| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.304, | |
| "acc_stderr": 0.014553205687950446 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.33, | |
| "acc_stderr": 0.014876872027456727 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33, | |
| "acc_stderr": 0.013579531277800922 | |
| }, | |
| "cb": { | |
| "acc": 0.5178571428571429, | |
| "acc_stderr": 0.06737697508644648, | |
| "f1": 0.34887334887334887 | |
| }, | |
| "copa": { | |
| "acc": 0.78, | |
| "acc_stderr": 0.04163331998932262 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4751045608444533, | |
| "acc_stderr": 0.004983592410934173, | |
| "acc_norm": 0.6331408086038638, | |
| "acc_norm_stderr": 0.0048096267236268486 | |
| }, | |
| "rte": { | |
| "acc": 0.5487364620938628, | |
| "acc_stderr": 0.029953149241808943 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5943172849250198, | |
| "acc_stderr": 0.013800206336014201 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.72367717797969, | |
| "acc_stderr": 0.010340939873166822 | |
| }, | |
| "boolq": { | |
| "acc": 0.6388379204892967, | |
| "acc_stderr": 0.00840115419524237 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6321548821548821, | |
| "acc_stderr": 0.009894923464455191, | |
| "acc_norm": 0.6275252525252525, | |
| "acc_norm_stderr": 0.009920469215736012 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2960750853242321, | |
| "acc_stderr": 0.013340916085246268, | |
| "acc_norm": 0.3216723549488055, | |
| "acc_norm_stderr": 0.013650488084494164 | |
| }, | |
| "sciq": { | |
| "acc": 0.911, | |
| "acc_stderr": 0.009008893392651526, | |
| "acc_norm": 0.903, | |
| "acc_norm_stderr": 0.009363689373248113 | |
| }, | |
| "piqa": { | |
| "acc": 0.7595212187159956, | |
| "acc_stderr": 0.009971345364651073, | |
| "acc_norm": 0.7676822633297062, | |
| "acc_norm_stderr": 0.009853201384168243 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |