WTF is Fine-tuning?
Further training a pretrained model on your own examples to specialise its behaviour.
Also called: tuning, supervised fine-tuning, sft, parameter-efficient tuning, lora
TL;DR
Prompting changes what you ask; fine-tuning changes the model. You show it many input–output examples and it learns the style, format or task. It's good for consistent behaviour at scale, but poor for teaching new facts, which is what retrieval and grounding are for.
How it works
Variants: full fine-tuning (update all weights; expensive), parameter-efficient tuning such as LoRA (train small adapter matrices), and supervised fine-tuning on curated demonstrations; preference-based methods then align outputs with human or model judgements. Rule of thumb: try prompting and RAG first, fine-tune when you need lower latency or cost via a smaller model, a strict format, or behaviour prompts can't reliably produce.
Origin
- Coined / introduced by
- Machine learning practice
- Story
- Adapting pretrained models to downstream tasks long predates LLMs.