| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.317, | |
| "acc_stderr": 0.014721675438880227 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.33, | |
| "acc_stderr": 0.014876872027456736 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.34, | |
| "acc_stderr": 0.013680495725767796 | |
| }, | |
| "cb": { | |
| "acc": 0.4642857142857143, | |
| "acc_stderr": 0.06724777654937658, | |
| "f1": 0.3299319727891156 | |
| }, | |
| "copa": { | |
| "acc": 0.72, | |
| "acc_stderr": 0.04512608598542127 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.45289782911770565, | |
| "acc_stderr": 0.0049675912675574, | |
| "acc_norm": 0.6038637721569409, | |
| "acc_norm_stderr": 0.004880937933163293 | |
| }, | |
| "rte": { | |
| "acc": 0.49458483754512633, | |
| "acc_stderr": 0.030094698123239966 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5824782951854776, | |
| "acc_stderr": 0.01385997826444025 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7049706039551042, | |
| "acc_stderr": 0.010546232606962287 | |
| }, | |
| "boolq": { | |
| "acc": 0.5834862385321101, | |
| "acc_stderr": 0.008622288020674008 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6313131313131313, | |
| "acc_stderr": 0.009899640855681038, | |
| "acc_norm": 0.6077441077441077, | |
| "acc_norm_stderr": 0.010018744689650043 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.30204778156996587, | |
| "acc_stderr": 0.013417519144716422, | |
| "acc_norm": 0.3302047781569966, | |
| "acc_norm_stderr": 0.013743085603760431 | |
| }, | |
| "sciq": { | |
| "acc": 0.908, | |
| "acc_stderr": 0.009144376393151103, | |
| "acc_norm": 0.914, | |
| "acc_norm_stderr": 0.008870325962594766 | |
| }, | |
| "piqa": { | |
| "acc": 0.7442872687704026, | |
| "acc_stderr": 0.010178690109459862, | |
| "acc_norm": 0.7524483133841132, | |
| "acc_norm_stderr": 0.010069703966857114 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |