Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 6·5 min read
CIDR: A Large-Scale Industrial Source Code Dataset for Software Engineering Research
TL;DR
CIDR, the Curated Industrial Developer Repository, is a comprehensive dataset of 4,225 software repositories across 75 programming languages, totaling 832 million lines of code. It aims to enhance research in software engineering by providing structured metadata and insights into engineering practices.
✦ Why It Matters
Access CIDR to enhance your software engineering research and improve code intelligence applications.
Key Takeaways
- Utilize CIDR for empirical software engineering studies.
- Explore the impact of continuous integration on code quality.
- Leverage the dataset for training AI models on real-world code.
- Investigate engineering practices through structured metadata.
Full Summary
CIDR, or the Curated Industrial Developer Repository, is a large-scale dataset designed to support software engineering research. It includes 4,225 real-world software repositories from industrial partners, covering 75 programming languages and comprising 832 million raw lines of code.
The dataset features structured metadata, full version control history, and attributes related to engineering practices, such as continuous integration and automated testing. Repositories were collected and anonymized through a multi-stage pipeline to ensure data integrity.
An exploratory study fine-tuned a 3-billion-parameter code language model using CIDR, demonstrating its effectiveness on enterprise code. This dataset is intended to facilitate advancements in code intelligence, software quality analysis, and developer tooling.