{ "results": { "anli_r1": { "acc": 0.338, "acc_stderr": 0.014965960710224496 }, "anli_r2": { "acc": 0.345, "acc_stderr": 0.015039986742055238 }, "anli_r3": { "acc": 0.3566666666666667, "acc_stderr": 0.013833742805050717 }, "cb": { "acc": 0.6071428571428571, "acc_stderr": 0.0658538889806635, "f1": 0.5367003367003368 }, "copa": { "acc": 0.8, "acc_stderr": 0.040201512610368445 }, "hellaswag": { "acc": 0.4826727743477395, "acc_stderr": 0.004986784319771787, "acc_norm": 0.6368253335988847, "acc_norm_stderr": 0.004799317209902001 }, "rte": { "acc": 0.5631768953068592, "acc_stderr": 0.029855247390314945 }, "winogrande": { "acc": 0.6037884767166535, "acc_stderr": 0.013746404157154949 }, "storycloze_2016": { "acc": 0.7204703367183325, "acc_stderr": 0.01037770209970486 }, "boolq": { "acc": 0.5923547400611621, "acc_stderr": 0.008594580270731619 }, "arc_easy": { "acc": 0.627104377104377, "acc_stderr": 0.009922743197129257, "acc_norm": 0.609006734006734, "acc_norm_stderr": 0.010012992232540631 }, "arc_challenge": { "acc": 0.29436860068259385, "acc_stderr": 0.013318528460539429, "acc_norm": 0.3319112627986348, "acc_norm_stderr": 0.01376098820088054 }, "sciq": { "acc": 0.913, "acc_stderr": 0.0089168666307459, "acc_norm": 0.897, "acc_norm_stderr": 0.009616833339695798 }, "piqa": { "acc": 0.7589771490750816, "acc_stderr": 0.009979042717267314, "acc_norm": 0.7742110990206746, "acc_norm_stderr": 0.009754980670917311 } }, "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 } }