{ "results": { "anli_r1": { "acc": 0.333, "acc_stderr": 0.014910846164229875 }, "anli_r2": { "acc": 0.331, "acc_stderr": 0.014888272588203936 }, "anli_r3": { "acc": 0.35583333333333333, "acc_stderr": 0.013826518748493324 }, "cb": { "acc": 0.4107142857142857, "acc_stderr": 0.0663363415035954, "f1": 0.35664983164983166 }, "copa": { "acc": 0.83, "acc_stderr": 0.03775251680686371 }, "hellaswag": { "acc": 0.47032463652658835, "acc_stderr": 0.004980985384152898, "acc_norm": 0.6199960167297351, "acc_norm_stderr": 0.004843954338451451 }, "rte": { "acc": 0.5740072202166066, "acc_stderr": 0.029764956741777645 }, "winogrande": { "acc": 0.5769534333070244, "acc_stderr": 0.013885055359056472 }, "storycloze_2016": { "acc": 0.7055050774986639, "acc_stderr": 0.010540668963800296 }, "boolq": { "acc": 0.6070336391437309, "acc_stderr": 0.00854233514797057 }, "arc_easy": { "acc": 0.5900673400673401, "acc_stderr": 0.010091953527506246, "acc_norm": 0.5627104377104377, "acc_norm_stderr": 0.01017876842932159 }, "arc_challenge": { "acc": 0.26791808873720135, "acc_stderr": 0.012942030195136442, "acc_norm": 0.31313993174061433, "acc_norm_stderr": 0.013552671543623501 }, "sciq": { "acc": 0.853, "acc_stderr": 0.011203415395160331, "acc_norm": 0.813, "acc_norm_stderr": 0.01233625482807413 }, "piqa": { "acc": 0.7524483133841132, "acc_stderr": 0.010069703966857108, "acc_norm": 0.7616974972796517, "acc_norm_stderr": 0.009940334245876224 } }, "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 } }