996834993362427,
        0.5063165426254272,
        0.5082049369812012,
        0.5085844993591309,
        0.5088592171669006,
        0.4793378710746765,
        0.5000401139259338,
        0.4973275661468506,
        0.5064547657966614,
        0.5045939087867737,
        0.5176864862442017,
        0.5041329264640808,
        0.5041984915733337,
        0.5056985020637512,
        0.505035936832428,
        0.5056606531143188,
        0.5041587352752686,
        0.5068693161010742,
        0.5088560581207275,
        0.5097446441650391,
        0.4991506040096283,
        0.5001731514930725,
        0.49990779161453247,
        0.5063710808753967,
        0.5075284242630005,
        0.5247361660003662,
        0.5051432251930237,
        0.5060253739356995,
        0.5043866038322449,
        0.5008167028427124,
        0.5056838989257812,
        0.4995962977409363,
        0.5074802041053772,
        0.5047229528427124,
        0.5123186707496643,
        0.5018339157104492,
        0.500941276550293,
        0.505850076675415,
        0.5027291178703308,
        0.5005195140838623,
        0.5043652653694153,
        0.5007486939430237,
        0.49879226088523865,
        0.5013629198074341,
        0.49876466393470764,
        0.5047147274017334,
        0.5119674801826477,
        0.5030670166015625,
        0.5000518560409546,
        0.5032029747962952,
        0.4997590482234955,
        0.5007814168930054,
        0.5014443397521973,
        0.5064855217933655,
        0.5007413029670715,
        0.5202938914299011,
        0.4987114667892456,
        0.5009844303131104,
        0.5032633543014526,
        0.5034523010253906,
        0.5052490234375,
        0.5170475244522095,
        0.506046712398529,
        0.5035299062728882,
        0.5167691111564636,
        0.5084772706031799,
        0.5100963711738586,
        0.5151799917221069,
        0.501897394657135,
        0.5037084817886353,
        0.5047932863235474,
        0.5006863474845886,
        0.5003185868263245,
        0.49774405360221863,
        0.5004344582557678,
        0.5101968050003052,
        0.5345442891120911,
        0.4961875379085541,
        0.5262426733970642,
        0.5332592129707336,
        0.503718376159668,
        0.5013817548751831,
        0.5064209699630737,
        0.5066655874252319,
        0.5051202178001404,
        0.48227638006210327,
        0.5056938529014587,
        0.5002160668373108,
        0.5002488493919373,
        0.5009569525718689,
        0.500316321849823,
        0.4029908776283264,
        0.5034173727035522,
        0.5017950534820557,
        0.5043558478355408,
        0.501910924911499,
        0.5044527649879456,
        0.5062637329101562,
        0.5072453618049622,
        0.49859926104545593,
        0.5066789984703064,
        0.5022022724151611,
        0.4921163320541382,
        0.496386855840683,
        0.5071248412132263,
        0.508571445941925,
        0.4778667986392975,
        0.4999915063381195,
        0.5093165636062622,
        0.503455400466919,
        0.5081764459609985,
        0.49851977825164795,
        0.514778196811676,
        0.4997222423553467,
        0.5026686787605286,
        0.49920448660850525,
        0.5115560293197632,
        0.5099359750747681,
        0.514392077922821,
        0.5105606913566589,
        0.5058308839797974,
        0.5068470239639282,
        0.4998343884944916,
        0.5017834305763245,
        0.4791451096534729,
        0.4998694062232971,
        0.5006538033485413,
        0.5062400102615356,
        0.5045655369758606,
        0.505126953125,
        0.5084541440010071,
        0.4967910051345825,
        0.5007727742195129,
        0.4673203229904175,
        0.5040602087974548,
        0.49878209829330444,
        0.5069664716720581,
        0.5090833306312561,
        0.5080189108848572,
        0.5005597472190857,
        0.5052945017814636,
        0.5080094933509827,
        0.506752073764801,
        0.496577650308609,
        0.5004936456680298,
        0.4537154734134674,
        0.509322464466095,
        0.5067769289016724,
        0.5025783777236938
    ],
    "eval_runtime": 52.0612,
    "eval_samples_per_second": 14.406,
    "eval_steps_per_second": 0.461
}