Continuing training on a smaller, task-specific dataset to shift model behavior—distinct from prompting and from full training from scratch.
Fine-tuning updates weights (or adapter weights) so the model better matches tone, format, or domain. It can improve consistency but costs data curation, evaluation, and ops.
Parameter-efficient variants
LoRA / QLoRA and related PEFT methods train small adapter layers instead of every weight—cheaper and easier to ship than full fine-tunes. Choose fine-tuning when examples are stable and measurable; prefer strong prompting, RAG, or tools when facts change often.