574m174b174b/evaluation/rankeval/lm1-574m-174b-results_lm-eval_global_step331103_2023-01-22-13-12-25_0shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.326,
            "acc_stderr": 0.014830507204541033
        },
        "anli_r2": {
            "acc": 0.337,
            "acc_stderr": 0.014955087918653609
        },
        "anli_r3": {
            "acc": 0.345,
            "acc_stderr": 0.01372842153945488
        },
        "cb": {
            "acc": 0.4107142857142857,
            "acc_stderr": 0.06633634150359541,
            "f1": 0.2824582824582824
        },
        "copa": {
            "acc": 0.7,
            "acc_stderr": 0.046056618647183814
        },
        "hellaswag": {
            "acc": 0.3451503684524995,
            "acc_stderr": 0.004744456628455116,
            "acc_norm": 0.41983668591913964,
            "acc_norm_stderr": 0.004925233680511587
        },
        "rte": {
            "acc": 0.5306859205776173,
            "acc_stderr": 0.03003973059219781
        },
        "winogrande": {
            "acc": 0.526440410418311,
            "acc_stderr": 0.014032823874407224
        },
        "storycloze_2016": {
            "acc": 0.638161411010155,
            "acc_stderr": 0.011112247531047543
        },
        "boolq": {
            "acc": 0.536697247706422,
            "acc_stderr": 0.008721469450159985
        },
        "arc_easy": {
            "acc": 0.5357744107744108,
            "acc_stderr": 0.010233488709726539,
            "acc_norm": 0.4692760942760943,
            "acc_norm_stderr": 0.01024039558481524
        },
        "arc_challenge": {
            "acc": 0.22610921501706485,
            "acc_stderr": 0.012224202097063283,
            "acc_norm": 0.26023890784982934,
            "acc_norm_stderr": 0.012821930225112564
        },
        "sciq": {
            "acc": 0.822,
            "acc_stderr": 0.012102167676183582,
            "acc_norm": 0.75,
            "acc_norm_stderr": 0.013699915608779773
        },
        "piqa": {
            "acc": 0.6724700761697497,
            "acc_stderr": 0.010949830482825483,
            "acc_norm": 0.6822633297062024,
            "acc_norm_stderr": 0.01086313324656929
        }
    },
    "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
    }
}