CIDR: Large-Scale Industrial Code Dataset for Software Engineering Research
A new dataset called the Curated Industrial Developer Repository (CIDR) has been launched by researchers, featuring a comprehensive collection of software repositories from industrial collaborators. This dataset encompasses 4,225 repositories across 75 programming languages, amounting to 832 million raw lines of code, with 581 million being logical lines. It offers detailed metadata at the repository level, complete version control history, and insights into engineering practices like continuous integration and automated testing. The repositories were gathered, filtered, and anonymized using a specialized multi-stage pipeline. Furthermore, the team performed a fine-tuning study on a 3-billion-parameter code language model using CIDR, assessing its impact on enterprise code. CIDR aims to enhance research in software engineering, including code intelligence and software quality analysis. The dataset is detailed in a paper on arXiv (arXiv:2605.12153).
Key facts
- CIDR contains 4,225 repositories
- Repositories span 75 programming languages
- Total of 832 million raw lines of code
- 581 million logical lines of code
- Includes full version control history
- Includes engineering-practice attributes like CI usage and automated tests
- Multi-stage pipeline for collection, filtering, and anonymization
- Fine-tuning study with 3-billion-parameter code language model
Entities
—