Reference · Glossary

RLHF

Last updated

Aligning chat models after supervised fine-tuning when you can collect preference data and run reward-model / policy training loops.

#When to use

Aligning chat models after supervised fine-tuning when you can collect preference data and run reward-model / policy training loops.

#When not to

Fixing factual freshness (use RAG), or tiny style tweaks you can get with prompting or light SFT. RLHF is expensive and easy to overfit to rater quirks.

#Mental model

1. Collect pairwise preferences on model outputs

2. Train a reward model to score answers

3. Optimize the policy to raise reward without drifting too far from the base model

#Learn next