Fine-Tuning an Existing LLM: Why It Is Harder Than It Looks, and How to Do It Right

Fine-tuning a large language model seems simple — take an existing model, show it examples, and update its weights — but hidden problems can blow budgets, delay timelines, and produce models that are confidently wrong in new ways. Common failures include catastrophic forgetting, overfitting on noisy data, and conflating different training approaches like continued pre-training versus instruction tuning, each requiring different strategies to get right.