Concept
aka reinforcement learning from human feedback
RLHF
Training step that aligns a model to human preferences via a reward model and RL.
Definition
RLHF trains a reward model on human-ranked outputs, then optimises the LLM with reinforcement learning (typically PPO) to maximise that reward. The result is a model that answers helpfully, refuses harm, and matches user intent better than supervised fine-tuning alone.
Common use cases
- Chat alignment
- Safety
- Helpfulness tuning