619m2b72b7/evaluation/rankeval/lm1-619m-2b7-results_lm-eval_global_step5111_2023-01-24-17-05-34_5shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.34,
            "acc_stderr": 0.014987482264363937
        },
        "anli_r2": {
            "acc": 0.338,
            "acc_stderr": 0.01496596071022449
        },
        "anli_r3": {
            "acc": 0.335,
            "acc_stderr": 0.013630871843821474
        },
        "cb": {
            "acc": 0.48214285714285715,
            "acc_stderr": 0.06737697508644648,
            "f1": 0.33543417366946776
        },
        "copa": {
            "acc": 0.59,
            "acc_stderr": 0.04943110704237102
        },
        "hellaswag": {
            "acc": 0.2678749253136825,
            "acc_stderr": 0.004419469983939183,
            "acc_norm": 0.2747460665206134,
            "acc_norm_stderr": 0.004454739415705043
        },
        "rte": {
            "acc": 0.5451263537906137,
            "acc_stderr": 0.029973636495415255
        },
        "winogrande": {
            "acc": 0.5067087608524072,
            "acc_stderr": 0.014051220692330352
        },
        "storycloze_2016": {
            "acc": 0.535542490646713,
            "acc_stderr": 0.011533182338113991
        },
        "boolq": {
            "acc": 0.6003058103975535,
            "acc_stderr": 0.008567275456584963
        },
        "arc_easy": {
            "acc": 0.35774410774410775,
            "acc_stderr": 0.009835772757343358,
            "acc_norm": 0.343013468013468,
            "acc_norm_stderr": 0.009740965666489224
        },
        "arc_challenge": {
            "acc": 0.17064846416382254,
            "acc_stderr": 0.010993654168413738,
            "acc_norm": 0.20648464163822525,
            "acc_norm_stderr": 0.011828865619002316
        },
        "sciq": {
            "acc": 0.669,
            "acc_stderr": 0.014888272588203943,
            "acc_norm": 0.664,
            "acc_norm_stderr": 0.014944140233795027
        },
        "piqa": {
            "acc": 0.5848748639825898,
            "acc_stderr": 0.011496520442659124,
            "acc_norm": 0.5826985854189336,
            "acc_norm_stderr": 0.01150515124433324
        }
    },
    "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
    }
}