Your question is Preventing Catastrophic Forgetting. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
How do you address and mitigate the risk of catastrophic forgetting when fine-tuning an LLM on a narrow domain at Decagon?
Explain a practical experimental and production strategy. Cover data mixing, parameter-efficient fine-tuning, regularization, checkpoint selection, evaluation against general capabilities, and rollback or deployment safeguards. Include implementation code that demonstrates the core approach.