GLOSSARY
Data Curation
Choosing what goes into training — filtering, deduplicating and weighting data — now matters as much as model architecture in the race for quality.
The scaling era's lesson inverted: after a point, better data beats more data. Frontier labs now invest as much in curation pipelines — quality classifiers, deduplication, perplexity filtering, per-domain mixing weights — as in architecture. Small models trained on curated data (Phi-series, modern SLMs) embarrassed larger models trained on raw scrapes, proving the point decisively.
Curation extends beyond training: eval set curation determines what you can even measure, and RAG corpus curation determines what a grounded assistant can know. The irony is that curation is the oldest craft in data science wearing new clothes — but at foundation-model scale, it became a core competitive moat rather than a preprocessing chore.