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

762 lines
21 KiB
JSON

{
"model_dir": "models/llama-30b",
"distribution_id": "code",
"timestamp": "2023-10-16 04:00:52",
"eval_score": -0.1590690923246832,
"eval_brier_score": 0.1590690923246832,
"eval_average_probability": 0.7297176488269179,
"eval_accuracy": 0.772,
"eval_probabilities": [
0.865585276172795,
0.9773957706765767,
0.9564308502622123,
0.9944157776426631,
0.9998663134512827,
0.9449621807530519,
0.9871225470817109,
0.9707339809310083,
0.8155294323812242,
0.9665373201707815,
0.8930967883552495,
0.9452510912056714,
0.6008240062862134,
0.9920379958677332,
0.01001663541847126,
0.5525515222357632,
0.5335329328784559,
0.3466101482113383,
0.9991215294815743,
0.9991853532579102,
0.9989786237094282,
0.8865456338670851,
0.9797187615648932,
0.9906321968025442,
0.7195378501673275,
0.8752436214135013,
0.8750841647000852,
0.9929174379480411,
0.999897501184391,
0.9987561174963461,
0.5898165560776766,
0.9996594333722278,
0.5870326803974834,
0.9565316665949202,
0.8606963251834787,
0.6006477742528479,
0.6875550260852005,
0.9761683144910753,
0.992073800876202,
0.9996736915258758,
0.989594812114068,
0.9954234487431214,
0.9400514123865497,
0.9529942062358079,
0.15055937136252054,
0.9851975716631765,
0.9839282838290992,
0.9985722585465531,
0.5037356076237601,
0.2936018723630829,
0.9449176418611015,
0.9999526978272653,
0.9999218409970736,
0.9996169412212422,
0.902751095682114,
0.6283046716307026,
0.9441222556823933,
0.9439026046537502,
0.1937933672568386,
0.8906544767138846,
0.7479310344592417,
0.41174477751291977,
0.37864904238283154,
0.9975084515509626,
0.027247168550746936,
0.028396796235330757,
0.9949415700907662,
0.7437473027605672,
0.7832839584850236,
0.9189311768953257,
0.8712724113161582,
0.410748629659082,
0.7233887665427714,
0.07024085856684703,
0.6022663682860887,
0.9999987184349068,
0.9999991139743007,
0.9999993381319867,
0.9748634576665798,
0.3369658526994852,
0.7866674708377066,
0.9778857985304562,
0.6179808881005591,
0.993048126119816,
0.8640894404252473,
0.9699355449107827,
0.999557886928252,
0.9999098176497544,
0.9998993826531438,
0.9979605968933957,
0.7282680655326821,
0.6222543773382858,
0.709996881055346,
0.47160670962647044,
0.7969192624380242,
0.6883914635182579,
0.3659653635492788,
0.08822200619076437,
0.22230098256786707,
0.8210043142582235,
0.5980731314715311,
0.10222399083872546,
0.9997857030619831,
0.9963244696718517,
0.9953491239508606,
0.9182472057392453,
0.8624076924734928,
0.3177403516231882,
0.4687212537132759,
0.4016288200224828,
0.5317175170107613,
0.9922720654192927,
0.9988501134647246,
0.9627971961284115,
0.8485502290500904,
0.5370043734209529,
0.7435707941985259,
0.9564453233344478,
0.9375782296853787,
0.6964985153261941,
0.8625019462101431,
0.9063429530862328,
0.9820228370048942,
0.9989476486488948,
0.9383716390044241,
0.9676481764638117,
0.36584773970671514,
0.5215658141642762,
0.4386152883425809,
0.451808504558308,
0.6976263604371398,
0.42547903286018574,
0.034265725042499424,
0.15486614191171755,
0.43345160240634584,
0.8074543748634991,
0.862369019686283,
0.30639242318858784,
0.8777355931652733,
0.6356242625423711,
0.12043738232961243,
0.9961711417768405,
0.2326717325516552,
0.9995912948942123,
0.9003115886921828,
0.13365105286036927,
0.2232920482908647,
0.9880308923110697,
0.8673065033250118,
0.793032955966061,
0.8487317262502871,
0.8881131492325699,
0.9379134249017141,
0.5517417271523752,
0.1640579851755035,
0.5494111260875908,
0.9353279416470461,
0.8945124208313621,
0.8991909360063897,
0.9999874668578922,
0.8043277056310597,
0.9981933756608066,
0.9806957706584771,
0.8622529659221203,
0.9585668228763412,
0.982451313442546,
0.9989913832286026,
0.977115257146372,
0.9402616602132192,
0.6377437725256999,
0.1750946594535844,
0.7108842960756674,
0.8359456624342285,
0.8128139646392883,
0.9670472004382846,
0.7848936903490653,
0.9537998485528254,
0.9949097883684533,
0.9999903595318043,
0.9999943639945854,
0.6393237691416244,
0.7168820690696766,
0.7300283865818121,
0.9125780651302793,
0.977793945660036,
0.5879604613000611,
0.4209290557858371,
0.6402248241779174,
0.5676163902166736,
0.9428283444910541,
0.8882351668955827,
0.9999643966441009,
0.959455345653817,
0.961782010763731,
0.9731105430892732,
0.9993540384603877,
0.9893523184663146,
0.9913699061021031,
0.4634899787397687,
0.37969894166985035,
0.9418903787518325,
0.970660691533003,
0.7911396233101344,
0.7937713816524241,
0.48767585496857957,
0.951244264218897,
0.9839740120293181,
0.5062188995966224,
0.7334533022267845,
0.22777563003918988,
0.8956532567869703,
0.7914679511097364,
0.9724709850196648,
0.18508130233922798,
0.9068522404138861,
0.07834712219030027,
0.9764101647688764,
0.854189163424768,
0.9736438719687706,
0.9626744031219994,
0.9891077112719766,
0.9732192015161719,
0.7895072067207702,
0.934927189033235,
0.36918045943479405,
0.8706128350887657,
0.8908233841113775,
0.9338518495060385,
0.16741434913332412,
0.9758957145543693,
0.9662337329725864,
0.7679633683475184,
0.8078489724958191,
0.718854326107169,
0.40011884549720567,
0.9791471882942799,
0.13642382777394244,
0.8884460607479006,
0.7777251594080585,
0.8579515683098636,
0.25186342014632723,
0.24055001763101566,
0.4084275127046153,
0.8774057212763716,
0.9423107519541691,
0.9973988453637206,
0.6453904560210009,
0.9250678245538883,
0.9999524979258956,
0.9981211709247517,
0.9884085543341244,
0.999491763847201,
0.14542079185793833,
0.7002599385235071,
0.5477249572571127,
0.9749107470491147,
0.9867139481880285,
0.9900141755043931,
0.8772366406902984,
0.9872228277258865,
0.43109484178808444,
0.9466825520564611,
0.9625563734640421,
0.9184097807943289,
0.9886376519480714,
0.7971822562338906,
0.9638956935875647,
0.08206305884601367,
0.1708432504334797,
0.22214210587601702,
0.828973939837307,
0.9985619190415318,
0.7037488829983157,
0.8712656992460667,
0.7503571140086611,
0.7913054818728879,
0.9941098328656675,
0.9888531146787803,
0.9033952604154943,
0.7973260106821236,
0.9943397393199421,
0.5966249475341596,
0.9903778149122514,
0.7463501159260733,
0.9375124333773298,
0.9999014720979136,
0.9999359420721903,
0.9998745498499777,
0.8906774687531451,
0.6840001665140496,
0.5871737073372172,
0.8842619484722096,
0.27713510855643086,
0.15049517797243497,
0.9919515978492212,
0.3090998095519538,
0.9796978096617251,
0.8311028373146951,
0.7384717470582225,
0.16290359516805009,
0.05847828810207691,
0.1340318434828104,
0.487502663242008,
0.7915066425799911,
0.9509724731902348,
0.9086683374471463,
0.9544062151584372,
0.9771165209446916,
0.9940084452950383,
0.3598387392529634,
0.8248128741884145,
0.8668787283660528,
0.9934755628017224,
0.827903191432848,
0.15260010703319346,
0.5490164557062183,
0.9146328455492555,
0.9669679661537451,
0.7287165424998371,
0.5752250108286378,
0.6345787028698879,
0.7816669590458887,
0.9025764127768487,
0.6849666031724251,
0.9527439662968084,
0.9914173079898124,
0.05239859908698572,
0.1483453064453717,
0.7882025263975911,
0.45541766957158636,
0.974338659163739,
0.9404116271796574,
0.943396195522646,
0.8504527533463012,
0.8818469962245828,
0.6752861989730866,
0.3463075393387244,
0.8823978858208489,
0.016687533891200243,
0.03231676543061222,
0.015271586356354406,
0.0838645857409643,
0.8784164998446018,
0.8794370112787808,
0.6377015266374626,
0.9807276550097423,
0.6026657062843954,
0.3693186900896438,
0.23185420960903058,
0.8271548934691508,
0.7524679448894303,
0.6551826677433223,
0.352831095162914,
0.4039328659559397,
0.9999550658055745,
0.850830801754491,
0.9996847663088594,
0.9964168641831362,
0.9998345038151849,
0.9709803885834178,
0.8443245211624093,
0.3254093041405294,
0.4606141838586737,
0.9277837235813383,
0.9426828345062812,
0.975318158693891,
0.9929155555761331,
0.996825511076603,
0.9991049879327967,
0.614530519590218,
0.239562618294339,
0.2489494397211315,
0.7580090989182666,
0.7632063275638039,
0.8924781489911914,
0.5749910492955237,
0.6427530146897918,
0.8691385976873388,
0.9730125442932755,
0.9909423158148329,
0.9889956053329232,
0.003034298505613604,
0.876162434340709,
0.573708468324599,
0.8565179282502171,
0.7621496171546426,
0.7499470991693254,
0.6002318243152288,
0.5365521845418829,
0.3214560211631441,
0.87206197882437,
0.8448538554866238,
0.8471951364810077,
0.9951156749797538,
0.9950207366608949,
0.7062015713848007,
0.8722325110309445,
0.22976501902839833,
0.6560418105105561,
0.9839775234183897,
0.9878679971007525,
0.9783243043114866,
0.3183868459116671,
0.08590956389680786,
0.05607987876219994,
0.9860873140364955,
0.4264614793436437,
0.5504520675091329,
0.9396840721694004,
0.09750254432007656,
0.8783792059019668,
0.179274101859526,
0.028760264372633015,
0.6495217427321213,
0.38260835796169346,
0.5403199080951819,
0.43907214148101437,
0.3610028931799174,
0.1505049612838144,
0.7826958892140992,
0.5039807875376151,
0.6812075284491598,
0.4480876450651862,
0.13691358792565964,
0.00816836551895217,
0.2107852211813091,
0.11614646760192864,
0.19781861788699484,
0.3142551437909209,
0.35988046681253644,
0.15188651539988798,
0.22853264745781354,
0.18452008934236136,
0.5274180533823878,
0.48884301501184124,
0.9979488097001925,
0.996926075935554,
0.9988398822181597,
0.9876052340280245,
0.6403818096563569,
0.9974181123297948,
0.9998513628206827,
0.9876713154091318,
0.9623471358569735,
0.9132702422819695,
0.0770907827949886,
0.8978787935148479,
0.0797882425626043,
0.9970990935702104,
0.012680783999087453,
0.6664210187173797,
0.9136422752403304,
0.9895529293686212,
0.8997187649645386,
0.49817932804354953,
0.9892261317863318,
0.013491138929861079,
0.7167241433207664,
0.9129703570254761,
0.8115211585355745,
0.915848746728656,
0.9444023964146127,
0.9929229302306637,
0.9799117454704712,
0.9808885227189111,
0.8994825932794169,
0.8954483206197079,
0.8931523246284045,
0.5091247874992019,
0.9930346636849791,
0.9975437306763737,
0.9787315444518552,
0.9608849643933669,
0.9999914537323892,
0.6777326854905521,
0.35046160699186024,
0.15476130175787473,
0.9942294919190947,
0.9619321037247928,
0.944251902975716,
0.5926412429625951,
0.2825180859960903,
0.9949443973789049,
0.9740983329877393,
0.5115623991279981,
0.07034231986808233,
0.9047180226986871,
0.6485391242643557,
0.9274938312967095,
0.9992702454952765,
0.8460651650308707,
0.9981749661655861,
0.9925309929526771,
0.8631300148712875,
0.9555249928969124,
0.7624586580462349,
0.7486367929641555,
0.9970002074120402,
0.47094315521448427,
0.9998375509086832,
0.43059140108048943,
0.7285857407082239,
0.13093232554187573,
0.9271760416756233,
0.8271449035818675,
0.8289678198048055,
0.8933051612790113,
0.5233400951901335,
0.4795003598021148,
0.4036535410353508,
0.9064893251874879,
0.9537045275330649,
0.9901632696428747,
0.2349026625171654,
0.26744428267126974,
0.9186021501280975,
0.9999824072033197,
0.9999234280864949,
0.9999285487208587,
0.11474431494163526,
0.07286684638324518,
0.13397937171602556,
0.7195190527753373,
0.6613727331733251,
0.778634775898496,
0.5064823236811165,
0.764899500250655,
0.30826157726019576,
0.9197619401107104,
0.9899871217466923,
0.9785788261794484,
0.8556485221162445,
0.9122757547168603,
0.7896248586948335,
0.8578887538320116,
0.6525438003414626,
0.9715298218506104,
0.06523359254013757,
0.18183465218137296,
0.7146089891667544,
0.944697115966319,
0.5583383249332168,
0.7601307818301078,
0.999880913457297,
0.9999510735397029,
0.9999937266374118,
0.9773696446396397,
0.48060445446027994,
0.9699344283846347,
0.9998142551828739,
0.15153780806379424,
0.8814477044153256,
0.9486508991829171,
0.6575313221362555,
0.1739143455563802,
0.7634609903375522,
0.5945252580429887,
0.5832491560385982,
0.2893730501592101,
0.8131157720311082,
0.517361108295863,
0.9670034162497246,
0.9846132940670793,
0.9164891518639514,
0.8846945656030815,
0.9338780331767492,
0.9352046570540643,
0.460715784783574,
0.7389556098586268,
0.6429572026126849,
0.9999831559862272,
0.9538628020070853,
0.9222706240998005,
0.4501642157960818,
0.9515880376033113,
0.07606412506655524,
0.999809088090837,
0.9999369445284082,
0.9998246368281436,
0.9358420755071477,
0.871799313826709,
0.9588127802022569,
0.5361363296418387,
0.11850522821635413,
0.16402835770250102,
0.3881198378904549,
0.6284090963903409,
0.5424945986976552,
0.4900584716908488,
0.9652445369988197,
0.8197486456066654,
0.9063550876727191,
0.48411424332109076,
0.830385963963741,
0.9984659674336241,
0.9998686739295881,
0.7419779998003979,
0.9618412386537764,
0.9978077205298157,
0.9499980131391914,
0.3776733387687345,
0.9484912251404509,
0.9485005613399725,
0.4865005508936662,
0.3550003577373352,
0.10332827738498518,
0.6379227178601956,
0.43071735976565917,
0.7809827539341005,
0.9834698317097491,
0.9904303928107854,
0.9553606705452571,
0.9639196360865422,
0.6431796510882659,
0.563051173492132,
0.563428212297628,
0.4925703150173368,
0.781721698981263,
0.21682497424594016,
0.6317543666219088,
0.7274165258177561,
0.6973583893958909,
0.3764043760755942,
0.29914200973065924,
0.971684410157777,
0.9482505439161781,
0.9343472272170404,
0.8400142958199706,
0.964528611936855,
0.8102887317480919,
0.9957535223124169,
0.9836341429020262,
0.8613835126802424,
0.8268354780949483,
0.999508659210971,
0.999497318318108,
0.9974110478127276,
0.05342477873489919,
0.5778719132124881,
0.3593150621942619,
0.4942124143099734,
0.5709204596028106,
0.11797350424009628,
0.9111030803045381,
0.3021688937424625,
0.9996569614812035,
0.9939043964294836,
0.4258378944975093,
0.992961316075492,
0.9558134559195168,
0.773853556751552,
0.9992232320713536,
0.9989129282389231,
0.9974376785566171,
0.9202659317848715,
0.29121970678535025,
0.9580414297945362,
0.896924039072563,
0.957576570617116,
0.9257533015828191,
0.9460317519564901,
0.9304566911428708,
0.9353689555658071,
0.8487470919164426,
0.37071599374364933,
0.9999072317772263,
0.9846871850916509,
0.9849811938995763,
0.9965105297598638,
0.8001872698836925,
0.14518503121911042,
0.2998346380966497,
0.7215268512178674,
0.36831795430524894,
0.7008802835741286,
0.7221416622260297,
0.8609080512412113,
0.3063754663410857,
0.3764781249038232,
0.9927666404290463,
0.9763695812576093,
0.9911248850734968,
0.9968180537538992,
0.9997991496735812,
0.9997387300391135,
0.9999674857731592,
0.7032981493830637,
0.9750434055532018,
0.15800697393863705,
0.83453145239678,
0.3913420857357588,
0.8074026887826818,
0.6365050879991034,
0.983649175401199,
0.9929012282369544,
0.9989883872434914,
0.8227920860736155,
0.1748335819815831,
0.8965928438363222,
0.9539271374375929,
0.972903024040289,
0.9523769324124723,
0.5413015825076228,
0.7475942933674604,
0.7358816528506914,
0.9727958008561244,
0.0983561750451788,
0.2534542011757853,
0.9663261924323444,
0.8897400107519734,
0.9438388430523927,
0.88617006476713,
0.5457727376998693,
0.9562425307352039,
0.873698796110355,
0.7489229130916484,
0.6779468728426679,
0.9421360938215322,
0.9436531598481688,
0.922033266325504,
0.7836979242548208,
0.532395872674072,
0.15919510109788518,
0.6058025350471926,
0.9918440154049065,
0.5292970626801169,
0.9133939319933456,
0.9666996312213855,
0.4342228942355352,
0.6918780781664017,
0.9270240323286508,
0.31922871717285195,
0.5944466992324624,
0.998212970394455,
0.9325924911938768,
0.32014013518813156,
0.05192283110166884,
0.38095028500152195,
0.8544330988589637,
0.7572667494353544,
0.9526827068743805,
0.9854570260680915,
0.9865832601942993,
0.9840651629907303,
0.9844790171420317,
0.9739425676304584,
0.9141892047246012,
0.9914330651557887,
0.8529566771040252,
0.9785065752362552
],
"eval_rms_calibration_error": 0.3057146181815999
}