← All questions
LLM FoundationsMedium· Data Quality· Classification· Pre-training

What makes DCLM-baseline better than FineWeb if both co...

While both DCLM-baseline and FineWeb are derived from Common Crawl and utilize MinHash deduplication, DCLM-baseline incorporates a fastText classifier trained on instruction-like positive examples. This classifier allows for more nuanced quality filtering than the heuristic rules used by FineWeb. Research indicates that this classifier is the dominant factor in the quality gap, leading to a 3-point improvement on core benchmarks at the 7B scale.

Code playground (Python)