1b11b51b5/evaluation/rankeval/lm1-1b1-1b5-results_lm-eval_global_step2891_2023-01-24-13-53-33_2shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.318,
            "acc_stderr": 0.014734079309311901
        },
        "anli_r2": {
            "acc": 0.338,
            "acc_stderr": 0.014965960710224473
        },
        "anli_r3": {
            "acc": 0.3275,
            "acc_stderr": 0.01355321116725195
        },
        "cb": {
            "acc": 0.48214285714285715,
            "acc_stderr": 0.06737697508644648,
            "f1": 0.33636151875145587
        },
        "copa": {
            "acc": 0.55,
            "acc_stderr": 0.04999999999999999
        },
        "hellaswag": {
            "acc": 0.26379207329217286,
            "acc_stderr": 0.00439787247185492,
            "acc_norm": 0.2671778530173272,
            "acc_norm_stderr": 0.004415816696303073
        },
        "rte": {
            "acc": 0.49097472924187724,
            "acc_stderr": 0.030091559826331334
        },
        "winogrande": {
            "acc": 0.5146014206787688,
            "acc_stderr": 0.014046492383275837
        },
        "storycloze_2016": {
            "acc": 0.5189738107963656,
            "acc_stderr": 0.011554104174019695
        },
        "boolq": {
            "acc": 0.5850152905198777,
            "acc_stderr": 0.008617716361921567
        },
        "arc_easy": {
            "acc": 0.3303872053872054,
            "acc_stderr": 0.009651430216428194,
            "acc_norm": 0.3253367003367003,
            "acc_norm_stderr": 0.00961342770899619
        },
        "arc_challenge": {
            "acc": 0.1621160409556314,
            "acc_stderr": 0.01077026738079835,
            "acc_norm": 0.2090443686006826,
            "acc_norm_stderr": 0.011882746987406446
        },
        "sciq": {
            "acc": 0.547,
            "acc_stderr": 0.015749255189977586,
            "acc_norm": 0.534,
            "acc_norm_stderr": 0.015782683329937614
        },
        "piqa": {
            "acc": 0.573993471164309,
            "acc_stderr": 0.011537375448519445,
            "acc_norm": 0.5652883569096845,
            "acc_norm_stderr": 0.011565943814308855
        }
    },
    "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
    }
}