1b11b51b5/evaluation/rankeval/lm1-1b1-1b5-results_lm-eval_global_step2891_2023-01-22-18-55-55_0shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.332,
            "acc_stderr": 0.01489959724281149
        },
        "anli_r2": {
            "acc": 0.308,
            "acc_stderr": 0.01460648312734276
        },
        "anli_r3": {
            "acc": 0.3233333333333333,
            "acc_stderr": 0.013508372867300228
        },
        "cb": {
            "acc": 0.4107142857142857,
            "acc_stderr": 0.0663363415035954,
            "f1": 0.22313546423135466
        },
        "copa": {
            "acc": 0.54,
            "acc_stderr": 0.05009082659620332
        },
        "hellaswag": {
            "acc": 0.26269667396932883,
            "acc_stderr": 0.004391995637542121,
            "acc_norm": 0.264389563831906,
            "acc_norm_stderr": 0.004401063265803207
        },
        "rte": {
            "acc": 0.5342960288808665,
            "acc_stderr": 0.030025579819366422
        },
        "winogrande": {
            "acc": 0.4980268350434096,
            "acc_stderr": 0.014052376259225636
        },
        "storycloze_2016": {
            "acc": 0.5307322287546766,
            "acc_stderr": 0.011540570846495544
        },
        "boolq": {
            "acc": 0.5529051987767584,
            "acc_stderr": 0.00869596306417273
        },
        "arc_easy": {
            "acc": 0.3421717171717172,
            "acc_stderr": 0.00973523677195875,
            "acc_norm": 0.31186868686868685,
            "acc_norm_stderr": 0.009505823345817666
        },
        "arc_challenge": {
            "acc": 0.16638225255972697,
            "acc_stderr": 0.010883248065964142,
            "acc_norm": 0.21245733788395904,
            "acc_norm_stderr": 0.011953482906582952
        },
        "sciq": {
            "acc": 0.548,
            "acc_stderr": 0.01574623586588068,
            "acc_norm": 0.515,
            "acc_norm_stderr": 0.01581217964181491
        },
        "piqa": {
            "acc": 0.5788900979325353,
            "acc_stderr": 0.01151970105915149,
            "acc_norm": 0.5690968443960827,
            "acc_norm_stderr": 0.011553893175901412
        }
    },
    "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
    }
}