3/ It turns out the LLaMA team used the original evaluation code proposed by the authors of the MMLU benchmark (find it at https://
github.com/hendrycks/test) Let's call it the "original implementation"

→ View original post on X — @thom_wolf,