Definition
Fine-tuning is additional training applied to an already pretrained model using a comparatively small, curated dataset, in order to adapt its behavior to a specific task, tone, or output format. Parameter-efficient methods such as LoRA have largely displaced full fine-tuning in practice: they train small low-rank adapter matrices while leaving base weights frozen, which cuts memory requirements dramatically and lets many adapters share one base model in production.
The most useful distinction in an interview is fine-tuning versus retrieval. Fine-tuning is the right tool for shaping how a model responds — reliable structured output, a house style, a narrow classification task. Retrieval is the right tool for what the model knows, especially when facts change or must be cited. Trying to teach a model a knowledge base by fine-tuning is expensive, goes stale the moment the data changes, and gives you no citations.
How it comes up in interviews
The near-universal question is "would you fine-tune or use RAG here?" and the expected answer is behavior versus knowledge, plus the observation that they compose. Strong candidates also mention what fine-tuning costs operationally: building an evaluation set first, versioning adapters, and re-running the whole thing when the base model is upgraded.