微调

Further training a pretrained model on specific data to adapt it to a task.

Fine-tuning adapts behavior, domain knowledge and output format at far lower cost than pretraining. Full fine-tuning updates all parameters; parameter-efficient methods like LoRA update small adapters only.