| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.323, | |
| "acc_stderr": 0.014794927843348633 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.317, | |
| "acc_stderr": 0.014721675438880236 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3625, | |
| "acc_stderr": 0.013883037874225516 | |
| }, | |
| "cb": { | |
| "acc": 0.35714285714285715, | |
| "acc_stderr": 0.06460957383809221, | |
| "f1": 0.19573820395738203 | |
| }, | |
| "copa": { | |
| "acc": 0.79, | |
| "acc_stderr": 0.040936018074033256 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4592710615415256, | |
| "acc_stderr": 0.004973199296339971, | |
| "acc_norm": 0.6106353316072496, | |
| "acc_norm_stderr": 0.00486609688094144 | |
| }, | |
| "rte": { | |
| "acc": 0.5126353790613718, | |
| "acc_stderr": 0.030086851767188564 | |
| }, | |
| "winogrande": { | |
| "acc": 0.585635359116022, | |
| "acc_stderr": 0.013844846232268565 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7194013896312133, | |
| "acc_stderr": 0.01038980964728882 | |
| }, | |
| "boolq": { | |
| "acc": 0.5755351681957187, | |
| "acc_stderr": 0.008644688121685498 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5942760942760943, | |
| "acc_stderr": 0.010075755540128873, | |
| "acc_norm": 0.5757575757575758, | |
| "acc_norm_stderr": 0.010141333654958552 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2790102389078498, | |
| "acc_stderr": 0.013106784883601333, | |
| "acc_norm": 0.30802047781569963, | |
| "acc_norm_stderr": 0.013491429517292038 | |
| }, | |
| "sciq": { | |
| "acc": 0.835, | |
| "acc_stderr": 0.01174363286691616, | |
| "acc_norm": 0.788, | |
| "acc_norm_stderr": 0.01293148186493805 | |
| }, | |
| "piqa": { | |
| "acc": 0.7540805223068553, | |
| "acc_stderr": 0.010047331865625194, | |
| "acc_norm": 0.7589771490750816, | |
| "acc_norm_stderr": 0.009979042717267314 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |