1b58b88b8/1b58b88b8pile/evaluation/rankeval/lm1-1b5-8b8-results_lm-eval_global_step16765_2023-01-22-18-55-55_0shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.325,
            "acc_stderr": 0.014818724459095524
        },
        "anli_r2": {
            "acc": 0.345,
            "acc_stderr": 0.015039986742055233
        },
        "anli_r3": {
            "acc": 0.3516666666666667,
            "acc_stderr": 0.013789711695404798
        },
        "cb": {
            "acc": 0.4642857142857143,
            "acc_stderr": 0.0672477765493766,
            "f1": 0.324184846352334
        },
        "copa": {
            "acc": 0.63,
            "acc_stderr": 0.04852365870939099
        },
        "hellaswag": {
            "acc": 0.2862975502887871,
            "acc_stderr": 0.004511063351278702,
            "acc_norm": 0.30501892053375823,
            "acc_norm_stderr": 0.004594744821762281
        },
        "rte": {
            "acc": 0.4620938628158845,
            "acc_stderr": 0.030009848912529113
        },
        "winogrande": {
            "acc": 0.5027624309392266,
            "acc_stderr": 0.014052271211616436
        },
        "storycloze_2016": {
            "acc": 0.5788348476750401,
            "acc_stderr": 0.011417808278216117
        },
        "boolq": {
            "acc": 0.5235474006116208,
            "acc_stderr": 0.008735351675636605
        },
        "arc_easy": {
            "acc": 0.4537037037037037,
            "acc_stderr": 0.010215708295494128,
            "acc_norm": 0.4036195286195286,
            "acc_norm_stderr": 0.010067368960348204
        },
        "arc_challenge": {
            "acc": 0.19539249146757678,
            "acc_stderr": 0.011586907189952911,
            "acc_norm": 0.24658703071672355,
            "acc_norm_stderr": 0.012595726268790124
        },
        "sciq": {
            "acc": 0.768,
            "acc_stderr": 0.013354937452281557,
            "acc_norm": 0.671,
            "acc_norm_stderr": 0.014865395385928354
        },
        "piqa": {
            "acc": 0.6147986942328618,
            "acc_stderr": 0.011354179751257075,
            "acc_norm": 0.6186071817192601,
            "acc_norm_stderr": 0.011332850406528672
        }
    },
    "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
    }
}