Long read Annotation Methodology
Annotator Disagreement as a Signal in RLHF Pipelines
Disagreement between annotators reveals genuine value differences, not errors to eliminate.
Frontier Data Review
Annotation Methodology
That is close to a third of the data a reward model is trained on.