421m32b32b/421m32b32bpile/evaluation/rankeval/lm1-421m-32b-results_lm-eval_global_step60336_2023-01-23-18-55-40_1shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.335,
            "acc_stderr": 0.014933117490932575
        },
        "anli_r2": {
            "acc": 0.353,
            "acc_stderr": 0.015120172605483708
        },
        "anli_r3": {
            "acc": 0.3275,
            "acc_stderr": 0.013553211167251942
        },
        "cb": {
            "acc": 0.48214285714285715,
            "acc_stderr": 0.06737697508644648,
            "f1": 0.3421052631578947
        },
        "copa": {
            "acc": 0.67,
            "acc_stderr": 0.04725815626252607
        },
        "hellaswag": {
            "acc": 0.2976498705437164,
            "acc_stderr": 0.004562902604938726,
            "acc_norm": 0.3331009759012149,
            "acc_norm_stderr": 0.0047035905585525025
        },
        "rte": {
            "acc": 0.5306859205776173,
            "acc_stderr": 0.03003973059219781
        },
        "winogrande": {
            "acc": 0.5280189423835833,
            "acc_stderr": 0.014030404213405786
        },
        "storycloze_2016": {
            "acc": 0.5911277391769107,
            "acc_stderr": 0.011368775493925615
        },
        "boolq": {
            "acc": 0.5532110091743119,
            "acc_stderr": 0.008695392261996197
        },
        "arc_easy": {
            "acc": 0.4831649831649832,
            "acc_stderr": 0.010253966261288888,
            "acc_norm": 0.4478114478114478,
            "acc_norm_stderr": 0.010203742451111506
        },
        "arc_challenge": {
            "acc": 0.19197952218430034,
            "acc_stderr": 0.011509598906598095,
            "acc_norm": 0.23378839590443687,
            "acc_norm_stderr": 0.012368225378507135
        },
        "sciq": {
            "acc": 0.824,
            "acc_stderr": 0.012048616898597516,
            "acc_norm": 0.78,
            "acc_norm_stderr": 0.01310617304066176
        },
        "piqa": {
            "acc": 0.6392818280739935,
            "acc_stderr": 0.011204064809088515,
            "acc_norm": 0.6458106637649619,
            "acc_norm_stderr": 0.011158755672626109
        }
    },
    "versions": {
        "anli_r1": 0,
        "anli_r2": 0,
        "anli_r3": 0,
        "cb": 1,
        "copa": 0,
        "hellaswag": 0,
        "rte": 0,
        "winogrande": 0,
        "storycloze_2016": 0,
        "boolq": 1,
        "arc_easy": 0,
        "arc_challenge": 0,
        "sciq": 0,
        "piqa": 0
    }
}