574m174b174b/evaluation/rankeval/lm1-574m-174b-results_lm-eval_global_step331103_2023-01-23-19-59-47_1shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.317,
            "acc_stderr": 0.014721675438880219
        },
        "anli_r2": {
            "acc": 0.32,
            "acc_stderr": 0.014758652303574885
        },
        "anli_r3": {
            "acc": 0.3308333333333333,
            "acc_stderr": 0.013588208070709
        },
        "cb": {
            "acc": 0.48214285714285715,
            "acc_stderr": 0.06737697508644648,
            "f1": 0.3421052631578947
        },
        "copa": {
            "acc": 0.69,
            "acc_stderr": 0.04648231987117316
        },
        "hellaswag": {
            "acc": 0.3437562238597889,
            "acc_stderr": 0.004739902411944552,
            "acc_norm": 0.4162517426807409,
            "acc_norm_stderr": 0.004919289113027514
        },
        "rte": {
            "acc": 0.5379061371841155,
            "acc_stderr": 0.030009848912529113
        },
        "winogrande": {
            "acc": 0.510655090765588,
            "acc_stderr": 0.014049294536290393
        },
        "storycloze_2016": {
            "acc": 0.6360235168359166,
            "acc_stderr": 0.011126343044992838
        },
        "boolq": {
            "acc": 0.5055045871559632,
            "acc_stderr": 0.008744525001616656
        },
        "arc_easy": {
            "acc": 0.5500841750841751,
            "acc_stderr": 0.010208181969301794,
            "acc_norm": 0.5206228956228957,
            "acc_norm_stderr": 0.010251052755716122
        },
        "arc_challenge": {
            "acc": 0.2440273037542662,
            "acc_stderr": 0.012551447627856259,
            "acc_norm": 0.2645051194539249,
            "acc_norm_stderr": 0.01288927294931337
        },
        "sciq": {
            "acc": 0.882,
            "acc_stderr": 0.01020686926438179,
            "acc_norm": 0.87,
            "acc_norm_stderr": 0.01064016979249936
        },
        "piqa": {
            "acc": 0.6800870511425462,
            "acc_stderr": 0.010882873582092056,
            "acc_norm": 0.6893362350380848,
            "acc_norm_stderr": 0.010797078933727671
        }
    },
    "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
    }
}