280m5b95b9/280m5b95b9pile/evaluation/rankeval/lm1-280m-5b9-results_lm-eval_global_step11269_2023-01-23-19-58-52_1shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.322,
            "acc_stderr": 0.014782913600996667
        },
        "anli_r2": {
            "acc": 0.335,
            "acc_stderr": 0.014933117490932573
        },
        "anli_r3": {
            "acc": 0.3333333333333333,
            "acc_stderr": 0.013613950010225606
        },
        "cb": {
            "acc": 0.4107142857142857,
            "acc_stderr": 0.0663363415035954,
            "f1": 0.2751322751322751
        },
        "copa": {
            "acc": 0.62,
            "acc_stderr": 0.048783173121456316
        },
        "hellaswag": {
            "acc": 0.26837283409679347,
            "acc_stderr": 0.004422070927212535,
            "acc_norm": 0.27504481179047996,
            "acc_norm_stderr": 0.004456242601950632
        },
        "rte": {
            "acc": 0.5415162454873647,
            "acc_stderr": 0.02999253538537331
        },
        "winogrande": {
            "acc": 0.5240726124704025,
            "acc_stderr": 0.014036189665395132
        },
        "storycloze_2016": {
            "acc": 0.5451630144307856,
            "acc_stderr": 0.011515167912227987
        },
        "boolq": {
            "acc": 0.5452599388379205,
            "acc_stderr": 0.00870915345549762
        },
        "arc_easy": {
            "acc": 0.380050505050505,
            "acc_stderr": 0.009960175831493116,
            "acc_norm": 0.36363636363636365,
            "acc_norm_stderr": 0.009870849346011767
        },
        "arc_challenge": {
            "acc": 0.1757679180887372,
            "acc_stderr": 0.011122850863120485,
            "acc_norm": 0.2158703071672355,
            "acc_norm_stderr": 0.012022975360030663
        },
        "sciq": {
            "acc": 0.707,
            "acc_stderr": 0.014399942998441271,
            "acc_norm": 0.659,
            "acc_norm_stderr": 0.014998131348402697
        },
        "piqa": {
            "acc": 0.5984766050054406,
            "acc_stderr": 0.01143732437339785,
            "acc_norm": 0.5903155603917302,
            "acc_norm_stderr": 0.011473932007187613
        }
    },
    "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
    }
}