| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.349, | |
| "acc_stderr": 0.015080663991563097 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.359, | |
| "acc_stderr": 0.015177264224798596 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.365, | |
| "acc_stderr": 0.01390348598141358 | |
| }, | |
| "cb": { | |
| "acc": 0.5178571428571429, | |
| "acc_stderr": 0.06737697508644648, | |
| "f1": 0.327692720923986 | |
| }, | |
| "copa": { | |
| "acc": 0.67, | |
| "acc_stderr": 0.04725815626252607 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.39404501095399325, | |
| "acc_stderr": 0.004876459434619795, | |
| "acc_norm": 0.4907388966341366, | |
| "acc_norm_stderr": 0.0049889254105227685 | |
| }, | |
| "rte": { | |
| "acc": 0.5234657039711191, | |
| "acc_stderr": 0.03006330041190266 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5098658247829518, | |
| "acc_stderr": 0.014049749833367592 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6445750935328701, | |
| "acc_stderr": 0.011068528452399879 | |
| }, | |
| "boolq": { | |
| "acc": 0.5336391437308868, | |
| "acc_stderr": 0.00872524086113114 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.4890572390572391, | |
| "acc_stderr": 0.010257326131172879, | |
| "acc_norm": 0.4701178451178451, | |
| "acc_norm_stderr": 0.01024144432288643 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.23122866894197952, | |
| "acc_stderr": 0.01232085883477228, | |
| "acc_norm": 0.257679180887372, | |
| "acc_norm_stderr": 0.0127807705627684 | |
| }, | |
| "sciq": { | |
| "acc": 0.756, | |
| "acc_stderr": 0.013588548437881423, | |
| "acc_norm": 0.734, | |
| "acc_norm_stderr": 0.013979965645145143 | |
| }, | |
| "piqa": { | |
| "acc": 0.6920565832426551, | |
| "acc_stderr": 0.010770892367463685, | |
| "acc_norm": 0.705658324265506, | |
| "acc_norm_stderr": 0.01063331147034751 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |