On code generation benchmarks, LLaMA-62B outperforms cont-PaLM (62B) as well as PaLM-540B.

→ View original post on X — @guillaumelample