1b11b51b5/evaluation/rankeval/lm1-1b1-1b5-results_lm-eval_global_step2891_2023-01-23-19-58-55_1shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.323,
            "acc_stderr": 0.014794927843348637
        },
        "anli_r2": {
            "acc": 0.319,
            "acc_stderr": 0.01474640486547348
        },
        "anli_r3": {
            "acc": 0.3358333333333333,
            "acc_stderr": 0.013639261190932884
        },
        "cb": {
            "acc": 0.48214285714285715,
            "acc_stderr": 0.06737697508644648,
            "f1": 0.3421052631578947
        },
        "copa": {
            "acc": 0.54,
            "acc_stderr": 0.05009082659620332
        },
        "hellaswag": {
            "acc": 0.2638916550487951,
            "acc_stderr": 0.004398404992933851,
            "acc_norm": 0.26419040031866164,
            "acc_norm_stderr": 0.004400000822742047
        },
        "rte": {
            "acc": 0.5306859205776173,
            "acc_stderr": 0.03003973059219781
        },
        "winogrande": {
            "acc": 0.5090765588003157,
            "acc_stderr": 0.014050170094497704
        },
        "storycloze_2016": {
            "acc": 0.5189738107963656,
            "acc_stderr": 0.011554104174019694
        },
        "boolq": {
            "acc": 0.5226299694189602,
            "acc_stderr": 0.008736093428015826
        },
        "arc_easy": {
            "acc": 0.34764309764309764,
            "acc_stderr": 0.009771868846830912,
            "acc_norm": 0.3186026936026936,
            "acc_norm_stderr": 0.009560775507673364
        },
        "arc_challenge": {
            "acc": 0.17235494880546076,
            "acc_stderr": 0.011037113093461295,
            "acc_norm": 0.2098976109215017,
            "acc_norm_stderr": 0.011900548748047446
        },
        "sciq": {
            "acc": 0.568,
            "acc_stderr": 0.01567232023733621,
            "acc_norm": 0.546,
            "acc_norm_stderr": 0.015752210388771844
        },
        "piqa": {
            "acc": 0.5723612622415669,
            "acc_stderr": 0.011543009623282828,
            "acc_norm": 0.5652883569096845,
            "acc_norm_stderr": 0.011565943814308855
        }
    },
    "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
    }
}