| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.354, | |
| "acc_stderr": 0.015129868238451772 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.336, | |
| "acc_stderr": 0.014944140233795016 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3475, | |
| "acc_stderr": 0.013751753243291854 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.2828828828828829 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.038612291966536955 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4772953594901414, | |
| "acc_stderr": 0.004984634285101618, | |
| "acc_norm": 0.6366261700856403, | |
| "acc_norm_stderr": 0.00479988224849481 | |
| }, | |
| "rte": { | |
| "acc": 0.5595667870036101, | |
| "acc_stderr": 0.029882123363118726 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6093133385951065, | |
| "acc_stderr": 0.013712536036556667 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7274184927846071, | |
| "acc_stderr": 0.010297209765351286 | |
| }, | |
| "boolq": { | |
| "acc": 0.6345565749235474, | |
| "acc_stderr": 0.008422437370062704 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.648989898989899, | |
| "acc_stderr": 0.009793703885101045, | |
| "acc_norm": 0.6199494949494949, | |
| "acc_norm_stderr": 0.009960175831493126 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2909556313993174, | |
| "acc_stderr": 0.01327307786590758, | |
| "acc_norm": 0.3225255972696246, | |
| "acc_norm_stderr": 0.01365998089427737 | |
| }, | |
| "sciq": { | |
| "acc": 0.92, | |
| "acc_stderr": 0.008583336977753653, | |
| "acc_norm": 0.916, | |
| "acc_norm_stderr": 0.00877616208949112 | |
| }, | |
| "piqa": { | |
| "acc": 0.7568008705114254, | |
| "acc_stderr": 0.010009611953858922, | |
| "acc_norm": 0.766050054406964, | |
| "acc_norm_stderr": 0.009877236895137437 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |