1b566b66b/1b566b66bpile/evaluation/lm1-1b5-66b-results_lm-eval_global_step125429_2023-01-24-13-53-33_2shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.325,
            "acc_stderr": 0.014818724459095524
        },
        "anli_r2": {
            "acc": 0.345,
            "acc_stderr": 0.015039986742055237
        },
        "anli_r3": {
            "acc": 0.33416666666666667,
            "acc_stderr": 0.013622434813136764
        },
        "cb": {
            "acc": 0.375,
            "acc_stderr": 0.06527912098338669,
            "f1": 0.26400759734093066
        },
        "copa": {
            "acc": 0.73,
            "acc_stderr": 0.044619604333847394
        },
        "hellaswag": {
            "acc": 0.3670583549093806,
            "acc_stderr": 0.004810175357871108,
            "acc_norm": 0.458972316271659,
            "acc_norm_stderr": 0.004972954732733358
        },
        "rte": {
            "acc": 0.5306859205776173,
            "acc_stderr": 0.030039730592197816
        },
        "winogrande": {
            "acc": 0.55327545382794,
            "acc_stderr": 0.0139724883716167
        },
        "storycloze_2016": {
            "acc": 0.6584714056654195,
            "acc_stderr": 0.010966327011405767
        },
        "boolq": {
            "acc": 0.5831804281345566,
            "acc_stderr": 0.008623192108843677
        },
        "arc_easy": {
            "acc": 0.5787037037037037,
            "acc_stderr": 0.010131882498193134,
            "acc_norm": 0.5854377104377104,
            "acc_norm_stderr": 0.010108889212447783
        },
        "arc_challenge": {
            "acc": 0.25597269624573377,
            "acc_stderr": 0.012753013241244521,
            "acc_norm": 0.28498293515358364,
            "acc_norm_stderr": 0.013191348179838793
        },
        "sciq": {
            "acc": 0.914,
            "acc_stderr": 0.008870325962594766,
            "acc_norm": 0.912,
            "acc_norm_stderr": 0.008963053962592074
        },
        "piqa": {
            "acc": 0.6871599564744287,
            "acc_stderr": 0.010817714425701112,
            "acc_norm": 0.6953210010881393,
            "acc_norm_stderr": 0.010738889044325161
        }
    },
    "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
    }
}