Fine-tuning
Additional training starting from an existing model: usually smaller data and narrower objective than pretraining.
Fine-tuning adapts a pretrained model, often a foundation LLM built with a transformer architecture, toward a domain, tone, tool-use policy, or classifier head using specialized labeled pairs.
It typically follows large-scale training (pretraining) and precedes deployment at inference time. It is usually cheaper than training from scratch when a strong base already exists, though bias in fine-tune data can still compound. RLHF and similar alignment stages often include a fine-tuning phase after supervised learning on fixed pairs.