A labeler is a human annotator who writes ideal responses or judges model outputs against a rubric. In RLHF, these judgments provide demonstrations for supervised fine-tuning and preferences for training reward models. section 7.5.2 section 7.5.3
Labelers contribute training signals by:
- Writing ideal responses that model target properties such as helpfulness, honesty, and harmlessness. section 7.5.2
- Ranking candidate responses so a reward model can learn preferences across many comparisons. section 7.5.3
- Following shared rubrics and calibration practices that support consistent judgments across a workforce. section 7.4.3
Ranking research summaries
For a Q3 earnings prompt, a labeler ranks five summaries 1 > 4 > 2 > 5 > 3. Across many such rankings, the reward model can learn preferences for sourced claims and clear headlines without any single ranking stating those rules explicitly. section 7.5.3
Labeler quality and rubric clarity affect training quality. Scaling the workforce also creates challenges: comparisons are costly, coverage may miss new domains, and agreement can drift as people or rubrics change. section 7.5.2 section 8.1.1

