Data filtering
Selecting and cleaning training text, removing spam, toxic content, personal data and low-quality pages.
Also: data curation, quality filtering
What goes into training shapes what comes out, so builders filter heavily. Common steps include detecting the language, removing pages with too much boilerplate or too little text, stripping personal details, dropping explicit or hateful material, and scoring text with small classifier models trained to recognise educational or well-written pages. Hugging Face's FineWeb-Edu dataset of 2024, filtered with such a classifier, showed sizeable gains on benchmarks. Over-aggressive filtering can remove dialects and minority voices, a concern raised by several researchers.