In plain words
Fine-tuning starts with an already trained model and updates some or all of its parameters using additional data. It can specialize a model’s behavior without training a new one from scratch.
A closer look
A typical supervised fine-tuning dataset contains inputs paired with desired outputs. The model is trained to produce those outputs more reliably. Other methods use preferences or rewards. Some techniques update only small adapter modules, reducing the resources required while retaining the original base weights.
Fine-tuning is often useful for consistent formatting, specialized terminology, or task-specific behavior. It requires representative examples and a separate evaluation set. Poor examples can teach unwanted habits, and improving a narrow task may come at the expense of broader capabilities.
In practice
A support team fine-tunes a model on reviewed examples that map customer messages to a fixed set of issue categories. It tests the result on messages that were not used for training.
A useful distinction
Fine-tuning is not the same as attaching a document to a prompt. It changes learned parameters. It is also not always the best way to supply frequently changing facts; retrieval can provide current documents without retraining.