2b246b46b/evaluation/lm1-2b2-46b-results_lm-eval_global_step44073_2023-01-24-13-53-38_2shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.328,
            "acc_stderr": 0.01485384248727033
        },
        "anli_r2": {
            "acc": 0.328,
            "acc_stderr": 0.014853842487270333
        },
        "anli_r3": {
            "acc": 0.325,
            "acc_stderr": 0.013526454480351025
        },
        "cb": {
            "acc": 0.4107142857142857,
            "acc_stderr": 0.0663363415035954,
            "f1": 0.29069767441860467
        },
        "copa": {
            "acc": 0.74,
            "acc_stderr": 0.0440844002276808
        },
        "hellaswag": {
            "acc": 0.351822346146186,
            "acc_stderr": 0.004765629263643526,
            "acc_norm": 0.43537143995220073,
            "acc_norm_stderr": 0.004947922692688831
        },
        "rte": {
            "acc": 0.49458483754512633,
            "acc_stderr": 0.030094698123239966
        },
        "winogrande": {
            "acc": 0.5438042620363063,
            "acc_stderr": 0.013998453610924324
        },
        "storycloze_2016": {
            "acc": 0.6424371993586317,
            "acc_stderr": 0.01108334116882779
        },
        "boolq": {
            "acc": 0.4828746177370031,
            "acc_stderr": 0.008739923994130054
        },
        "arc_easy": {
            "acc": 0.5618686868686869,
            "acc_stderr": 0.010180937100600076,
            "acc_norm": 0.5441919191919192,
            "acc_norm_stderr": 0.010219631763437851
        },
        "arc_challenge": {
            "acc": 0.25170648464163825,
            "acc_stderr": 0.012682496334042961,
            "acc_norm": 0.28924914675767915,
            "acc_norm_stderr": 0.013250012579393443
        },
        "sciq": {
            "acc": 0.902,
            "acc_stderr": 0.009406619184621224,
            "acc_norm": 0.901,
            "acc_norm_stderr": 0.009449248027662732
        },
        "piqa": {
            "acc": 0.6822633297062024,
            "acc_stderr": 0.010863133246569285,
            "acc_norm": 0.6806311207834603,
            "acc_norm_stderr": 0.010877964076613742
        }
    },
    "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
    }
}