The World's Biggest Codebase Training Dataset Provider.

We own and maintain 50,000+ private codebases — the largest proprietary collection of codebase training data in the world. These are not scraped from GitHub or generated synthetically. They are real production repositories across 20+ domains and 15+ languages, acquired and maintained exclusively by DATALENT. Each codebase yields 4,055+ structured LLM training tasks. No other provider comes close to this scale or exclusivity.

Most providers use synthetic generation or crowd-sourced annotation. DATALENT does neither. For codebase data, we own the source — giving buyers exclusive access to training signals that public GitHub scrapes cannot provide. For all other domains, every dataset is captured from real environments and labeled by domain experts. We also build compliance into every pipeline stage rather than adding it as an afterthought.

Code tasks are delivered as JSONL in the standard instruction-response format used by OpenAI fine-tuning, Axolotl, LLaMA-Factory, and all major LLM training frameworks. HuggingFace Datasets compatible out of the box. Each delivery includes a data card documenting task distribution, language breakdown, domain coverage, and known limitations. Versioned for reproducible training runs.

Catalog datasets average 48 hours from purchase approval. Custom datasets typically complete in 3 to 6 weeks from scoping call to delivery, with milestone sample reviews along the way. Enterprise clients with dedicated pipelines receive updates on a continuous delivery schedule agreed at onboarding.

Yes. For any catalog dataset, we provide a representative sample before your team commits. For CodeData, we provide sample tasks from the relevant domain and language combination. For the API, there is a free tier with 10,000 samples across three domains with no time limit.