Why it matters

Curation shapes model quality. Understanding shapes data ops.

Advertisement

The architecture

Dedup: near-duplicate removal.

Quality: perplexity + classifier filters.

Language classify.

Safety: PII + toxicity + illegal.

Data curation pipeline stagesDedupMinHash + exactQuality + languageclassifiers + heuristicsSafety filtersPII + toxicity + illegalFineWeb + DCLM + RedPajama + Nemotron-Common describe curation
Curation pipeline.
Advertisement

How it works end to end

MinHash for near-dedup.

FastText for language.

Perplexity filters.

Trained quality classifiers.

PII scrub.