1b191b91b/1b191b91bpile/evaluation/rankeval/lm1-1b1-91b-results_lm-eval_global_step173500_2023-01-23-19-59-49_1shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.329,
            "acc_stderr": 0.014865395385928367
        },
        "anli_r2": {
            "acc": 0.318,
            "acc_stderr": 0.014734079309311901
        },
        "anli_r3": {
            "acc": 0.34,
            "acc_stderr": 0.013680495725767797
        },
        "cb": {
            "acc": 0.48214285714285715,
            "acc_stderr": 0.06737697508644648,
            "f1": 0.3421052631578947
        },
        "copa": {
            "acc": 0.69,
            "acc_stderr": 0.04648231987117316
        },
        "hellaswag": {
            "acc": 0.3619796853216491,
            "acc_stderr": 0.004795908282584554,
            "acc_norm": 0.4466241784505079,
            "acc_norm_stderr": 0.004961268387512966
        },
        "rte": {
            "acc": 0.5270758122743683,
            "acc_stderr": 0.0300523034631437
        },
        "winogrande": {
            "acc": 0.5374901341752171,
            "acc_stderr": 0.014012928183336574
        },
        "storycloze_2016": {
            "acc": 0.6509887760555852,
            "acc_stderr": 0.011022640519108541
        },
        "boolq": {
            "acc": 0.4996941896024465,
            "acc_stderr": 0.008745053340723166
        },
        "arc_easy": {
            "acc": 0.5597643097643098,
            "acc_stderr": 0.010186228624515655,
            "acc_norm": 0.5446127946127947,
            "acc_norm_stderr": 0.010218861787618721
        },
        "arc_challenge": {
            "acc": 0.24744027303754265,
            "acc_stderr": 0.01261035266329267,
            "acc_norm": 0.27986348122866894,
            "acc_norm_stderr": 0.013119040897725923
        },
        "sciq": {
            "acc": 0.893,
            "acc_stderr": 0.009779910359847167,
            "acc_norm": 0.89,
            "acc_norm_stderr": 0.009899393819724439
        },
        "piqa": {
            "acc": 0.6887921653971708,
            "acc_stderr": 0.01080226387804584,
            "acc_norm": 0.6860718171926007,
            "acc_norm_stderr": 0.010827928134189646
        }
    },
    "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
    }
}