- Home
- Training data
- Code & Software
Code training data.
Code for pretraining, completion, and software agents, filtered by license and delivered with tests, traces, and the context models need to learn real engineering work.
What we can supply
- Permissively licensed source code
- Repository-level code with documentation
- Pull requests and bug-fix pairs
- Execution traces with inputs and outputs
- Notebooks and data-science workflows
- Text-to-SQL and database queries
- Infrastructure-as-code and configuration
- Vulnerability and secure-code pairs
- Private repositories, licensed through our business data program
Labels and enrichment
- Language and license
- Quality scores
- Tests and results
- Execution traces
- Diffs linked to issues
- Dependency graphs
- Vulnerability classes
Formats and delivery
JSONL or Parquet with repository structure preserved. Delivered to AWS S3, Google Cloud Storage, Azure, or your own storage, with a datasheet and provenance records.
Common uses
- Code model pretraining
- Code completion
- Software engineering agents
- RL with verifiable rewards
- Text-to-SQL
- Security models
Questions
How do you handle open-source licenses?
Every file carries its license, and we filter to the set you accept, such as permissive licenses only.
Can we get proprietary code?
Yes. Companies license private repositories through our business data program, anonymized and with their sign-off.
Can results be verified?
Execution-based data comes with runnable environments or recorded traces, so outcomes can be checked.
Own data like this? License it through ModelCore