619m22b22b/evaluation/rankeval/lm1-619m-22b-results_lm-eval_global_step41007_2023-01-24-13-57-03_3shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.326,
            "acc_stderr": 0.014830507204541024
        },
        "anli_r2": {
            "acc": 0.358,
            "acc_stderr": 0.015167928865407557
        },
        "anli_r3": {
            "acc": 0.3375,
            "acc_stderr": 0.013655897185463665
        },
        "cb": {
            "acc": 0.35714285714285715,
            "acc_stderr": 0.0646095738380922,
            "f1": 0.23561507936507933
        },
        "copa": {
            "acc": 0.69,
            "acc_stderr": 0.04648231987117316
        },
        "hellaswag": {
            "acc": 0.30233021310495917,
            "acc_stderr": 0.004583289072937737,
            "acc_norm": 0.3353913563035252,
            "acc_norm_stderr": 0.004711622011148457
        },
        "rte": {
            "acc": 0.5415162454873647,
            "acc_stderr": 0.029992535385373314
        },
        "winogrande": {
            "acc": 0.5193370165745856,
            "acc_stderr": 0.01404197273371297
        },
        "storycloze_2016": {
            "acc": 0.6002137894174239,
            "acc_stderr": 0.011327813397531862
        },
        "boolq": {
            "acc": 0.5299694189602446,
            "acc_stderr": 0.0087293318183149
        },
        "arc_easy": {
            "acc": 0.4882154882154882,
            "acc_stderr": 0.010256933475911015,
            "acc_norm": 0.4671717171717172,
            "acc_norm_stderr": 0.010237645778853851
        },
        "arc_challenge": {
            "acc": 0.2226962457337884,
            "acc_stderr": 0.012158314774829924,
            "acc_norm": 0.25170648464163825,
            "acc_norm_stderr": 0.012682496334042961
        },
        "sciq": {
            "acc": 0.834,
            "acc_stderr": 0.011772110370812192,
            "acc_norm": 0.809,
            "acc_norm_stderr": 0.012436787112179486
        },
        "piqa": {
            "acc": 0.6436343852013058,
            "acc_stderr": 0.011174109865864703,
            "acc_norm": 0.6436343852013058,
            "acc_norm_stderr": 0.011174109865864729
        }
    },
    "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
    }
}