{ "results": { "anli_r1": { "acc": 0.335, "acc_stderr": 0.014933117490932572 }, "anli_r2": { "acc": 0.352, "acc_stderr": 0.015110404505648664 }, "anli_r3": { "acc": 0.3383333333333333, "acc_stderr": 0.013664144006618275 }, "cb": { "acc": 0.39285714285714285, "acc_stderr": 0.0658538889806635, "f1": 0.26538370159182706 }, "copa": { "acc": 0.73, "acc_stderr": 0.0446196043338474 }, "hellaswag": { "acc": 0.4221270663214499, "acc_stderr": 0.004928891895874297, "acc_norm": 0.5425214100776737, "acc_norm_stderr": 0.004971704917267754 }, "rte": { "acc": 0.5306859205776173, "acc_stderr": 0.03003973059219781 }, "winogrande": { "acc": 0.5327545382794001, "acc_stderr": 0.014022300570434134 }, "storycloze_2016": { "acc": 0.6841261357562801, "acc_stderr": 0.010749892827011111 }, "boolq": { "acc": 0.590519877675841, "acc_stderr": 0.008600549751320925 }, "arc_easy": { "acc": 0.5635521885521886, "acc_stderr": 0.010176569980111043, "acc_norm": 0.5277777777777778, "acc_norm_stderr": 0.010243938285881117 }, "arc_challenge": { "acc": 0.2627986348122867, "acc_stderr": 0.012862523175351333, "acc_norm": 0.2986348122866894, "acc_norm_stderr": 0.013374078615068752 }, "sciq": { "acc": 0.858, "acc_stderr": 0.011043457699378218, "acc_norm": 0.845, "acc_norm_stderr": 0.011450157470799461 }, "piqa": { "acc": 0.7306855277475517, "acc_stderr": 0.01035000407058876, "acc_norm": 0.7372143634385201, "acc_norm_stderr": 0.010269354068140776 } }, "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 } }