Files
GENIES/results/evaluations/cra/llama-7b-code_easy/code_hard.json
T
2023-11-11 19:44:09 +00:00

762 lines
21 KiB
JSON

{
"model_dir": "models/llama-7b",
"distribution_id": "code_hard",
"timestamp": "2023-10-16 03:12:41",
"eval_score": -0.20125630779949452,
"eval_brier_score": 0.20125630779949452,
"eval_average_probability": 0.648183155400546,
"eval_accuracy": 0.708,
"eval_probabilities": [
0.7609667459214036,
0.24194455234859572,
0.16989501326017875,
0.6808680523273165,
0.5158028995119683,
0.5935491722035372,
0.8333000946744872,
0.9321447057822994,
0.590392489000511,
0.6164653403929407,
0.9877460266582603,
0.8889401391101079,
0.7936457962067766,
0.6607248241094703,
0.34857412901854146,
0.507523673304505,
0.8538608198989938,
0.6742771711804998,
0.967693512890976,
0.4315889498654635,
0.617178132568255,
0.978507416172474,
0.9608497851686639,
0.9819239299909316,
0.8914661637148508,
0.9108290767080788,
0.7509061809258418,
0.5997850790930059,
0.5831026851842016,
0.7108852701645691,
0.7864586214227081,
0.5793081161192384,
0.29486595919132674,
0.9302038659796957,
0.9347306974268811,
0.8279555758497175,
0.569082791520314,
0.6498959276836734,
0.7035608567518259,
0.5537759958157024,
0.5397438856573775,
0.6633516574054055,
0.9819119533473911,
0.9201328977431604,
0.9611005172848464,
0.6346789625646169,
0.1193072795940503,
0.8435900350338318,
0.9847478264158657,
0.9528617245033676,
0.9549029254561586,
0.16380366312716488,
0.9607913358291592,
0.9687359499449505,
0.8553235972516258,
0.8985739131386813,
0.8677954471976786,
0.20013719958697757,
0.7777920817273327,
0.03954346302547727,
0.2992214388978221,
0.12576852803439814,
0.08309012364525781,
0.46249265982257814,
0.31664387805508853,
0.46165639023588534,
0.1394395994699055,
0.5619274798107191,
0.09829580913463681,
0.9202931950856695,
0.7715694026714914,
0.8469967569714103,
0.41881681622915923,
0.9940577061121822,
0.7598824935786876,
0.5152609410771741,
0.28257703544173096,
0.7033030907775293,
0.46982301708903895,
0.32441246266491763,
0.5035206142998729,
0.34974802661644183,
0.7253290053371826,
0.7665049560960827,
0.5210362390038262,
0.7164078720790722,
0.6575091844171002,
0.831713141042286,
0.24029274540323786,
0.19959702574377516,
0.9633939488222572,
0.888472941431049,
0.7604024794832427,
0.7145608666781613,
0.6175552629549493,
0.6910717513619242,
0.6791065813940579,
0.9707897852985067,
0.08234914072224536,
0.602254507787441,
0.9401744169720796,
0.9814449808696227,
0.2781012869755798,
0.1633779626199949,
0.20910294705031535,
0.8039682037609619,
0.9482841692616957,
0.8865587961022355,
0.8672193355484263,
0.41607226440852507,
0.4193719623147755,
0.6219922178072879,
0.5005758043121732,
0.3932127604189417,
0.5056250788832765,
0.22802839439082692,
0.3470158485320731,
0.8039515178427841,
0.5385855480717147,
0.7390781072868884,
0.6937432001952248,
0.40508235719804886,
0.0939045482914902,
0.2800587875290519,
0.20581384540073022,
0.49496969230613497,
0.9317685126459074,
0.7371109468236867,
0.4456207586641525,
0.014660130910106995,
0.009030169493304733,
0.007543369498045911,
0.12828583981738742,
0.24951026191710335,
0.2973174124856288,
0.931045773334582,
0.7898386507243049,
0.9101661678985117,
0.38538602789455156,
0.6372395785118159,
0.546627374215434,
0.5148160229404838,
0.29275058065008663,
0.40879048957434116,
0.7631974951322635,
0.8918756137301116,
0.9730214456099529,
0.7268532965707548,
0.7322878982188369,
0.2783943289596998,
0.5694693134625521,
0.5560439705341146,
0.7482823677992181,
0.06831923018711265,
0.09827732077478864,
0.058481938324490876,
0.6235215327101167,
0.35545365857592454,
0.6629339268922926,
0.17875463759556937,
0.2226118581843345,
0.1639162746030952,
0.9305779603654294,
0.7507686065267613,
0.42124353375988566,
0.6128259248283074,
0.967303495004081,
0.9536914958128208,
0.3184077247302475,
0.19320564180808175,
0.0884213063977757,
0.7916030249191432,
0.8363673348141071,
0.4865036058711045,
0.6730613409207776,
0.6605189482679351,
0.7147376628202918,
0.7249489122068027,
0.1784743167898628,
0.28428375605750356,
0.7969234441262387,
0.9736941692713974,
0.9673689835845887,
0.692096978449282,
0.8666816075477526,
0.7288646506187964,
0.059378652176816545,
0.43944544017748344,
0.9008717116192994,
0.7484583124177199,
0.9850107655144261,
0.9452802209910864,
0.8836758900540755,
0.19499724582692296,
0.30673017752260034,
0.7292970219971682,
0.6534929407278874,
0.8231400485557355,
0.8027214224005622,
0.46324525988825244,
0.3297470599743637,
0.8792097188596829,
0.7986247698420591,
0.6178625330556652,
0.5267028626345921,
0.7309466767861472,
0.865369069167656,
0.7112039184225369,
0.43594313039202337,
0.4387341496874574,
0.2955895487604397,
0.23540695441772502,
0.7404494218299164,
0.4808001373404163,
0.818282670055976,
0.4326230535852861,
0.8537502429991082,
0.6796554181373482,
0.5297698376919813,
0.862074948381015,
0.8563892056830845,
0.8488322830779219,
0.8307792656909504,
0.8138849081571563,
0.7510907217278906,
0.28526816366423907,
0.34935540023468414,
0.8861550405193324,
0.6502127306247196,
0.8143905670545153,
0.7294011661145456,
0.5641175011513239,
0.908251375906229,
0.6984628401691622,
0.3188734579958145,
0.3147648205818502,
0.43176325151745765,
0.6847906100233434,
0.6186772638803079,
0.6658780983951396,
0.9511396915192469,
0.49584120935059384,
0.7196597849908719,
0.9251017284691517,
0.9539282761049529,
0.9563970863325733,
0.5236543728999206,
0.40658247912502987,
0.16639526100049773,
0.897402027589502,
0.9406515524026489,
0.9597206687059753,
0.2505834516496459,
0.17020744752686512,
0.1579983569213522,
0.24608531653069465,
0.26142150520616403,
0.4440380640789504,
0.21258055468993456,
0.24256040028956238,
0.21316953787388704,
0.9320667004544779,
0.9455688311173912,
0.9658206147815231,
0.8858643088898835,
0.7751391818960249,
0.9310869882897195,
0.9551461582670577,
0.9664037254556886,
0.9840501679554685,
0.7553464983539765,
0.9148266440058818,
0.9355473499564902,
0.09027876471502641,
0.9907768594768144,
0.49031082038607104,
0.8891243273551682,
0.8975870970935338,
0.9164640000603773,
0.2869020264632077,
0.24902627254692605,
0.8119752519407429,
0.9915927467674462,
0.9863038507007076,
0.9894838964539503,
0.9546312093858798,
0.8311812817751251,
0.5358308555710635,
0.2961736207229586,
0.29328359733588605,
0.590917565576298,
0.1118282275471413,
0.42761245298717054,
0.5290792178271961,
0.32634741438731096,
0.18315186706290854,
0.3674461232226651,
0.5956793154072992,
0.7251964954130976,
0.38858964362253035,
0.9571563299973784,
0.9180983344912163,
0.7386410228460811,
0.9216389556669887,
0.2636565392824686,
0.9540874323640658,
0.8635895527552317,
0.8505905017752684,
0.914605796256452,
0.09215441221392284,
0.05513602411628151,
0.3970233311510961,
0.4287934924780758,
0.10406400476629415,
0.6791285847201185,
0.8682624804378329,
0.7611749751188055,
0.19001467807172653,
0.9599965157546717,
0.9472770913499281,
0.9391427810608381,
0.4213680477543462,
0.12838813689996975,
0.24384938242825713,
0.8818502720981959,
0.7724834632017414,
0.6590705561930286,
0.9063375220933666,
0.8103522810065785,
0.9963711759084751,
0.43302338927033124,
0.5244673654679066,
0.2188426669232022,
0.9377826391353495,
0.4950935126597608,
0.6353870055000642,
0.2759763172289674,
0.6630497386411577,
0.27434356407257615,
0.8520414560107543,
0.8516950336002809,
0.9464261452573287,
0.6194120939849221,
0.9924000264991591,
0.5312550387287618,
0.7238378781934056,
0.7458098584608361,
0.49965042809101246,
0.9544687153018986,
0.9043996323350576,
0.9142108060848305,
0.8270625522283674,
0.9001989619084348,
0.8929212550123763,
0.3829947285320914,
0.29295968287709645,
0.30525851292520084,
0.986554109603778,
0.9805819230958528,
0.9768784064487211,
0.6548491021571848,
0.16398717337974153,
0.7206286835165392,
0.6389080905785555,
0.6634952294943118,
0.8514978441955029,
0.8874448094141256,
0.7674464100380528,
0.8304580087163871,
0.9487072540850043,
0.6555385934560632,
0.667644522576461,
0.8805161663041147,
0.9118029382992082,
0.8276299539054474,
0.7750064966106075,
0.6601138839963663,
0.9086746996588421,
0.7004534830010027,
0.5880964177840097,
0.6748997049762215,
0.29732776101923314,
0.24993892965894024,
0.3594598292870754,
0.875615567579338,
0.7714058381110508,
0.8696790492851896,
0.2071653452296952,
0.5584294081694877,
0.760671918999266,
0.5684713721965705,
0.5373477334628749,
0.9853414651534345,
0.609740654129889,
0.43128145137152774,
0.6689353828118159,
0.5338165244817306,
0.2655453079723536,
0.7087089123532034,
0.2899249111087195,
0.46649355030204187,
0.7610878630686742,
0.9638209910026986,
0.9520573272511698,
0.9667735491625505,
0.9817331818041285,
0.977749335242767,
0.9479922606170526,
0.7297398290356949,
0.573814274075013,
0.9619706692787794,
0.18878827182913757,
0.33441909244344337,
0.2192769230472981,
0.8797099920442887,
0.9005981588806679,
0.8849818315015396,
0.6466934855880071,
0.7590845447594073,
0.6638981334659884,
0.7966848343590206,
0.9104800042399755,
0.9320864809192747,
0.449632934330973,
0.710528658041782,
0.4180805723769827,
0.08427234594400049,
0.21362619983881126,
0.08774439411699828,
0.8133614345932204,
0.6470579187705826,
0.8551484824610682,
0.7320846742840753,
0.6773560548004304,
0.3263779897118192,
0.8557199275745507,
0.9086253616968889,
0.8068755860559056,
0.7499967522084698,
0.7396209213716158,
0.7441597843784691,
0.8424865150403492,
0.81362822418554,
0.893496963859568,
0.7826611938322507,
0.5862189110592171,
0.7328828373485266,
0.16340871090677744,
0.3134638982759463,
0.4233369078215574,
0.8470158724284126,
0.6086357673494929,
0.47031894547589337,
0.9167580027477278,
0.22633692178229972,
0.789374671870706,
0.7924090858452475,
0.15909352844430844,
0.4466048884723334,
0.5511933375447367,
0.9151479102592575,
0.9494367138999459,
0.07833216537250998,
0.7723880226731231,
0.282028541328857,
0.7440388026516802,
0.9863944385078699,
0.9097944227545942,
0.23883229781669163,
0.31840548012987646,
0.35803852541352543,
0.7414951637967776,
0.538705833950583,
0.8101365746555095,
0.6896162233092394,
0.6612636024150667,
0.4601794792221855,
0.6393009588309408,
0.7284832852392172,
0.9508737522628326,
0.925327338197487,
0.7223899238071699,
0.46507665277520627,
0.1071981141748759,
0.08377559123287781,
0.08183349278320882,
0.6263073721500532,
0.8054052626059492,
0.7504965414992475,
0.18909262318423872,
0.4560824284392062,
0.9841299854659489,
0.08824717302067894,
0.3500524907234668,
0.4334778148700124,
0.6462295225924456,
0.26162828436457697,
0.8182989381424602,
0.9434080704805526,
0.8938917010529,
0.9124668001816777,
0.7298565029308324,
0.2988916010689575,
0.9522201156457071,
0.32097891899341285,
0.8709464132770129,
0.08545911063796108,
0.6627496029486368,
0.6597953698198948,
0.6546878254674875,
0.3420124403298237,
0.547816891454386,
0.3608460036840394,
0.17911180004055471,
0.6015615174727154,
0.5215230654382337,
0.8039893707141091,
0.685547563867229,
0.6800362529655424,
0.936226585076551,
0.6549560672456967,
0.6946867096485387,
0.8791439483253212,
0.8630860540972782,
0.7796908556534672,
0.935083196884654,
0.9111818546873786,
0.9896297210873896,
0.9335125616394844,
0.8619522249519567,
0.988573714626054,
0.7277776749545575,
0.46664673450151595,
0.34038126474367303,
0.21049515123976328,
0.11762927774301304,
0.2556928286761375,
0.8683053877675305,
0.5349920701237614,
0.7156377704320693,
0.8965847444134097,
0.8068629687961348,
0.9753885958403357,
0.8435945596820319,
0.8652906834903649,
0.7281143798218713,
0.29290861090635156,
0.882070805735071,
0.9681347912778316,
0.6557753744564101,
0.7108343933797094,
0.6809877033048755,
0.5771220621619626,
0.8721384815837734,
0.9669073709081634,
0.9199820636270692,
0.9305083904561979,
0.8715325198548064,
0.964027720125954,
0.5115015581750488,
0.5616880546863268,
0.812964319620895,
0.8778290006829228,
0.9253056716285967,
0.14906892188718135,
0.1425915508340354,
0.13853497409757787,
0.9543689701372816,
0.9174581726639295,
0.9549705196700131,
0.992794444980678,
0.964433879216108,
0.6276623155334775,
0.014691331716927492,
0.49168586790660956,
0.6827896817174637,
0.9388973374809811,
0.7781075214403275,
0.8634346235720473,
0.9758382685994564,
0.9758042192860235,
0.40457903727447925,
0.9695599628979688,
0.9952104664553767,
0.9895357868143428,
0.6369061321531773,
0.28821459839981456,
0.7085254069550441,
0.5149992612542845,
0.44782419806067525,
0.6181805010883544,
0.7818570651789657,
0.8780317256997854,
0.9109656994187622,
0.6899554249092175,
0.5237774178796315,
0.9717219127937988,
0.9424009770952657,
0.9287105341521141,
0.8971412423968876,
0.7509656857259683,
0.6831613403857524,
0.13299645753409853,
0.6868365455404232,
0.7120546078051168,
0.9910052387920201,
0.9549517964435146,
0.8664366373013512,
0.9776229695418658,
0.6678744812942468,
0.7183476464811684,
0.35369382579025566,
0.06310795111971679,
0.02894891349250206,
0.04609836274419943,
0.9926526784673266,
0.9882102673281089,
0.9879812092043503,
0.9431750982246204,
0.9335479755458984,
0.8299649586281128,
0.7837374046301608,
0.8426380001888937,
0.2813123175655066,
0.7489991893467622,
0.9311309205684482,
0.947412139282352,
0.525051752244098,
0.21049926714089484,
0.6639334160819961,
0.8835388213656741,
0.983553432247489,
0.7585458337225387,
0.9319671226176798,
0.9206347656141423,
0.9348594115528938,
0.6622314007222595,
0.8949403977927949,
0.8461825786057738,
0.9005994379845901,
0.6729120809155403,
0.9545137334397801,
0.14644230226756866,
0.9691098502462763,
0.9026523620990681,
0.261835537314141,
0.702522360967702,
0.08870365615088815,
0.9278088899518523,
0.8477638511538935,
0.7662255868178899,
0.905019473790464,
0.9030717394875726,
0.8297788293249304,
0.6534523885346651,
0.5932177859848278,
0.550527823550965,
0.9084244306176261,
0.9661111015453898,
0.9128807112344137,
0.6977589981775658,
0.555755694113991,
0.29951314073135044,
0.9391557370036057,
0.8657612111411328,
0.9427077042007725,
0.6304338365715358,
0.5836826635352753,
0.9585703046532212,
0.7399123103859757,
0.5802190596458235,
0.5500108623181733,
0.3836049303392427,
0.7684791192704544,
0.544627806565091,
0.19175554759426086,
0.0454396381013635,
0.15015358469446982,
0.1614982989066253,
0.37144788046404775,
0.2355058963007669,
0.9918860726578104,
0.9691744584337569,
0.9780580742331236,
0.6602642958730388,
0.7884305026767802,
0.8827899678232829,
0.9412337683814423,
0.18840506843678417,
0.1714606504375133,
0.4578419699312799,
0.6157854345892305,
0.45118026489778396,
0.9622401695834794,
0.9391643365829404,
0.9414343409900524,
0.468448969153889,
0.6782470000928683,
0.6697706504628764,
0.972669631377518,
0.9732890256383601,
0.9547963520695597,
0.9577997994812919,
0.9424617977392408,
0.9251556903166929,
0.95225520143593,
0.45654653059578437,
0.8837690737987257,
0.8514067778594121,
0.634863393709372,
0.9374947906534981,
0.9313202224705054,
0.8575062633244128,
0.9361801684154094,
0.6947400729713021,
0.9050100570374717,
0.7274922392764744,
0.48134326859642157,
0.3629198602340617,
0.8418487545131663,
0.5844210927665507,
0.8195084637997195,
0.9263586142991528,
0.7136141360774206,
0.6783900013165143,
0.2809769577616779,
0.5940845352338837,
0.8628402575249661,
0.8394963090175116,
0.9217352135826163,
0.9747159762862757,
0.9754590675419706,
0.9248437744885374,
0.9535446710011016,
0.9920299316726252,
0.12405955357792399,
0.23735558544762492,
0.3153689959560394,
0.31237929878685755,
0.363803295488637,
0.7923736933087444,
0.24957713172656404,
0.1462553471359904,
0.15025839397530039,
0.9357939006473937,
0.9325916054198435,
0.9491781845437208,
0.8530288469555186,
0.7659364222205822,
0.5537717923658096
],
"eval_rms_calibration_error": 0.28920527449290784
}