146m14b14b/evaluation/rankeval/lm1-146m-14b-results_lm-eval_global_step21553_2023-01-24-13-53-29_2shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.32,
            "acc_stderr": 0.014758652303574876
        },
        "anli_r2": {
            "acc": 0.334,
            "acc_stderr": 0.014922019523732963
        },
        "anli_r3": {
            "acc": 0.32916666666666666,
            "acc_stderr": 0.013570806258433633
        },
        "cb": {
            "acc": 0.35714285714285715,
            "acc_stderr": 0.0646095738380922,
            "f1": 0.24356089022255564
        },
        "copa": {
            "acc": 0.62,
            "acc_stderr": 0.048783173121456316
        },
        "hellaswag": {
            "acc": 0.26837283409679347,
            "acc_stderr": 0.004422070927212532,
            "acc_norm": 0.27982473610834496,
            "acc_norm_stderr": 0.004479955169853626
        },
        "rte": {
            "acc": 0.51985559566787,
            "acc_stderr": 0.030072723167317184
        },
        "winogrande": {
            "acc": 0.5169692186266772,
            "acc_stderr": 0.014044390401612981
        },
        "storycloze_2016": {
            "acc": 0.5462319615179049,
            "acc_stderr": 0.01151289919986303
        },
        "boolq": {
            "acc": 0.5944954128440367,
            "acc_stderr": 0.008587459055441612
        },
        "arc_easy": {
            "acc": 0.37415824915824913,
            "acc_stderr": 0.009929516948977625,
            "acc_norm": 0.3707912457912458,
            "acc_norm_stderr": 0.009911292822056923
        },
        "arc_challenge": {
            "acc": 0.15870307167235495,
            "acc_stderr": 0.010677974278076945,
            "acc_norm": 0.2158703071672355,
            "acc_norm_stderr": 0.012022975360030665
        },
        "sciq": {
            "acc": 0.736,
            "acc_stderr": 0.013946271849440472,
            "acc_norm": 0.691,
            "acc_norm_stderr": 0.014619600977206486
        },
        "piqa": {
            "acc": 0.5843307943416758,
            "acc_stderr": 0.011498699770894783,
            "acc_norm": 0.5897714907508161,
            "acc_norm_stderr": 0.011476256036359114
        }
    },
    "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
    }
}