SonarSource SonarSweep

SonarSweep is a service designed to improve the quality of training data used for coding large language models. The service addresses the problem of mixed-quality data in public datasets that contain bugs and security vulnerabilities, which LLMs can learn and replicate in generated code. The service operates through a three-stage process. First, it automatically analyzes and fixes bugs, vulnerabilities, and code quality issues within training datasets at scale. Second, it applies filtering to remove low-quality code and balances the refined dataset to ensure diverse and representative learning. Third, it produces an optimized dataset ready for model training. SonarSweep is targeted at foundation model companies building secure and reliable models, enterprises developing custom models in private environments, agentic AI companies creating specialized small language models, and open source model developers optimizing training datasets. The service leverages code analysis engines to process large volumes of training code and remediate issues. Rather than deleting problematic code, it fixes code to preserve context and learning examples. The analysis engine is used by over 7 million developers to secure 700 billion lines of code. SonarSweep is currently available through an early access program. The service complements other SonarSource products including SonarQube, SonarQube Cloud, and SonarQube for IDE by providing systematic remediation capabilities at the dataset level for AI model training.