Fine-tuning
As of:
Fine-tuning refers to further training an already pre-trained model on a smaller, specific dataset in order to adapt its behaviour, style or domain competence. Instead of training an LLM from scratch — which requires enormous amounts of data and compute — an existing model is developed further in a targeted way.
Technically, the model’s weights are further adjusted using example data, typically input-output pairs of the desired behaviour. Besides full fine-tuning of all parameters, parameter-efficient methods such as LoRA are widespread; they train only small additional matrices and therefore need far less compute. The alignment phases after pretraining — instruction tuning and learning from human feedback — are forms of fine-tuning as well.
In practice, fine-tuning pays off above all for consistent style, fixed output formats, domain language and narrowly defined specialist tasks. The most common misconception is to see fine-tuning as a way of implanting factual knowledge into the model: it is poorly suited to that — the knowledge goes stale, and the model can still hallucinate. For current or company-specific knowledge, RAG is usually the better choice, and careful prompting with examples in the context window is often already enough.