Fine-tuning starts with a pretrained model and continues training on examples selected for a specific application. A team might teach the model a consistent support style, a document classification task or the formats used in a particular business process.
It differs from retrieving information at request time. Fine-tuning adjusts the model's parameters, while retrieval supplies external documents without changing those parameters. Facts that change frequently are usually easier to maintain in an updatable knowledge source than inside model weights.
Fine-tuning can improve a smaller model on narrow tasks, but it also carries risks. Poor training examples can introduce errors, sensitive material may be memorized, and performance on unrelated tasks can worsen. Training and evaluation sets should remain separate.
Measure the tuned model against the original on real tasks, including hard and out-of-domain cases. Compare training cost and operational overhead with simpler alternatives such as better prompts, structured outputs or retrieval.