1b58b88b8/1b58b88b8pile/evaluation/rankeval/lm1-1b5-8b8-results_lm-eval_global_step16765_2023-01-24-13-57-09_3shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.321,
            "acc_stderr": 0.014770821817934645
        },
        "anli_r2": {
            "acc": 0.35,
            "acc_stderr": 0.015090650341444233
        },
        "anli_r3": {
            "acc": 0.3433333333333333,
            "acc_stderr": 0.01371263383046586
        },
        "cb": {
            "acc": 0.5714285714285714,
            "acc_stderr": 0.06672848092813058,
            "f1": 0.3720462543991956
        },
        "copa": {
            "acc": 0.64,
            "acc_stderr": 0.04824181513244218
        },
        "hellaswag": {
            "acc": 0.28560047799243177,
            "acc_stderr": 0.004507768029590114,
            "acc_norm": 0.30860386377215693,
            "acc_norm_stderr": 0.004609731925736893
        },
        "rte": {
            "acc": 0.5270758122743683,
            "acc_stderr": 0.030052303463143706
        },
        "winogrande": {
            "acc": 0.5209155485398579,
            "acc_stderr": 0.01404018549421294
        },
        "storycloze_2016": {
            "acc": 0.5665419561731694,
            "acc_stderr": 0.011459581799087402
        },
        "boolq": {
            "acc": 0.5629969418960244,
            "acc_stderr": 0.008675365793227086
        },
        "arc_easy": {
            "acc": 0.45286195286195285,
            "acc_stderr": 0.010214087372211392,
            "acc_norm": 0.4276094276094276,
            "acc_norm_stderr": 0.010151683397430679
        },
        "arc_challenge": {
            "acc": 0.20477815699658702,
            "acc_stderr": 0.011792544338513409,
            "acc_norm": 0.24914675767918087,
            "acc_norm_stderr": 0.012639407111926433
        },
        "sciq": {
            "acc": 0.793,
            "acc_stderr": 0.012818553557843981,
            "acc_norm": 0.76,
            "acc_norm_stderr": 0.013512312258920845
        },
        "piqa": {
            "acc": 0.6175190424374319,
            "acc_stderr": 0.011339019654272345,
            "acc_norm": 0.6164309031556039,
            "acc_norm_stderr": 0.01134512873411628
        }
    },
    "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
    }
}