574m174b174b/evaluation/rankeval/lm1-574m-174b-results_lm-eval_global_step331103_2023-01-24-13-57-05_3shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.327,
            "acc_stderr": 0.014842213153411242
        },
        "anli_r2": {
            "acc": 0.353,
            "acc_stderr": 0.01512017260548369
        },
        "anli_r3": {
            "acc": 0.3408333333333333,
            "acc_stderr": 0.013688600793296939
        },
        "cb": {
            "acc": 0.5,
            "acc_stderr": 0.06741998624632421,
            "f1": 0.42649620505163616
        },
        "copa": {
            "acc": 0.68,
            "acc_stderr": 0.046882617226215034
        },
        "hellaswag": {
            "acc": 0.3430591515634336,
            "acc_stderr": 0.004737608340163389,
            "acc_norm": 0.4219279028082055,
            "acc_norm_stderr": 0.004928578106026371
        },
        "rte": {
            "acc": 0.5342960288808665,
            "acc_stderr": 0.030025579819366426
        },
        "winogrande": {
            "acc": 0.5098658247829518,
            "acc_stderr": 0.014049749833367592
        },
        "storycloze_2016": {
            "acc": 0.6349545697487974,
            "acc_stderr": 0.011133301783914869
        },
        "boolq": {
            "acc": 0.481039755351682,
            "acc_stderr": 0.008738765179491934
        },
        "arc_easy": {
            "acc": 0.5563973063973064,
            "acc_stderr": 0.010194308914521135,
            "acc_norm": 0.5467171717171717,
            "acc_norm_stderr": 0.010214901516731604
        },
        "arc_challenge": {
            "acc": 0.23464163822525597,
            "acc_stderr": 0.012383873560768676,
            "acc_norm": 0.27474402730375425,
            "acc_norm_stderr": 0.013044617212771227
        },
        "sciq": {
            "acc": 0.889,
            "acc_stderr": 0.009938701010583726,
            "acc_norm": 0.883,
            "acc_norm_stderr": 0.010169287802713329
        },
        "piqa": {
            "acc": 0.6789989118607181,
            "acc_stderr": 0.010892641574707899,
            "acc_norm": 0.6800870511425462,
            "acc_norm_stderr": 0.010882873582092062
        }
    },
    "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
    }
}