Mitigating Memorization in LLMs: @dair_ai observed this paper presents a modification of another-token prediction aim called goldfish loss that can help mitigate the verbatim generatio