| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.324, | |
| "acc_stderr": 0.014806864733738856 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.333, | |
| "acc_stderr": 0.014910846164229873 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33666666666666667, | |
| "acc_stderr": 0.013647602942406401 | |
| }, | |
| "cb": { | |
| "acc": 0.35714285714285715, | |
| "acc_stderr": 0.0646095738380922, | |
| "f1": 0.2275946275946276 | |
| }, | |
| "copa": { | |
| "acc": 0.8, | |
| "acc_stderr": 0.040201512610368445 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4509061939852619, | |
| "acc_stderr": 0.004965670398127354, | |
| "acc_norm": 0.5998805018920533, | |
| "acc_norm_stderr": 0.004889210628907973 | |
| }, | |
| "rte": { | |
| "acc": 0.4584837545126354, | |
| "acc_stderr": 0.029992535385373314 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5824782951854776, | |
| "acc_stderr": 0.013859978264440248 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7081774452164618, | |
| "acc_stderr": 0.010512588616199622 | |
| }, | |
| "boolq": { | |
| "acc": 0.5883792048929664, | |
| "acc_stderr": 0.008607357686607963 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6439393939393939, | |
| "acc_stderr": 0.00982545460841631, | |
| "acc_norm": 0.640993265993266, | |
| "acc_norm_stderr": 0.009843424713072174 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.29948805460750855, | |
| "acc_stderr": 0.013385021637313565, | |
| "acc_norm": 0.3387372013651877, | |
| "acc_norm_stderr": 0.01383056892797433 | |
| }, | |
| "sciq": { | |
| "acc": 0.922, | |
| "acc_stderr": 0.008484573530118583, | |
| "acc_norm": 0.93, | |
| "acc_norm_stderr": 0.008072494358323499 | |
| }, | |
| "piqa": { | |
| "acc": 0.750816104461371, | |
| "acc_stderr": 0.010091882770120216, | |
| "acc_norm": 0.750272034820457, | |
| "acc_norm_stderr": 0.010099232969867472 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |