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

{
    "model_dir": "models/openllama-3b",
    "distribution_id": "wrong_arc",
    "timestamp": "2023-10-20 01:10:23",
    "eval_loss": 0.713603138923645,
    "eval_score": -0.2604227066040039,
    "eval_brier_score": 0.2604227066040039,
    "eval_average_probability": 0.48988017439842224,
    "eval_accuracy": 0.244,
    "eval_probabilities": [
        0.4871862530708313,
        0.4881269931793213,
        0.4684944450855255,
        0.481976181268692,
        0.4841439127922058,
        0.4874495565891266,
        0.48795777559280396,
        0.4829939603805542,
        0.4818950593471527,
        0.503623366355896,
        0.5053632259368896,
        0.5084194540977478,
        0.4929749667644501,
        0.4957391321659088,
        0.4832057058811188,
        0.49733787775039673,
        0.48589348793029785,
        0.5113143920898438,
        0.49171358346939087,
        0.5021988749504089,
        0.5186639428138733,
        0.47192147374153137,
        0.48859986662864685,
        0.4775811731815338,
        0.4668554365634918,
        0.49508213996887207,
        0.4963756501674652,
        0.49291929602622986,
        0.5045862197875977,
        0.49326765537261963,
        0.4959131181240082,
        0.496704638004303,
        0.49327513575553894,
        0.5027367472648621,
        0.5166918635368347,
        0.4582791030406952,
        0.4814019501209259,
        0.485125869512558,
        0.4903547763824463,
        0.49874451756477356,
        0.48993581533432007,
        0.46368175745010376,
        0.4886102080345154,
        0.487223356962204,
        0.48593634366989136,
        0.5003162026405334,
        0.48458331823349,
        0.5028654336929321,
        0.5444562435150146,
        0.47074443101882935,
        0.5052387714385986,
        0.492236852645874,
        0.4959132969379425,
        0.4766242802143097,
        0.4995802044868469,
        0.46299678087234497,
        0.4984775185585022,
        0.48708343505859375,
        0.5129186511039734,
        0.4951343536376953,
        0.4691915810108185,
        0.4773304760456085,
        0.4941478371620178,
        0.5010058283805847,
        0.49587762355804443,
        0.4720380902290344,
        0.4916825592517853,
        0.500735342502594,
        0.5242191553115845,
        0.47962480783462524,
        0.5006029605865479,
        0.5004259347915649,
        0.4757077097892761,
        0.4898621737957001,
        0.4919328987598419,
        0.5181395411491394,
        0.47911936044692993,
        0.4932674169540405,
        0.5001049637794495,
        0.5009469389915466,
        0.4952067732810974,
        0.4845515191555023,
        0.495017409324646,
        0.4930589199066162,
        0.4773653745651245,
        0.4855053722858429,
        0.48002007603645325,
        0.5084033608436584,
        0.4980103373527527,
        0.49130555987358093,
        0.49260714650154114,
        0.500511109828949,
        0.481354296207428,
        0.49229317903518677,
        0.4600612223148346,
        0.5211605429649353,
        0.502425491809845,
        0.49343574047088623,
        0.5017443895339966,
        0.4758780598640442,
        0.4974922239780426,
        0.4609968960285187,
        0.47961291670799255,
        0.4825732409954071,
        0.48155030608177185,
        0.4861832559108734,
        0.5013216733932495,
        0.48793330788612366,
        0.4620995819568634,
        0.4835435748100281,
        0.4626553952693939,
        0.4947565793991089,
        0.48447293043136597,
        0.48785200715065,
        0.4923180043697357,
        0.4661085307598114,
        0.4809713065624237,
        0.48937109112739563,
        0.47135382890701294,
        0.4915151596069336,
        0.45796939730644226,
        0.500161349773407,
        0.5087539553642273,
        0.5148164629936218,
        0.4916517436504364,
        0.4890440106391907,
        0.506800651550293,
        0.47400644421577454,
        0.4858291745185852,
        0.4662706255912781,
        0.46891942620277405,
        0.46907466650009155,
        0.47487545013427734,
        0.4899934232234955,
        0.5001229047775269,
        0.49309131503105164,
        0.4823770225048065,
        0.48497167229652405,
        0.512969434261322,
        0.4818212389945984,
        0.5007370710372925,
        0.48888352513313293,
        0.4549937844276428,
        0.4858066141605377,
        0.48472926020622253,
        0.47980621457099915,
        0.4806237816810608,
        0.47106361389160156,
        0.49751996994018555,
        0.492817223072052,
        0.5050973296165466,
        0.48550713062286377,
        0.507538378238678,
        0.5025631189346313,
        0.5051950812339783,
        0.5104390978813171,
        0.49535825848579407,
        0.4942142963409424,
        0.5114203095436096,
        0.47967883944511414,
        0.5048908591270447,
        0.5041651129722595,
        0.45791736245155334,
        0.49299851059913635,
        0.4908706545829773,
        0.4674992859363556,
        0.5335745215415955,
        0.48192065954208374,
        0.5052374601364136,
        0.47684434056282043,
        0.4770922064781189,
        0.5037971138954163,
        0.4848896861076355,
        0.46035799384117126,
        0.48817557096481323,
        0.48727086186408997,
        0.4430428445339203,
        0.47781458497047424,
        0.4951154291629791,
        0.5058280229568481,
        0.49037325382232666,
        0.49169719219207764,
        0.495022177696228,
        0.5002468228340149,
        0.497315913438797,
        0.48562681674957275,
        0.4750918447971344,
        0.4702063500881195,
        0.5054723024368286,
        0.4980982840061188,
        0.47961142659187317,
        0.4986233711242676,
        0.5094085335731506,
        0.47487619519233704,
        0.5006697773933411,
        0.49024826288223267,
        0.4822885990142822,
        0.4757611155509949,
        0.4957931935787201,
        0.5048408508300781,
        0.47144395112991333,
        0.49869948625564575,
        0.4906611442565918,
        0.5001755952835083,
        0.4965263307094574,
        0.49637553095817566,
        0.49219250679016113,
        0.46787625551223755,
        0.487187922000885,
        0.5140500664710999,
        0.48086872696876526,
        0.4967770576477051,
        0.49363696575164795,
        0.4815727174282074,
        0.48441192507743835,
        0.4913637936115265,
        0.4857942759990692,
        0.4956826865673065,
        0.5034660696983337,
        0.4843102693557739,
        0.4859721064567566,
        0.48880448937416077,
        0.4972941279411316,
        0.5038987994194031,
        0.48899024724960327,
        0.48087137937545776,
        0.4987112283706665,
        0.4546473026275635,
        0.4745439887046814,
        0.47993218898773193,
        0.48593851923942566,
        0.4863012433052063,
        0.4777888059616089,
        0.49433058500289917,
        0.5088313221931458,
        0.489531934261322,
        0.48840633034706116,
        0.48101288080215454,
        0.5040384531021118,
        0.5025137066841125,
        0.5005448460578918,
        0.4665708541870117,
        0.4935063123703003,
        0.4933200478553772,
        0.4889622926712036,
        0.48492246866226196,
        0.4921472370624542,
        0.4960475265979767,
        0.5041992664337158,
        0.48809078335762024
    ],
    "eval_runtime": 14.784,
    "eval_samples_per_second": 16.91,
    "eval_steps_per_second": 0.541
}