Training open-source LLMs on ChatGPT output is a really bad idea.gist.github.com114 pointslaprise3 years ago