619m2b72b7/evaluation/rankeval/lm1-619m-2b7-results_lm-eval_global_step5111_2023-01-24-13-57-06_3shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.325,
            "acc_stderr": 0.014818724459095524
        },
        "anli_r2": {
            "acc": 0.347,
            "acc_stderr": 0.01506047203170662
        },
        "anli_r3": {
            "acc": 0.32916666666666666,
            "acc_stderr": 0.01357080625843363
        },
        "cb": {
            "acc": 0.42857142857142855,
            "acc_stderr": 0.06672848092813058,
            "f1": 0.29871345029239765
        },
        "copa": {
            "acc": 0.6,
            "acc_stderr": 0.049236596391733084
        },
        "hellaswag": {
            "acc": 0.2704640509858594,
            "acc_stderr": 0.004432917403755054,
            "acc_norm": 0.2746464847639912,
            "acc_norm_stderr": 0.004454237797448344
        },
        "rte": {
            "acc": 0.5090252707581228,
            "acc_stderr": 0.030091559826331334
        },
        "winogrande": {
            "acc": 0.5146014206787688,
            "acc_stderr": 0.014046492383275832
        },
        "storycloze_2016": {
            "acc": 0.5334045964724746,
            "acc_stderr": 0.011536599118298168
        },
        "boolq": {
            "acc": 0.5926605504587156,
            "acc_stderr": 0.00859357330260705
        },
        "arc_easy": {
            "acc": 0.35774410774410775,
            "acc_stderr": 0.00983577275734336,
            "acc_norm": 0.351010101010101,
            "acc_norm_stderr": 0.009793703885101049
        },
        "arc_challenge": {
            "acc": 0.16296928327645052,
            "acc_stderr": 0.010793073338782481,
            "acc_norm": 0.19368600682593856,
            "acc_norm_stderr": 0.01154842540997854
        },
        "sciq": {
            "acc": 0.656,
            "acc_stderr": 0.015029633724408943,
            "acc_norm": 0.623,
            "acc_norm_stderr": 0.01533317012577987
        },
        "piqa": {
            "acc": 0.5870511425462459,
            "acc_stderr": 0.011487658725079095,
            "acc_norm": 0.5756256800870512,
            "acc_norm_stderr": 0.011531612758871062
        }
    },
    "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
    }
}