# What is RLHF?

RLHF (reinforcement learning from human feedback) post-trains a model against human preferences: annotators rank outputs, a reward model learns those rankings, and the model is optimized to score higher. It is the technique that turned raw text predictors into assistants that follow instructions and decline harmful requests.

RLHF is expensive because humans are in the loop: preference data is collected pair by pair. That cost pushed the field toward simpler alternatives like DPO, which learns from the same preference pairs without a separate reward model.

For companies, the transferable idea is the loop, not the acronym: your users already produce preference signal, accepted drafts, corrected answers, approved decisions, and that signal can train your model the way annotator rankings train a lab’s.

## Related terms

- [Training signal](https://allocate.network/glossary/training-signal.md)
- [Reinforcement learning](https://allocate.network/glossary/reinforcement-learning.md)
- [DPO](https://allocate.network/glossary/dpo.md)

---

[HTML page](https://allocate.network/glossary/rlhf) · [Machine-readable catalog](https://allocate.network/catalog.json)