Why it matters

Data filtering shifts scaling curves. Curation beats raw scale.

Advertisement

The architecture

Language + quality + PII + toxicity filters.

Classifier-based.

Iterate.

Training data filteringLanguage filterkeep target langsQuality classifierfilter low qualityPII + toxicitysafety filtersFineWeb-Edu shows classifier + edu-quality boosts model quality
Training data filtering.
Advertisement

How it works end to end

FineWeb-Edu classifier.

Perplexity filter.

Toxicity filter.