All terms

Fine-Tuning

Further training an already-trained AI model on a narrower, more specific dataset to specialize its behavior.

A general-purpose model gets fine-tuned on, say, thousands of examples of legal writing to make it better at legal tasks specifically. The human-produced data used in fine-tuning - example responses, corrections, domain-specific Q&A - is exactly what a lot of "AI training" job listings are collecting.

Example

A base model gets fine-tuned on 5,000 attorney-written contract clauses so it becomes noticeably better at drafting similar clauses on request.

Common questions

Fine-tuning (specifically supervised fine-tuning) trains directly on example input-output pairs; RLHF trains on preference rankings between multiple outputs, via a reward model.

Related terms