Files
GENIES/results/evaluations/cra/llama-30b-code_hard/code_hard.json
T
2023-11-11 19:44:09 +00:00

761 lines
21 KiB
JSON

{
"model_dir": "models/llama-30b",
"distribution_id": "code_hard",
"timestamp": "2023-10-16 03:58:45",
"eval_score": -0.1334372951555614,
"eval_brier_score": 0.1334372951555614,
"eval_average_probability": 0.7362497092942535,
"eval_accuracy": 0.7986666666666666,
"eval_probabilities": [
0.8824457256499152,
0.8881720951578254,
0.8851485096619639,
0.7388402287492402,
0.6176371786035809,
0.4972160959925112,
0.994525850977727,
0.9997676107183779,
0.9935845904610504,
0.797657683578919,
0.9587217653237139,
0.5419652193639839,
0.9978996078605208,
0.9996861823161843,
0.9990682927489047,
0.669084515783624,
0.7520615967430649,
0.8526906235688417,
0.8299333324186342,
0.9908958995861157,
0.9835700442293178,
0.9880431216706049,
0.9984302909148559,
0.9989245424366238,
0.9978766543271567,
0.9997459854758276,
0.9991948000097618,
0.9981697617240867,
0.9990953956207593,
0.9986871789049447,
0.1666694715917879,
0.4017959894637052,
0.28244997006947103,
0.9588580412539249,
0.9449643867413615,
0.9499775381379276,
0.9997362507181415,
0.9990852720041066,
0.9994900280974564,
0.8975205896942637,
0.8932409426772963,
0.9204877611895157,
0.9997267398404794,
0.9997402609110593,
0.9997176879011744,
0.48424211592363936,
0.33210227304480955,
0.5762981208693823,
0.5104076611117719,
0.5381471311872937,
0.6971111126276626,
0.669299883462682,
0.33477087943973227,
0.8017589339473328,
0.7912266461603787,
0.8444720508544731,
0.8253340006350192,
0.3118307750735441,
0.614538822779475,
0.6302028525490315,
0.5647044228853009,
0.3474054459896302,
0.4490580362728691,
0.5877282252523495,
0.5860431579353366,
0.3967505771700981,
0.484954087167741,
0.7021366803564999,
0.5939018657243611,
0.9473674929321019,
0.9314738747349357,
0.9386672345324544,
0.956183921633118,
0.9993217941503445,
0.9995984818352621,
0.9215250393838048,
0.4725746262200191,
0.8902575369557507,
0.6605091300806641,
0.6327518320663321,
0.5855113118278288,
0.2885168469802081,
0.5464554871190985,
0.16247883073196207,
0.9979706765601709,
0.999672779944757,
0.5590391135729673,
0.8486249771402374,
0.7840836010048031,
0.5192030082593669,
0.999430359407696,
0.9997110821774353,
0.9993266917872133,
0.6285481486021672,
0.5738145877738858,
0.6914918802938984,
0.7732747904447579,
0.7194187275403789,
0.2984031617107301,
0.7085678386887124,
0.6460144655955715,
0.749670640915268,
0.08149548945195362,
0.06871533708167926,
0.06992649659912244,
0.2023661045023507,
0.7649022965360016,
0.38992381559747924,
0.8736984560402826,
0.5354381510211537,
0.5249132930613318,
0.7187806479087238,
0.5204297415363417,
0.5122694007596635,
0.8068900234317288,
0.5244073458374072,
0.729706578652533,
0.919555879294309,
0.9294209821496197,
0.9657478578024722,
0.4568206575035796,
0.457066769826201,
0.8221728673810533,
0.32548065014440763,
0.6288598871067578,
0.90436708410395,
0.9995086715548673,
0.9951052262594554,
0.995796007911519,
0.997023516117536,
0.9975857687593332,
0.9991288540293727,
0.20775990020797883,
0.37002533463805354,
0.24424222914660948,
0.9998068457946594,
0.9929882032990839,
0.9998626263108074,
0.49625629056740483,
0.5295838742630599,
0.5341980481409612,
0.7299593965167972,
0.5104531334577155,
0.6015748446157728,
0.6974672722610367,
0.5349950260374174,
0.5333642149788058,
0.9998209815004391,
0.9997849344861931,
0.9997768322254347,
0.9618470977970477,
0.9651998498948643,
0.9666404590139525,
0.1745547031344451,
0.09163631659448321,
0.07349509576962515,
0.9996281713658329,
0.44871018005313196,
0.9992934176953676,
0.6318051763304771,
0.5976767339074371,
0.7182216130059548,
0.8651733606659888,
0.8370758987287927,
0.36716010710823915,
0.1911866600073725,
0.7421709472654958,
0.37036144384116715,
0.6372470286301293,
0.6228528525656009,
0.7002925482664428,
0.6427357575020065,
0.3726267443356145,
0.6486392172262695,
0.9986596553695343,
0.5774724034770289,
0.6165583125613037,
0.8298670962430997,
0.7949785813325357,
0.7876105835379751,
0.9996477309103763,
0.9987555422457456,
0.9970161208259636,
0.37820736787065656,
0.3977566738454775,
0.8402363315184942,
0.8036268618167713,
0.7962350001039173,
0.8329253372584686,
0.5589931769103725,
0.7670239554023806,
0.419236534688333,
0.5393502561465725,
0.3273616188846285,
0.22962048602265384,
0.3622017949291051,
0.38522430569990346,
0.4490784999587407,
0.9997920731087865,
0.9995734023635767,
0.9964177019490287,
0.999852369922919,
0.9998890296324116,
0.9999171288166231,
0.6449276927708003,
0.3716028443368158,
0.2843716680588817,
0.7805383292170007,
0.9295571954626937,
0.9218394605176441,
0.6555119422670702,
0.7206173315682728,
0.824217742430899,
0.604645130369865,
0.6080209609903626,
0.3129141727993612,
0.7159116162556773,
0.722199707991078,
0.4386765922579461,
0.9995017632353672,
0.9950631300302643,
0.9995756409337059,
0.8718952747376012,
0.7003717671914629,
0.7091383266007262,
0.8615831592088454,
0.34071518662181854,
0.7219213150807396,
0.8766491625712983,
0.9326091911723113,
0.9157345268477782,
0.5070095775684986,
0.5818606583371343,
0.8978977314973612,
0.11034942841757506,
0.12413209046329599,
0.12517877930286253,
0.45102552660236483,
0.24900735963106102,
0.1882864362083363,
0.9474320189914547,
0.10215833935435528,
0.7158929583738576,
0.9997346254672513,
0.9986281437223179,
0.9986896255060639,
0.6699172195557287,
0.649218232395712,
0.7458536018921134,
0.9997620535618733,
0.999569065337001,
0.9997373942352313,
0.5525660383882263,
0.5557365334629571,
0.5830836898414367,
0.8395982928024529,
0.7415774064787422,
0.8548820338941786,
0.7038702602671063,
0.7555448167241307,
0.7577056520532183,
0.9989258754458294,
0.9984216491115911,
0.9993786576419172,
0.46942959676737017,
0.7316801483154336,
0.7986790584964951,
0.9999322517490357,
0.9999350034785265,
0.9999396079397469,
0.8961576277121592,
0.7038661565810986,
0.8144131381170504,
0.32222140194194077,
0.567646141401168,
0.2349756233301789,
0.952793789360116,
0.9988035521611255,
0.9990280179803419,
0.9772132196785195,
0.826004633412347,
0.6509526071400467,
0.9996912133467776,
0.9996554355741204,
0.9996075541309676,
0.9992797700985484,
0.9996061600070792,
0.999356918099644,
0.662558952257565,
0.6543442848950881,
0.7970818152930137,
0.35757302581808115,
0.4903575021114065,
0.4684504215981018,
0.8380574458149361,
0.8838381924598369,
0.9448840866911044,
0.5153141066352613,
0.6671006713376492,
0.6671437528502769,
0.9988108585742326,
0.997430820326485,
0.9545774903258322,
0.9926771970634248,
0.9252706749179341,
0.976420465812612,
0.999220781371935,
0.9991392591272762,
0.9975753202103983,
0.39939188882268883,
0.6361460306852226,
0.7253282909631915,
0.7920088335998681,
0.6497410242226337,
0.5156480562052,
0.8520728724971842,
0.41613635604415333,
0.16282682447124217,
0.9749484458538858,
0.9229505713806758,
0.9606377280907503,
0.9423784170351281,
0.38410266739203186,
0.41003658646781405,
0.6528113667437929,
0.5708878336398566,
0.5193381557333324,
0.8033956993538819,
0.8845218195615184,
0.7934314338418409,
0.18368600689134829,
0.30875475562151555,
0.29412814811582433,
0.7947501870007029,
0.708701074599945,
0.8573262183465142,
0.3121409633235642,
0.1598126116804265,
0.31865272732720545,
0.26467962225208136,
0.3976462962466086,
0.4362153913811375,
0.7803186370927991,
0.7516104073696541,
0.8685450311223576,
0.3363201407550265,
0.40039085050466183,
0.22721810499332118,
0.9997406614279317,
0.9995027605350935,
0.9998323835768571,
0.8790678377618037,
0.9638288686730592,
0.8668690704316947,
0.9558564486567575,
0.9472569281728065,
0.9541541256418801,
0.9998505474683999,
0.9998317128199128,
0.9998486433794059,
0.6558293663857708,
0.7926136461306388,
0.7370631680414708,
0.49400596258689666,
0.632382718965002,
0.5734033653539832,
0.9750777566591591,
0.9317239915732193,
0.9694028414735485,
0.8212219096618687,
0.33679274002357606,
0.7300269798161076,
0.7338446467325754,
0.1278700666788427,
0.6750688336324736,
0.2629215043677331,
0.9521036296227401,
0.8111564802490954,
0.9765771506250003,
0.9634918727719175,
0.9625080088582615,
0.7876653174285831,
0.8048204751254855,
0.8383057853739339,
0.9506275021713966,
0.964787646763094,
0.9301146200718918,
0.6809295201299397,
0.7980590956567453,
0.6875856112375125,
0.6788855824779582,
0.5759951411104492,
0.999474666014341,
0.4544046739981149,
0.6968177855893237,
0.3451094488420351,
0.8905465618170224,
0.8910513513619982,
0.929781230636189,
0.9998461313955368,
0.9998016762661487,
0.9999230351533634,
0.9995380779015772,
0.9997996854364529,
0.9994824289823421,
0.9998815071663164,
0.9998086340249102,
0.9997979187595429,
0.9995859430824416,
0.998841899755077,
0.9989684403881902,
0.99772968400641,
0.9846445151179608,
0.9899367266723422,
0.9993966731537158,
0.9992411408448779,
0.9980143951671183,
0.8392547883749621,
0.9521108642997674,
0.8813782869397766,
0.9997057574662117,
0.9997696637721641,
0.9997872175050215,
0.27842286271725697,
0.7037050101525825,
0.20495365334006435,
0.9964651983696147,
0.9723186408105704,
0.9967661035745764,
0.9994267124543543,
0.9990944885943471,
0.9990810607557843,
0.997958445136105,
0.998609034925759,
0.9969402053020618,
0.5913509414178613,
0.57682917324539,
0.7830837685287542,
0.8340613460036841,
0.7021632567016649,
0.7561540999130771,
0.9997448062439378,
0.999416953092081,
0.9997321140290704,
0.5778743034225248,
0.3463548710041224,
0.36106895762633207,
0.4516731451095336,
0.6407899384957656,
0.3847800631791648,
0.9141972196929999,
0.9278397077349269,
0.906111866337818,
0.53588439933195,
0.66604079073545,
0.8665000020312776,
0.8019743512069486,
0.5610582463511673,
0.6834493883168903,
0.9516342297465157,
0.9444932762379136,
0.815083423535889,
0.5745100524366598,
0.5668492941249168,
0.5267282771299924,
0.8966616759604699,
0.9932184115438554,
0.9602295200124213,
0.9443110288136627,
0.947173270658532,
0.9424157253552811,
0.9113910856058,
0.773772960429835,
0.948649647314568,
0.5283604947295584,
0.6351108390494865,
0.7073880601188394,
0.8956752727156292,
0.7403961845625697,
0.4748437246485168,
0.45208548681671024,
0.6787590514839761,
0.7239210288693312,
0.40695041634484086,
0.2793720715477065,
0.2082816829661083,
0.9996298892578025,
0.9990098773639949,
0.9997210349633279,
0.58261566371888,
0.5815625772839943,
0.788812744876603,
0.8294219318833449,
0.8336592261959851,
0.6365229314557772,
0.9993223958398684,
0.9989496110664029,
0.9992298301202456,
0.44056977311581913,
0.38768398268471627,
0.39466181792008104,
0.8835088244635854,
0.3822367486620748,
0.7715526869232956,
0.9394792246760583,
0.7908881975745589,
0.2558764631626439,
0.9720690258938571,
0.9720591021533429,
0.97390911370652,
0.5706857483813044,
0.8285221462858054,
0.7648616181228316,
0.6942777465757637,
0.7671345479172348,
0.7599728608694806,
0.9993758257966872,
0.9998225101889526,
0.9995268148897726,
0.385549884437246,
0.43444167677807366,
0.4983180361382448,
0.6857537825098168,
0.7057135268527227,
0.6979066184820679,
0.7382969367746959,
0.6707465949281171,
0.9394177079524932,
0.9945335744187993,
0.9910488738149332,
0.9835006724544209,
0.9217859157654827,
0.805902020884232,
0.847955640584462,
0.6934569602917637,
0.7767479503225785,
0.607202385015975,
0.8041790091480485,
0.8100738302996784,
0.7264501091449782,
0.9997481872858606,
0.9993062496337027,
0.9996125870173937,
0.9995101726083869,
0.9996868835607688,
0.9982133130657433,
0.7957124517281084,
0.4557317858352209,
0.552262467804124,
0.9434127893454901,
0.23238384052061933,
0.8795814312142862,
0.6319332091892224,
0.8073640792498,
0.792417298819557,
0.7027084775939415,
0.836362834878442,
0.8944488718691598,
0.8105532213114255,
0.8057968462289554,
0.8251041873243393,
0.9995676332879829,
0.9995813217173554,
0.9996971720371415,
0.3502764306121178,
0.36360163956572156,
0.4188969272827734,
0.4148091359415258,
0.6658187779619343,
0.999215987429465,
0.9993523050107851,
0.9908832411129743,
0.8663510021584603,
0.23064709511101644,
0.6199692480161129,
0.9855537849092626,
0.18011810866530875,
0.39026476860516424,
0.49028406224591226,
0.9171299782979608,
0.8865345614643615,
0.684213348989758,
0.99970008915794,
0.9993048528212554,
0.9997010266734101,
0.2758696506354964,
0.4435282029755816,
0.327897111700554,
0.24564455938748053,
0.19304240244187884,
0.17714885107653952,
0.8164271664999404,
0.8138335028898243,
0.7552466314014541,
0.42809745975665614,
0.27620808161409416,
0.4244287591180015,
0.999578737747118,
0.9997567418684771,
0.9997416237292026,
0.5192232805600892,
0.8438430102722053,
0.45833049870203224,
0.9168731318417636,
0.8912721898543784,
0.8041303158916641,
0.518606977866199,
0.8894640396305876,
0.9153835695681748,
0.24023129820296618,
0.999623827223495,
0.16286811729620698,
0.19932330635338313,
0.2066195925274349,
0.280807000498719,
0.9201173101899689,
0.873271359482698,
0.8765489525434189,
0.6691587645628455,
0.867162093447605,
0.6435413824488915,
0.27588680248215225,
0.6796378771244546,
0.6998599470561778,
0.9997685979176881,
0.9998414582783094,
0.9998572066327087,
0.8400560909610628,
0.6994273055900334,
0.6752208577607884,
0.6113072323506167,
0.6699448204978657,
0.6814044391328263,
0.8098005180535869,
0.65833796696366,
0.694391560905788,
0.8195264835091954,
0.8731509808071154,
0.9042506118269624,
0.6004942574205436,
0.675300363623454,
0.7175050429532315,
0.9262599465392737,
0.9950574973549742,
0.9991654455662375,
0.7870442567021623,
0.39040470040547004,
0.6969065788168227,
0.9677848662195212,
0.9493695491480878,
0.9262323105816003,
0.4093931513055624,
0.43076708407562836,
0.4716787019718713,
0.9210444151123908,
0.9192931052318097,
0.8852895467415316,
0.9996212673405107,
0.9995580744766234,
0.9998754347430274,
0.8212397030122961,
0.6064790413252394,
0.43371093964453905,
0.9993965372044223,
0.9993502673669963,
0.9996858808881637,
0.6953793592000725,
0.943467117340319,
0.7629785368391938,
0.6899668286083106,
0.3709530510821317,
0.6977802237425788,
0.5852851939671919,
0.5667239640555052,
0.7024605642965004,
0.6328932113058444,
0.2620202256862308,
0.46794525821763333,
0.5804015074668117,
0.5576213838934553,
0.6718108369258075,
0.9999014461063227,
0.9998512016852299,
0.9999278365045776,
0.5047499495477554,
0.47108452316958815,
0.4633423430404021,
0.7239226819628426,
0.4467181118512972,
0.296551706183831,
0.9791370321531192,
0.9821609295129347,
0.9726338582513931,
0.9997575679502393,
0.9996145145913061,
0.9997389913857065,
0.9958211495581951,
0.9935504950821797,
0.9906796363337099,
0.7037398432183858,
0.6558558740251067,
0.6498785953525238,
0.9997154239901205,
0.999467393978126,
0.9960077337362665,
0.8342510831682098,
0.8653052413260633,
0.7508285075698964,
0.9987488070502946,
0.5965796011803082,
0.9995371225935034,
0.9996265183972564,
0.9990452558115929,
0.999571656077788,
0.2529283547865787,
0.7764947300760385,
0.7786737520666781,
0.7263781076934765,
0.6116173550649253,
0.5988593509083462,
0.9993261804974033,
0.9995450146511298,
0.9998005779685567,
0.49202650329016895,
0.4264156884426978,
0.19280753962840855,
0.941138231211119,
0.9996739955151408,
0.9996845997185915,
0.9841458662073007,
0.8832860661149717,
0.9292267304477149,
0.4304028206724121,
0.5180026682686568,
0.6237772662585035,
0.5138168048956061,
0.7342330725058087,
0.22603121720150543,
0.1800781367842741,
0.18033585481153347,
0.9972962291099047,
0.3802625392513917,
0.6659906127995988,
0.7702758555663081,
0.7924911080088364,
0.8125443013992801,
0.8299279196545282,
0.984410207864405,
0.9993975449281757,
0.9809434690614522
]
}