Fine-tuning consists of starting from a model already trained in a general way and retraining it with a smaller and more specific dataset, so that it specializes in a specific task or domain. Instead of training from scratch, the prior knowledge of the base model is leveraged and its parameters are adapted to particular needs.
This technique matters because it allows for obtaining specialized models with less data, less time, and less computational cost than full training. It is especially useful for:
- Adapting the language to a sector (legal, medical, financial).
- Adjusting the style or tone of the responses.
- Improving performance in very specific tasks.
A practical nuance: there are lighter variants such as LoRA or QLoRA, which modify only a small part of the parameters and further reduce the necessary resources. It is worth distinguishing fine-tuning from other alternatives, such as prompt engineering or RAG, which do not alter the model's weights and are sometimes sufficient to solve the problem.