619m2b72b7/evaluation/lm1-619m-2b7-results_lm-eval_global_step5111_2023-01-24-13-53-29_2shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.312,
            "acc_stderr": 0.014658474370509
        },
        "anli_r2": {
            "acc": 0.343,
            "acc_stderr": 0.015019206922356951
        },
        "anli_r3": {
            "acc": 0.3433333333333333,
            "acc_stderr": 0.01371263383046586
        },
        "cb": {
            "acc": 0.375,
            "acc_stderr": 0.06527912098338669,
            "f1": 0.26285714285714284
        },
        "copa": {
            "acc": 0.61,
            "acc_stderr": 0.04902071300001975
        },
        "hellaswag": {
            "acc": 0.26847241585341564,
            "acc_stderr": 0.00442259026238513,
            "acc_norm": 0.2740489942242581,
            "acc_norm_stderr": 0.004451222241494057
        },
        "rte": {
            "acc": 0.5234657039711191,
            "acc_stderr": 0.030063300411902666
        },
        "winogrande": {
            "acc": 0.5138121546961326,
            "acc_stderr": 0.014047122916440412
        },
        "storycloze_2016": {
            "acc": 0.5339390700160342,
            "acc_stderr": 0.01153576488164141
        },
        "boolq": {
            "acc": 0.5782874617737003,
            "acc_stderr": 0.008637194202160971
        },
        "arc_easy": {
            "acc": 0.36826599326599324,
            "acc_stderr": 0.009897286209010894,
            "acc_norm": 0.3468013468013468,
            "acc_norm_stderr": 0.009766326091716005
        },
        "arc_challenge": {
            "acc": 0.15443686006825938,
            "acc_stderr": 0.010560149230392597,
            "acc_norm": 0.197098976109215,
            "acc_norm_stderr": 0.011625047669880624
        },
        "sciq": {
            "acc": 0.67,
            "acc_stderr": 0.014876872027456736,
            "acc_norm": 0.635,
            "acc_norm_stderr": 0.0152317762262649
        },
        "piqa": {
            "acc": 0.5767138193688792,
            "acc_stderr": 0.011527699473614478,
            "acc_norm": 0.5832426550598476,
            "acc_norm_stderr": 0.011503015163618312
        }
    },
    "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
    }
}