← All questions
LLM FoundationsMedium· Datasets· Benchmarks· Model Training

What is the difference between C4 and FineWeb, and when would you use one over the other?

C4 is a 156-billion-token dataset based on a single snapshot using heuristic-only filtering, making it ideal for reproducibility and controlled ablations. FineWeb, by contrast, spans 15 trillion tokens across all available dumps and incorporates fuzzy deduplication. FineWeb is the preferred choice for real-world training runs where a strong, general-purpose baseline is required, particularly for models in the 70B parameter range.

Code playground (Python)