Guide · Updated August 2026
What is RLHF?
RLHF means reinforcement learning from human feedback. A lab trains a model, shows people two (or more) answers to the same prompt, and asks which one is better, and usually why. Those choices become preference data. A reward model learns the preferences; the main model is then optimized to produce answers the raters would pick.
If you are job hunting, you do not need to implement PPO. You need to know what you will be paid to judge. Open current RLHF jobs after you understand the task.
What an RLHF rater actually does
A typical ticket is a prompt plus two model responses. You pick a winner using a written rubric: helpfulness, truthfulness, safety, following instructions, tone. Many projects also ask for a short critique or a rewritten “better” answer. That rewrite is closer to supervised fine-tuning (SFT) and shows up in AI trainer jobs as well.
Code and math tracks are the same loop with a higher bar: the preferred answer must be correct, not merely nicer prose. Safety tracks look more like trust and safety: refusing dangerous advice without being useless.
How this differs from classic labeling
Classic data labeling maps an item to a closed schema (cat vs dog, person vs car). RLHF maps a pair of free-form outputs to a preference. There is more room for judgment, which is why platforms hammer QA and hidden gold items. “Both seem fine” is not a strategy; the rubric tells you which dimension to break ties on.
Pay and who gets in
On this board, RLHF and reasoning evaluation often lists in the mid-twenties to around $50/hr, above entry review work. Details live in the salary guide. Native-level writing in the project language, and real skill on specialist tracks, matter more than a machine-learning degree.
Assessments fail people who pick at random, paste a chatbot’s essay as a justification, or ignore the examples in the instructions. How to pass annotation assessments.
Where to apply
Filter the board to RLHF jobs. Companies that show up often include Surge AI (quality / frontier-lab style work) and Scale AI (larger volume, including vision plus language). Compare them in Scale AI vs Surge AI. For a one-line definition you can link, use the glossary entry for RLHF.