sample_response.json

{
    "id": 1234,
    "results": {
        "BPS": {
            "acc": 0.49
        },
        "CheGeKa": {
            "em": 0.0,
            "f1": 0.001201923076923077
        },
        "LCS": {
            "acc": 0.078
        },
        "MathLogicQA": {
            "acc": 0.2379702537182852
        },
        "MultiQ": {
            "em": 0.005555555555555556,
            "f1": 0.021870219014724045
        },
        "PARus": {
            "acc": 0.492
        },
        "RCB": {
            "acc": 0.3378995433789954,
            "f1_macro": 0.3383335745645775
        },
        "RWSD": {
            "acc": 0.4846153846153846
        },
        "SimpleAr": {
            "acc": 0.0
        },
        "USE": {
            "grade_norm": 0.060784313725490195
        },
        "ruDetox": {
            "fl": 0.5578551230338372,
            "j": 0.3816129346637701,
            "sim": 0.8054497441265253,
            "sta": 0.8405460153513176
        },
        "ruEthics": {
            "correct.justice": 0.05020856170875817,
            "correct.law": 0.07830396004971925,
            "correct.moral": -0.08346819418405328,
            "correct.utilitarianism": -0.024393037349895674,
            "correct.virtue": 0.037727498455149874,
            "ethical.justice": -0.07713103749826518,
            "ethical.law": -0.001934545879364499,
            "ethical.moral": -9.105511969567648e-05,
            "ethical.utilitarianism": 0.0037156535319434305,
            "ethical.virtue": 0.0738710178906162,
            "good.justice": -0.02232405272967756,
            "good.law": 0.01951635912572728,
            "good.moral": 0.006374961981576146,
            "good.utilitarianism": 0.06492166530158856,
            "good.virtue": -0.06111365371617636
        },
        "ruHHH": {
            "acc": 0.5337078651685393,
            "acc.harmless": 0.5172413793103449,
            "acc.helpful": 0.4915254237288136,
            "acc.honest": 0.5901639344262295
        },
        "ruHateSpeech": {
            "acc": 0.4981132075471698,
            "acc.другое": 0.4426229508196721,
            "acc.женщины": 0.5740740740740741,
            "acc.лгбт": 0.29411764705882354,
            "acc.мигранты": 0.42857142857142855,
            "acc.мужчины": 0.5142857142857142,
            "acc.национальность": 0.4594594594594595
        },
        "ruHumanEval": {
            "pass@1": 0.0,
            "pass@10": 0.0,
            "pass@5": 0.0
        },
        "ruMMLU": {
            "acc": 0.24141519250780438,
            "acc.abstract_algebra": 0.5,
            "acc.anatomy": 0.4,
            "acc.astronomy": 0.4,
            "acc.business_ethics": 0.2,
            "acc.clinical_knowledge": 0.2727272727272727,
            "acc.college_biology": 0.1111111111111111,
            "acc.college_chemistry": 0.2727272727272727,
            "acc.college_computer_science": 0.0,
            "acc.college_mathematics": 0.4,
            "acc.college_medicine": 0.2549019607843137,
            "acc.college_physics": 0.3,
            "acc.computer_security": 0.2,
            "acc.conceptual_physics": 0.3,
            "acc.econometrics": 0.2727272727272727,
            "acc.electrical_engineering": 0.3,
            "acc.elementary_mathematics": 0.2,
            "acc.formal_logic": 0.5,
            "acc.global_facts": 0.2,
            "acc.high_school_biology": 0.2857142857142857,
            "acc.high_school_chemistry": 0.3,
            "acc.high_school_computer_science": 0.3333333333333333,
            "acc.high_school_european_history": 0.30303030303030304,
            "acc.high_school_geography": 0.24050632911392406,
            "acc.high_school_government_and_politics": 0.2962962962962963,
            "acc.high_school_macroeconomics": 0.20588235294117646,
            "acc.high_school_mathematics": 0.0,
            "acc.high_school_microeconomics": 0.2,
            "acc.high_school_physics": 0.0,
            "acc.high_school_psychology": 0.0625,
            "acc.high_school_statistics": 0.4,
            "acc.high_school_us_history": 0.2,
            "acc.high_school_world_history": 0.4375,
            "acc.human_aging": 0.3,
            "acc.human_sexuality": 0.3,
            "acc.international_law": 0.3333333333333333,
            "acc.jurisprudence": 0.15384615384615385,
            "acc.logical_fallacies": 0.3,
            "acc.machine_learning": 0.2,
            "acc.management": 0.2,
            "acc.marketing": 0.2857142857142857,
            "acc.medical_genetics": 0.5454545454545454,
            "acc.miscellaneous": 0.18181818181818182,
            "acc.moral_disputes": 0.2,
            "acc.moral_scenarios": 0.2,
            "acc.nutrition": 0.2857142857142857,
            "acc.philosophy": 0.17647058823529413,
            "acc.prehistory": 0.3,
            "acc.professional_accounting": 0.2,
            "acc.professional_law": 0.25,
            "acc.professional_medicine": 0.1,
            "acc.professional_psychology": 0.2,
            "acc.public_relations": 0.35714285714285715,
            "acc.security_studies": 0.0,
            "acc.sociology": 0.2,
            "acc.us_foreign_policy": 0.2,
            "acc.virology": 0.0625,
            "acc.world_religions": 0.21153846153846154
        },
        "ruModAr": {
            "acc": 0.0
        },
        "ruMultiAr": {
            "acc": 0.0
        },
        "ruOpenBookQA": {
            "acc": 0.2325,
            "f1_macro": 0.23250468078345715
        },
        "ruTiE": {
            "acc": 0.49767441860465117
        },
        "ruWorldTree": {
            "acc": 0.21714285714285714,
            "f1_macro": 0.217112546769622
        },
        "total_score": 0.17058262521762277
    },
    "status": "OK"
}