Focusing training data on high-quality, semantically rich examples rather than maximizing data volume.