574m174b174b/evaluation/rankeval/lm1-574m-174b-results_lm-eval_global_step331103_2023-01-24-17-05-33_5shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.318,
            "acc_stderr": 0.014734079309311901
        },
        "anli_r2": {
            "acc": 0.351,
            "acc_stderr": 0.0151005637983164
        },
        "anli_r3": {
            "acc": 0.32916666666666666,
            "acc_stderr": 0.013570806258433625
        },
        "cb": {
            "acc": 0.4642857142857143,
            "acc_stderr": 0.06724777654937658,
            "f1": 0.3288355822088955
        },
        "copa": {
            "acc": 0.68,
            "acc_stderr": 0.04688261722621504
        },
        "hellaswag": {
            "acc": 0.34256124278032263,
            "acc_stderr": 0.004735962781136079,
            "acc_norm": 0.4219279028082055,
            "acc_norm_stderr": 0.00492857810602637
        },
        "rte": {
            "acc": 0.555956678700361,
            "acc_stderr": 0.02990739633379599
        },
        "winogrande": {
            "acc": 0.5311760063141279,
            "acc_stderr": 0.014025142640639513
        },
        "storycloze_2016": {
            "acc": 0.6440406199893105,
            "acc_stderr": 0.01107225418438284
        },
        "boolq": {
            "acc": 0.4889908256880734,
            "acc_stderr": 0.008742934884517644
        },
        "arc_easy": {
            "acc": 0.5593434343434344,
            "acc_stderr": 0.010187264635711986,
            "acc_norm": 0.5467171717171717,
            "acc_norm_stderr": 0.010214901516731604
        },
        "arc_challenge": {
            "acc": 0.24232081911262798,
            "acc_stderr": 0.012521593295800118,
            "acc_norm": 0.27986348122866894,
            "acc_norm_stderr": 0.013119040897725923
        },
        "sciq": {
            "acc": 0.903,
            "acc_stderr": 0.009363689373248128,
            "acc_norm": 0.906,
            "acc_norm_stderr": 0.009233052000787731
        },
        "piqa": {
            "acc": 0.6751904243743199,
            "acc_stderr": 0.01092629623829403,
            "acc_norm": 0.6789989118607181,
            "acc_norm_stderr": 0.010892641574707903
        }
    },
    "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
    }
}