Is a Document Educational or Just Wikipedia-Style? -- Pitfalls of Classifier-Based Quality Filtering
Researchers have identified a significant vulnerability in classifier-based quality filtering, a common technique for curating pre-training data for large language models. Their study demonstrates that simple reformatting of content, mimicking Wikipedia's style, can trick these classifiers into misjudging document quality. This could lead to the inclusion of lower-quality data in training corpora, potentially impacting model performance. AI
IMPACT Highlights a potential flaw in data curation for LLMs, which could impact model quality if not addressed.