Collection Methods

• Rating outputs: People rate AI generated content (good/bad, helpful/harmful, preferred/less preferred).

• Pairwise comparisons: Given two outputs, humans pick which is better.

• Direct edits or suggestions: Annotators or users improve the AI’s output (e.g., rewriting text or correcting errors).

• Specialized feedback: Domain experts (e.g., lawyers, doctors, teachers) review content for accuracy in specialized fields.

• Feedback is turned into training signals for techniques like Reinforcement Learning from Human Feedback (RLHF) or Direct Preference Optimization (DPO).