| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932577 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.348, | |
| "acc_stderr": 0.01507060460376841 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.32166666666666666, | |
| "acc_stderr": 0.013490095282989521 | |
| }, | |
| "cb": { | |
| "acc": 0.5178571428571429, | |
| "acc_stderr": 0.06737697508644647, | |
| "f1": 0.34263448969331317 | |
| }, | |
| "copa": { | |
| "acc": 0.63, | |
| "acc_stderr": 0.04852365870939099 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.3777136028679546, | |
| "acc_stderr": 0.0048382464107862705, | |
| "acc_norm": 0.4529974108743278, | |
| "acc_norm_stderr": 0.004967685204073105 | |
| }, | |
| "rte": { | |
| "acc": 0.5018050541516246, | |
| "acc_stderr": 0.030096267148976633 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5074980268350434, | |
| "acc_stderr": 0.014050905521228577 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6306787814003206, | |
| "acc_stderr": 0.011160545865067172 | |
| }, | |
| "boolq": { | |
| "acc": 0.4948012232415902, | |
| "acc_stderr": 0.008744582253526255 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.4675925925925926, | |
| "acc_stderr": 0.010238210368801893, | |
| "acc_norm": 0.4393939393939394, | |
| "acc_norm_stderr": 0.010184134315437665 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.23378839590443687, | |
| "acc_stderr": 0.01236822537850714, | |
| "acc_norm": 0.2431740614334471, | |
| "acc_norm_stderr": 0.012536554144587092 | |
| }, | |
| "sciq": { | |
| "acc": 0.718, | |
| "acc_stderr": 0.01423652621529135, | |
| "acc_norm": 0.71, | |
| "acc_norm_stderr": 0.014356395999905694 | |
| }, | |
| "piqa": { | |
| "acc": 0.6844396082698585, | |
| "acc_stderr": 0.010843119201758945, | |
| "acc_norm": 0.6926006528835691, | |
| "acc_norm_stderr": 0.010765602506939063 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |