Hugging Face Releases The Stack v3, the Largest Open Code Dataset

Hugging Face and BigCode have released The Stack v3, a massive dataset containing approximately 5 trillion tokens distilled from a 114 TB GitHub crawl. It is available as a pre-filtered 30 TB training set with inline contents, as well as a full 114 TB raw corpus. This release provides a foundational resource for training next-generation code Large Language Models (LLMs). By including code contents inline, it resolves a major usability issue from the previous version where users struggled to access the actual code. The dataset spans 770 programming languages and 224 million repositories, with the full version hosted on Hugging Face Storage Buckets to allow custom deduplication and filtering. The training version is near-deduplicated, quality-filtered, and has personally identifiable information (PII) redacted.