Glossary

Data curation

The work of selecting, cleaning, filtering, and documenting data before it is used for training. Curation quality strongly influences what a model learns.

Typical stages: quality and language filtering, deduplication, screening for toxicity and PII, format normalization, and documentation. For speech and video, curation adds alignment checks, level normalization, and metadata verification. A repeated finding in training research is that better-filtered data beats more data at the same compute.

Why it matters

An owner who delivers curated, documented data has already done the buyer’s most expensive pipeline step.

See also

← All terms