How do you address and mitigate the risk of catastrophic forgetting when fine-tuning an LLM on a narrow domain at Decagon?
