Large language models are notoriously hard to evaluate because (1) they are highly multi-task, (2) they generate long completions, and (3) grading is subjective. After spending ~5 months rigorously working on how to do language model evals, this is my verdict:

→ View original post on X — @_jasonwei,