1b11b51b5/evaluation/rankeval/lm1-1b1-1b5-results_lm-eval_global_step2891_2023-01-24-13-57-07_3shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.314,
            "acc_stderr": 0.014683991951087976
        },
        "anli_r2": {
            "acc": 0.336,
            "acc_stderr": 0.014944140233795023
        },
        "anli_r3": {
            "acc": 0.32166666666666666,
            "acc_stderr": 0.013490095282989521
        },
        "cb": {
            "acc": 0.4107142857142857,
            "acc_stderr": 0.0663363415035954,
            "f1": 0.286512928022362
        },
        "copa": {
            "acc": 0.6,
            "acc_stderr": 0.049236596391733084
        },
        "hellaswag": {
            "acc": 0.2657837084246166,
            "acc_stderr": 0.004408468107262735,
            "acc_norm": 0.2660824536944832,
            "acc_norm_stderr": 0.00441004753083503
        },
        "rte": {
            "acc": 0.4729241877256318,
            "acc_stderr": 0.030052303463143706
        },
        "winogrande": {
            "acc": 0.5114443567482242,
            "acc_stderr": 0.014048804199859332
        },
        "storycloze_2016": {
            "acc": 0.5173703901656868,
            "acc_stderr": 0.011555452669106639
        },
        "boolq": {
            "acc": 0.6021406727828746,
            "acc_stderr": 0.008560641169303364
        },
        "arc_easy": {
            "acc": 0.3265993265993266,
            "acc_stderr": 0.009623047038267647,
            "acc_norm": 0.31734006734006737,
            "acc_norm_stderr": 0.009550648343947771
        },
        "arc_challenge": {
            "acc": 0.17747440273037543,
            "acc_stderr": 0.01116513876964396,
            "acc_norm": 0.20819112627986347,
            "acc_norm_stderr": 0.011864866118448069
        },
        "sciq": {
            "acc": 0.562,
            "acc_stderr": 0.01569721001969469,
            "acc_norm": 0.552,
            "acc_norm_stderr": 0.015733516566347833
        },
        "piqa": {
            "acc": 0.5788900979325353,
            "acc_stderr": 0.01151970105915149,
            "acc_norm": 0.5680087051142546,
            "acc_norm_stderr": 0.01155740721010026
        }
    },
    "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
    }
}