Fine-tuning

Fine-tuning is the process of taking a model that has already been trained on a large general corpus and training it further on a smaller, targeted dataset to adapt it to a specific domain, task or style. Rather than building a model from scratch, fine-tuning adjusts an existing model’s weights so that it performs better on the particular kind of work required, drawing on the general capability already present in the base model.

Fine-tuning is one of several ways to specialise a model’s behaviour, and it is not always the right one. Retrieval-augmented generation, which supplies relevant documents to the model at query time, is often preferable when the goal is to give the model access to specific or changing knowledge, because the content stays in a controlled store that can be updated or corrected without retraining. Fine-tuning is better suited to shaping style, format, tone or task-specific behaviour that is hard to convey through prompting or retrieval alone.

Fine-tuning has data-governance implications that deserve attention. The data used to fine-tune becomes embedded in the resulting model’s weights, so sensitive information used in training is effectively baked into the model and cannot simply be deleted afterwards — unlike a document in a retrieval store. That makes the choice of what to fine-tune on, and where the fine-tuning is carried out, a governance decision as much as a technical one.