1b58b88b8/1b58b88b8pile/evaluation/rankeval/lm1-1b5-8b8-results_lm-eval_global_step16765_2023-01-24-17-05-36_5shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.341,
            "acc_stderr": 0.014998131348402697
        },
        "anli_r2": {
            "acc": 0.349,
            "acc_stderr": 0.0150806639915631
        },
        "anli_r3": {
            "acc": 0.35333333333333333,
            "acc_stderr": 0.01380457216231493
        },
        "cb": {
            "acc": 0.5357142857142857,
            "acc_stderr": 0.06724777654937658,
            "f1": 0.3364632237871675
        },
        "copa": {
            "acc": 0.64,
            "acc_stderr": 0.04824181513244218
        },
        "hellaswag": {
            "acc": 0.2863971320454093,
            "acc_stderr": 0.0045115330394062255,
            "acc_norm": 0.30940051782513445,
            "acc_norm_stderr": 0.004613018101185303
        },
        "rte": {
            "acc": 0.49458483754512633,
            "acc_stderr": 0.030094698123239966
        },
        "winogrande": {
            "acc": 0.5138121546961326,
            "acc_stderr": 0.014047122916440419
        },
        "storycloze_2016": {
            "acc": 0.5686798503474078,
            "acc_stderr": 0.011452834647978421
        },
        "boolq": {
            "acc": 0.5669724770642202,
            "acc_stderr": 0.008666251305518052
        },
        "arc_easy": {
            "acc": 0.45580808080808083,
            "acc_stderr": 0.01021963176343785,
            "acc_norm": 0.43434343434343436,
            "acc_norm_stderr": 0.010170943451269428
        },
        "arc_challenge": {
            "acc": 0.2098976109215017,
            "acc_stderr": 0.011900548748047442,
            "acc_norm": 0.24658703071672355,
            "acc_norm_stderr": 0.012595726268790136
        },
        "sciq": {
            "acc": 0.806,
            "acc_stderr": 0.012510816141264374,
            "acc_norm": 0.783,
            "acc_norm_stderr": 0.01304151375727071
        },
        "piqa": {
            "acc": 0.6169749727965179,
            "acc_stderr": 0.011342081709082855,
            "acc_norm": 0.6104461371055495,
            "acc_norm_stderr": 0.011377657455130238
        }
    },
    "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
    }
}