574m174b174b/evaluation/lm1-574m-174b-results_lm-eval_global_step331103_2023-01-24-13-53-29_2shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.308,
            "acc_stderr": 0.014606483127342765
        },
        "anli_r2": {
            "acc": 0.346,
            "acc_stderr": 0.015050266127564443
        },
        "anli_r3": {
            "acc": 0.3233333333333333,
            "acc_stderr": 0.013508372867300224
        },
        "cb": {
            "acc": 0.32142857142857145,
            "acc_stderr": 0.06297362289056341,
            "f1": 0.23303167420814477
        },
        "copa": {
            "acc": 0.75,
            "acc_stderr": 0.04351941398892446
        },
        "hellaswag": {
            "acc": 0.3419637522405895,
            "acc_stderr": 0.004733980470799225,
            "acc_norm": 0.4202350129456284,
            "acc_norm_stderr": 0.004925877705771198
        },
        "rte": {
            "acc": 0.5090252707581228,
            "acc_stderr": 0.030091559826331334
        },
        "winogrande": {
            "acc": 0.516179952644041,
            "acc_stderr": 0.0140451261309786
        },
        "storycloze_2016": {
            "acc": 0.6435061464457509,
            "acc_stderr": 0.011075964871050996
        },
        "boolq": {
            "acc": 0.4932721712538226,
            "acc_stderr": 0.008744263273827433
        },
        "arc_easy": {
            "acc": 0.5526094276094277,
            "acc_stderr": 0.010202832385415646,
            "acc_norm": 0.5484006734006734,
            "acc_norm_stderr": 0.010211600726405236
        },
        "arc_challenge": {
            "acc": 0.24658703071672355,
            "acc_stderr": 0.012595726268790125,
            "acc_norm": 0.2841296928327645,
            "acc_norm_stderr": 0.013179442447653886
        },
        "sciq": {
            "acc": 0.891,
            "acc_stderr": 0.009859828407037188,
            "acc_norm": 0.884,
            "acc_norm_stderr": 0.010131468138757005
        },
        "piqa": {
            "acc": 0.676278563656148,
            "acc_stderr": 0.010916765010708781,
            "acc_norm": 0.675734494015234,
            "acc_norm_stderr": 0.010921539041347985
        }
    },
    "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
    }
}