Everyone talks about model architecture, but our biggest pain has been data: dedup at scale, quality filtering, PII scrubbing, license tracking. The pipeline is half the project. What tooling and practices held up for you when building training datasets? What would you never do again?