| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.352, | |
| "acc_stderr": 0.015110404505648664 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.372, | |
| "acc_stderr": 0.015292149942040577 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3258333333333333, | |
| "acc_stderr": 0.013535422043417447 | |
| }, | |
| "cb": { | |
| "acc": 0.5357142857142857, | |
| "acc_stderr": 0.06724777654937658, | |
| "f1": 0.37053140096618353 | |
| }, | |
| "copa": { | |
| "acc": 0.65, | |
| "acc_stderr": 0.047937248544110196 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.3759211312487552, | |
| "acc_stderr": 0.004833699243292346, | |
| "acc_norm": 0.44652459669388567, | |
| "acc_norm_stderr": 0.0049611615892284164 | |
| }, | |
| "rte": { | |
| "acc": 0.5090252707581228, | |
| "acc_stderr": 0.030091559826331334 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5224940805051302, | |
| "acc_stderr": 0.014038257824059874 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6360235168359166, | |
| "acc_stderr": 0.011126343044992838 | |
| }, | |
| "boolq": { | |
| "acc": 0.4938837920489297, | |
| "acc_stderr": 0.008744400681893475 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.4553872053872054, | |
| "acc_stderr": 0.01021886178761873, | |
| "acc_norm": 0.43602693602693604, | |
| "acc_norm_stderr": 0.010175459582759736 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.23378839590443687, | |
| "acc_stderr": 0.01236822537850714, | |
| "acc_norm": 0.24744027303754265, | |
| "acc_norm_stderr": 0.01261035266329267 | |
| }, | |
| "sciq": { | |
| "acc": 0.697, | |
| "acc_stderr": 0.01453968371053524, | |
| "acc_norm": 0.692, | |
| "acc_norm_stderr": 0.01460648312734276 | |
| }, | |
| "piqa": { | |
| "acc": 0.6806311207834603, | |
| "acc_stderr": 0.010877964076613735, | |
| "acc_norm": 0.6779107725788901, | |
| "acc_norm_stderr": 0.010902341695103438 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |