An LLM you’re working on starts generating offensive or factually incorrect outputs. How would you diagnose and address the issue?