Rejection sampling aligns language model training with accepted tokens

Original: Thanks for sharing my former work at http://JD.com 🥰

Source: jd.com ↗