Selecting, deduplicating, and filtering data to improve quality and reduce bias in training sets.