280m5b95b9/280m5b95b9pile/evaluation/rankeval/lm1-280m-5b9-results_lm-eval_global_step11269_2023-01-24-13-57-03_3shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.321,
            "acc_stderr": 0.014770821817934654
        },
        "anli_r2": {
            "acc": 0.333,
            "acc_stderr": 0.014910846164229857
        },
        "anli_r3": {
            "acc": 0.3233333333333333,
            "acc_stderr": 0.013508372867300219
        },
        "cb": {
            "acc": 0.4107142857142857,
            "acc_stderr": 0.0663363415035954,
            "f1": 0.27519379844961245
        },
        "copa": {
            "acc": 0.58,
            "acc_stderr": 0.04960449637488583
        },
        "hellaswag": {
            "acc": 0.27106154152559253,
            "acc_stderr": 0.004435993492583858,
            "acc_norm": 0.2756423023302131,
            "acc_norm_stderr": 0.004459241474518783
        },
        "rte": {
            "acc": 0.49458483754512633,
            "acc_stderr": 0.030094698123239966
        },
        "winogrande": {
            "acc": 0.526440410418311,
            "acc_stderr": 0.01403282387440723
        },
        "storycloze_2016": {
            "acc": 0.5398182789951897,
            "acc_stderr": 0.011525709570367517
        },
        "boolq": {
            "acc": 0.591743119266055,
            "acc_stderr": 0.008596583869583204
        },
        "arc_easy": {
            "acc": 0.37962962962962965,
            "acc_stderr": 0.009958037725468558,
            "acc_norm": 0.35984848484848486,
            "acc_norm_stderr": 0.00984848484848485
        },
        "arc_challenge": {
            "acc": 0.18088737201365188,
            "acc_stderr": 0.011248574467407024,
            "acc_norm": 0.21160409556313994,
            "acc_norm_stderr": 0.011935916358632852
        },
        "sciq": {
            "acc": 0.718,
            "acc_stderr": 0.014236526215291343,
            "acc_norm": 0.661,
            "acc_norm_stderr": 0.014976758771620345
        },
        "piqa": {
            "acc": 0.5957562568008705,
            "acc_stderr": 0.01144989176300747,
            "acc_norm": 0.5843307943416758,
            "acc_norm_stderr": 0.011498699770894797
        }
    },
    "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
    }
}