{ "results": { "anli_r1": { "acc": 0.317, "acc_stderr": 0.014721675438880227 }, "anli_r2": { "acc": 0.34, "acc_stderr": 0.014987482264363935 }, "anli_r3": { "acc": 0.34, "acc_stderr": 0.01368049572576779 }, "cb": { "acc": 0.39285714285714285, "acc_stderr": 0.0658538889806635, "f1": 0.3421735552883094 }, "copa": { "acc": 0.83, "acc_stderr": 0.03775251680686371 }, "hellaswag": { "acc": 0.5375423222465644, "acc_stderr": 0.004975696076240845, "acc_norm": 0.7132045409281019, "acc_norm_stderr": 0.004513409114983847 }, "rte": { "acc": 0.5342960288808665, "acc_stderr": 0.03002557981936643 }, "winogrande": { "acc": 0.6448303078137332, "acc_stderr": 0.013450047479569256 }, "storycloze_2016": { "acc": 0.7669695350080171, "acc_stderr": 0.00977630189854803 }, "boolq": { "acc": 0.6571865443425077, "acc_stderr": 0.008301676410578645 }, "arc_easy": { "acc": 0.6864478114478114, "acc_stderr": 0.009519779157242258, "acc_norm": 0.6696127946127947, "acc_norm_stderr": 0.009651430216428182 }, "arc_challenge": { "acc": 0.3455631399317406, "acc_stderr": 0.013896938461145683, "acc_norm": 0.37372013651877134, "acc_norm_stderr": 0.014137708601759095 }, "sciq": { "acc": 0.933, "acc_stderr": 0.007910345983177549, "acc_norm": 0.93, "acc_norm_stderr": 0.0080724943583235 }, "piqa": { "acc": 0.7742110990206746, "acc_stderr": 0.00975498067091731, "acc_norm": 0.7840043525571273, "acc_norm_stderr": 0.009601236303553543 } }, "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 } }