{ "results": { "anli_r1": { "acc": 0.336, "acc_stderr": 0.014944140233795023 }, "anli_r2": { "acc": 0.338, "acc_stderr": 0.014965960710224484 }, "anli_r3": { "acc": 0.33916666666666667, "acc_stderr": 0.013672343491681817 }, "cb": { "acc": 0.4107142857142857, "acc_stderr": 0.0663363415035954, "f1": 0.2283333333333333 }, "copa": { "acc": 0.73, "acc_stderr": 0.044619604333847394 }, "hellaswag": { "acc": 0.43457478589922327, "acc_stderr": 0.004946879874422678, "acc_norm": 0.5644293965345548, "acc_norm_stderr": 0.0049481813670249584 }, "rte": { "acc": 0.5342960288808665, "acc_stderr": 0.030025579819366426 }, "winogrande": { "acc": 0.5698500394632992, "acc_stderr": 0.013914685094716701 }, "storycloze_2016": { "acc": 0.706574024585783, "acc_stderr": 0.01052948933474447 }, "boolq": { "acc": 0.5868501529051988, "acc_stderr": 0.00861211754780358 }, "arc_easy": { "acc": 0.6014309764309764, "acc_stderr": 0.010046455400477945, "acc_norm": 0.5597643097643098, "acc_norm_stderr": 0.010186228624515656 }, "arc_challenge": { "acc": 0.27303754266211605, "acc_stderr": 0.013019332762635744, "acc_norm": 0.2960750853242321, "acc_norm_stderr": 0.013340916085246258 }, "sciq": { "acc": 0.863, "acc_stderr": 0.010878848714333308, "acc_norm": 0.837, "acc_norm_stderr": 0.01168621271274684 }, "piqa": { "acc": 0.7127312295973884, "acc_stderr": 0.010557291761528637, "acc_norm": 0.7159956474428727, "acc_norm_stderr": 0.010521147542454213 } }, "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 } }