1b11b51b5/evaluation/rankeval/lm1-1b1-1b5-results_lm-eval_global_step2891_2023-01-24-13-57-05_4shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.324,
            "acc_stderr": 0.014806864733738859
        },
        "anli_r2": {
            "acc": 0.34,
            "acc_stderr": 0.014987482264363937
        },
        "anli_r3": {
            "acc": 0.335,
            "acc_stderr": 0.013630871843821479
        },
        "cb": {
            "acc": 0.48214285714285715,
            "acc_stderr": 0.0673769750864465,
            "f1": 0.3356643356643356
        },
        "copa": {
            "acc": 0.57,
            "acc_stderr": 0.049756985195624284
        },
        "hellaswag": {
            "acc": 0.26538538139812784,
            "acc_stderr": 0.004406358190678484,
            "acc_norm": 0.2664807807209719,
            "acc_norm_stderr": 0.004412149415717922
        },
        "rte": {
            "acc": 0.5126353790613718,
            "acc_stderr": 0.030086851767188564
        },
        "winogrande": {
            "acc": 0.5074980268350434,
            "acc_stderr": 0.014050905521228573
        },
        "storycloze_2016": {
            "acc": 0.5227151256012827,
            "acc_stderr": 0.01155049419200895
        },
        "boolq": {
            "acc": 0.6146788990825688,
            "acc_stderr": 0.008511930879680647
        },
        "arc_easy": {
            "acc": 0.32786195286195285,
            "acc_stderr": 0.009632587076170018,
            "acc_norm": 0.3085016835016835,
            "acc_norm_stderr": 0.009477472342978126
        },
        "arc_challenge": {
            "acc": 0.1689419795221843,
            "acc_stderr": 0.010949795652485024,
            "acc_norm": 0.2090443686006826,
            "acc_norm_stderr": 0.011882746987406448
        },
        "sciq": {
            "acc": 0.561,
            "acc_stderr": 0.015701131345400774,
            "acc_norm": 0.554,
            "acc_norm_stderr": 0.015726771166750357
        },
        "piqa": {
            "acc": 0.5701849836779108,
            "acc_stderr": 0.011550322268694083,
            "acc_norm": 0.5522306855277476,
            "acc_norm_stderr": 0.011601999796866812
        }
    },
    "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
    }
}