points
3. The primary held-out datasets we’ve been using since GPT-3.5 are sourced from our internal codebase, and those are the numbers reported here (for a fixed snapshot).