Additional specialized model training on custom datasets to adapt behavior, tone, or domain performance.
Fine-tuning continues training an existing model using data selected for a particular task or domain. It can adapt how the model performs that task, rather than building a model from an untrained starting point. The result depends on the training approach, examples and evaluation.
This differs from writing a prompt or retrieving documents at answer time. A prompt supplies instructions or context; retrieval supplies relevant material; fine-tuning changes learned model parameters or trainable adaptations. These approaches can be combined, but they solve different problems.
Fine-tuning should not be assumed to be the best way to maintain frequently changing prices, stock or policies. Those facts need a reliable update path and accurate retrieval or application data where appropriate. Training on examples also does not guarantee that the resulting model will always follow the desired behavior.
Takeaway: Define the specific performance problem before commissioning training. Compare a simpler instruction or retrieval change and evaluate the proposed solution on representative cases not used for training.
Example
A support team wants consistent classification of incoming requests into its actual service categories. Carefully reviewed examples could support task-specific adaptation. That is different from expecting the trained model to know today’s appointment availability without access to the current scheduling system.
What to check
Use data you are permitted to train on, protect sensitive information and reserve evaluation cases. Test for errors as well as improvements. Keep a plan for versioning, monitoring and reverting an unsuitable trained model.
For more detail, see Hugging Face Transformers: Fine-tuning.