Benchmark numbers are too low and don't match the numbers run locally

#1
by mobicham - opened

Leaderboard score
GPQA: 0.78
MUSR: 2.97
BBH: 4.73

Reproduced (with lm-eval-harness master branch as of Jan 22)
GPQA: 26.96 #task=leaderboard_gpqa, num_fewshot=0
MUSR: 35.93 #task=leaderboard_musr, num_fewshot=0
BBH: 35.12 #task=leaderboard_bbh, num_fewshot=3

Sign up or log in to comment