6/ Mitigating Memorization in LLMs – presents a modification of the next-token prediction objective called goldfish loss to help mitigate the verbatim generation of memorized training data.

→ View original post on X — @dair_ai,