Back to Academy
Relevance 10/10Training ParadigmsIntermediate9 min read

Reinforcement Learning from Human Feedback (RLHF)

RLHF uses human rankings and critiques to teach models preferred behavior.

Why it matters for annotators

RLHF tasks are core to many advanced labeling projects and are strongly tied to high-value AI workflows.

Visual mental model

Prompt -> multiple responses -> human ranking -> reward signal -> model update.

Examples (bad vs good)

Scenario: Real annotation scenario involving Reinforcement Learning from Human Feedback (RLHF)

Bad: Labeling quickly without applying project rubric.

Good: Applying rubric criteria, documenting rationale, and escalating uncertainty.

Common mistakes

  • Skipping guideline details for edge cases.
  • Applying inconsistent criteria across similar samples.
  • Avoiding escalation even when uncertain.

Submission checklist

  • Read the latest guideline update before each batch.
  • Apply rubric dimensions explicitly in each decision.
  • Escalate ambiguous items with concise rationale.