1b58b88b8/1b58b88b8pile/evaluation/rankeval/lm1-1b5-8b8-results_lm-eval_global_step16765_2023-01-24-13-57-09_4shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.358,
            "acc_stderr": 0.015167928865407557
        },
        "anli_r2": {
            "acc": 0.352,
            "acc_stderr": 0.015110404505648668
        },
        "anli_r3": {
            "acc": 0.33666666666666667,
            "acc_stderr": 0.013647602942406394
        },
        "cb": {
            "acc": 0.42857142857142855,
            "acc_stderr": 0.06672848092813058,
            "f1": 0.27354497354497354
        },
        "copa": {
            "acc": 0.66,
            "acc_stderr": 0.04760952285695237
        },
        "hellaswag": {
            "acc": 0.28450507866958774,
            "acc_stderr": 0.004502563079349396,
            "acc_norm": 0.308105954989046,
            "acc_norm_stderr": 0.004607669909914983
        },
        "rte": {
            "acc": 0.49097472924187724,
            "acc_stderr": 0.030091559826331334
        },
        "winogrande": {
            "acc": 0.489344909234412,
            "acc_stderr": 0.014049294536290403
        },
        "storycloze_2016": {
            "acc": 0.5708177445216461,
            "acc_stderr": 0.011445869980247976
        },
        "boolq": {
            "acc": 0.5663608562691131,
            "acc_stderr": 0.008667690464344678
        },
        "arc_easy": {
            "acc": 0.45580808080808083,
            "acc_stderr": 0.010219631763437851,
            "acc_norm": 0.4292929292929293,
            "acc_norm_stderr": 0.01015667807591109
        },
        "arc_challenge": {
            "acc": 0.20648464163822525,
            "acc_stderr": 0.011828865619002316,
            "acc_norm": 0.24658703071672355,
            "acc_norm_stderr": 0.012595726268790127
        },
        "sciq": {
            "acc": 0.797,
            "acc_stderr": 0.012726073744598288,
            "acc_norm": 0.763,
            "acc_norm_stderr": 0.013454070462577962
        },
        "piqa": {
            "acc": 0.6158868335146899,
            "acc_stderr": 0.011348160741479145,
            "acc_norm": 0.6039173014145811,
            "acc_norm_stderr": 0.011411089031912477
        }
    },
    "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
    }
}