4b284b12bc4/evaluation/4b284b12bc4_3_lm-eval_global_step80108_2023-01-30-11-26-31_3shots_backup.json

{
    "results": {
        "anli_r1": {
            "acc": 0.332,
            "acc_stderr": 0.014899597242811485
        },
        "anli_r2": {
            "acc": 0.334,
            "acc_stderr": 0.014922019523732963
        },
        "anli_r3": {
            "acc": 0.35,
            "acc_stderr": 0.013774667009018554
        },
        "cb": {
            "acc": 0.6071428571428571,
            "acc_stderr": 0.0658538889806635,
            "f1": 0.42400932400932395
        },
        "copa": {
            "acc": 0.81,
            "acc_stderr": 0.03942772444036622
        },
        "hellaswag": {
            "acc": 0.47241585341565423,
            "acc_stderr": 0.004982182323923561,
            "acc_norm": 0.6199960167297351,
            "acc_norm_stderr": 0.004843954338451449
        },
        "rte": {
            "acc": 0.5379061371841155,
            "acc_stderr": 0.030009848912529113
        },
        "winogrande": {
            "acc": 0.5737963693764798,
            "acc_stderr": 0.013898585965412338
        },
        "storycloze_2016": {
            "acc": 0.7124532335649385,
            "acc_stderr": 0.010466744473098363
        },
        "boolq": {
            "acc": 0.5587155963302752,
            "acc_stderr": 0.008684548127832637
        },
        "arc_easy": {
            "acc": 0.5955387205387206,
            "acc_stderr": 0.010070746648278783,
            "acc_norm": 0.5740740740740741,
            "acc_norm_stderr": 0.010146568651002255
        },
        "arc_challenge": {
            "acc": 0.2815699658703072,
            "acc_stderr": 0.013143376735009022,
            "acc_norm": 0.3122866894197952,
            "acc_norm_stderr": 0.013542598541688067
        },
        "sciq": {
            "acc": 0.841,
            "acc_stderr": 0.01156947936827129,
            "acc_norm": 0.796,
            "acc_norm_stderr": 0.012749374359024384
        },
        "piqa": {
            "acc": 0.7513601741022851,
            "acc_stderr": 0.01008451123429685,
            "acc_norm": 0.7578890097932536,
            "acc_norm_stderr": 0.009994371269104397
        }
    },
    "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
    }
}