221m60b60b/evaluation/lm1-221m-60b-results_lm-eval_global_step115203_2023-01-24-13-53-29_2shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.313,
            "acc_stderr": 0.014671272822977885
        },
        "anli_r2": {
            "acc": 0.334,
            "acc_stderr": 0.014922019523732963
        },
        "anli_r3": {
            "acc": 0.33,
            "acc_stderr": 0.01357953127780092
        },
        "cb": {
            "acc": 0.4107142857142857,
            "acc_stderr": 0.06633634150359541,
            "f1": 0.24493628437290407
        },
        "copa": {
            "acc": 0.64,
            "acc_stderr": 0.04824181513244218
        },
        "hellaswag": {
            "acc": 0.2894841665006971,
            "acc_stderr": 0.004525960965551705,
            "acc_norm": 0.3194582752439753,
            "acc_norm_stderr": 0.004653138360948109
        },
        "rte": {
            "acc": 0.45126353790613716,
            "acc_stderr": 0.029953149241808946
        },
        "winogrande": {
            "acc": 0.510655090765588,
            "acc_stderr": 0.014049294536290393
        },
        "storycloze_2016": {
            "acc": 0.5873864243719936,
            "acc_stderr": 0.011384472322969045
        },
        "boolq": {
            "acc": 0.5140672782874618,
            "acc_stderr": 0.008741593202770605
        },
        "arc_easy": {
            "acc": 0.45580808080808083,
            "acc_stderr": 0.01021963176343785,
            "acc_norm": 0.4377104377104377,
            "acc_norm_stderr": 0.010179856486006916
        },
        "arc_challenge": {
            "acc": 0.19880546075085323,
            "acc_stderr": 0.011662850198175543,
            "acc_norm": 0.22525597269624573,
            "acc_norm_stderr": 0.01220783999540732
        },
        "sciq": {
            "acc": 0.825,
            "acc_stderr": 0.012021627157731972,
            "acc_norm": 0.793,
            "acc_norm_stderr": 0.012818553557843983
        },
        "piqa": {
            "acc": 0.6360174102285092,
            "acc_stderr": 0.011225875703487176,
            "acc_norm": 0.6316648531011969,
            "acc_norm_stderr": 0.011254089354334354
        }
    },
    "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
    }
}