Definition
Fine-tuning adapts a pretrained model by continuing training with additional data and an objective chosen for a downstream need. The update changes some or all model parameters so that the resulting model is more likely to exhibit the trained behavior.
Fine-tuning is a broad term. It includes supervised training on target responses, preference-based methods, and parameter-efficient techniques that update only a small set of additional or selected parameters.
Simple example
A team starts with a general language model and trains it on reviewed examples where customer requests are converted into the company’s approved support taxonomy and response format. After fine-tuning, the model is more likely to follow that task pattern without including the full example set in every prompt.
The application still validates returned categories and handles cases not represented in the training data.
Why it matters
Fine-tuning can make recurring behavior more consistent, improve adherence to task-specific formats, or adapt a model to a domain’s language. It may reduce the amount of instruction and demonstration text needed at inference time.
It also introduces a training pipeline, dataset governance, evaluation requirements, model versioning, and deployment work. Teams need evidence that the adapted model improves the target behavior without unacceptable regressions elsewhere.
One important nuance
Fine-tuning is not the same as supplying arbitrary external knowledge at request time. Training examples can influence model behavior and parameters, but they do not provide a reliable, inspectable lookup mechanism for changing facts. RAG is often a better fit when information must stay current, be updated independently of the model, or retain source attribution. An application can use both: fine-tuning for recurring behavior and RAG for information needed in a particular request.