4b284b12bc4/evaluation/4b284b12bc4_4_lm-eval_global_step80108_2023-01-30-11-26-32_4shots_backup.json

{
    "results": {
        "anli_r1": {
            "acc": 0.337,
            "acc_stderr": 0.014955087918653603
        },
        "anli_r2": {
            "acc": 0.349,
            "acc_stderr": 0.015080663991563102
        },
        "anli_r3": {
            "acc": 0.36666666666666664,
            "acc_stderr": 0.013916893275819938
        },
        "cb": {
            "acc": 0.44642857142857145,
            "acc_stderr": 0.067031892279424,
            "f1": 0.3176100628930817
        },
        "copa": {
            "acc": 0.8,
            "acc_stderr": 0.040201512610368445
        },
        "hellaswag": {
            "acc": 0.4722166899024099,
            "acc_stderr": 0.004982072108448081,
            "acc_norm": 0.6184027086237801,
            "acc_norm_stderr": 0.004847857546957481
        },
        "rte": {
            "acc": 0.5379061371841155,
            "acc_stderr": 0.03000984891252911
        },
        "winogrande": {
            "acc": 0.56353591160221,
            "acc_stderr": 0.013938569465677023
        },
        "storycloze_2016": {
            "acc": 0.7194013896312133,
            "acc_stderr": 0.010389809647288821
        },
        "boolq": {
            "acc": 0.5636085626911315,
            "acc_stderr": 0.008674000467432068
        },
        "arc_easy": {
            "acc": 0.6039562289562289,
            "acc_stderr": 0.010035580962097942,
            "acc_norm": 0.5702861952861953,
            "acc_norm_stderr": 0.010157908005763674
        },
        "arc_challenge": {
            "acc": 0.2790102389078498,
            "acc_stderr": 0.013106784883601346,
            "acc_norm": 0.3165529010238908,
            "acc_norm_stderr": 0.013592431519068077
        },
        "sciq": {
            "acc": 0.842,
            "acc_stderr": 0.011539894677559568,
            "acc_norm": 0.789,
            "acc_norm_stderr": 0.012909130321042092
        },
        "piqa": {
            "acc": 0.7431991294885746,
            "acc_stderr": 0.010192864802278045,
            "acc_norm": 0.7568008705114254,
            "acc_norm_stderr": 0.010009611953858915
        }
    },
    "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
    }
}