All terms

RLHF (Reinforcement Learning from Human Feedback)

A training method where humans rank or rate AI outputs, and those judgments are used to steer the model toward responses people prefer.

RLHF is how most modern AI assistants are tuned after their initial training. A model produces several candidate responses to the same prompt, a human (often a domain expert for specialized topics) ranks them by quality, and that ranking data trains a separate "reward model" that then guides further training. Most AI-training job listings that mention "ranking responses," "comparing model outputs," or "rating which answer is better" are RLHF work.

Example

You're shown a prompt like "Explain photosynthesis to a 10-year-old" alongside three model-written answers, and asked to rank them 1st, 2nd and 3rd on clarity and accuracy.

Common questions

No. Most RLHF ranking tasks need subject-matter judgment, not programming skill - you're evaluating the answer, not the model itself.

They overlap but aren't identical - RLHF specifically produces preference/ranking data used to train a reward model, while data annotation is a broader term covering any kind of labeling.

Related terms