146m14b14b/evaluation/rankeval/lm1-146m-14b-results_lm-eval_global_step21553_2023-01-24-17-05-33_5shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.343,
            "acc_stderr": 0.015019206922356951
        },
        "anli_r2": {
            "acc": 0.338,
            "acc_stderr": 0.014965960710224487
        },
        "anli_r3": {
            "acc": 0.3333333333333333,
            "acc_stderr": 0.013613950010225598
        },
        "cb": {
            "acc": 0.5178571428571429,
            "acc_stderr": 0.06737697508644648,
            "f1": 0.3607787942763312
        },
        "copa": {
            "acc": 0.62,
            "acc_stderr": 0.048783173121456316
        },
        "hellaswag": {
            "acc": 0.26976697868950406,
            "acc_stderr": 0.004429315788310528,
            "acc_norm": 0.2778331009759012,
            "acc_norm_stderr": 0.004470152081675125
        },
        "rte": {
            "acc": 0.5667870036101083,
            "acc_stderr": 0.029826764082138264
        },
        "winogrande": {
            "acc": 0.5240726124704025,
            "acc_stderr": 0.014036189665395125
        },
        "storycloze_2016": {
            "acc": 0.5547835382148584,
            "acc_stderr": 0.01149281951929236
        },
        "boolq": {
            "acc": 0.609480122324159,
            "acc_stderr": 0.008532845556631466
        },
        "arc_easy": {
            "acc": 0.3968855218855219,
            "acc_stderr": 0.010039236800583202,
            "acc_norm": 0.36784511784511786,
            "acc_norm_stderr": 0.009894923464455196
        },
        "arc_challenge": {
            "acc": 0.1680887372013652,
            "acc_stderr": 0.01092771504612486,
            "acc_norm": 0.21416382252559726,
            "acc_norm_stderr": 0.011988383205966494
        },
        "sciq": {
            "acc": 0.737,
            "acc_stderr": 0.013929286594259715,
            "acc_norm": 0.715,
            "acc_norm_stderr": 0.014282120955200478
        },
        "piqa": {
            "acc": 0.5908596300326442,
            "acc_stderr": 0.011471593460443316,
            "acc_norm": 0.588683351468988,
            "acc_norm_stderr": 0.011480860577192817
        }
    },
    "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
    }
}