New Framework for Quality Governance of Code Large Models: A Systematic Review from Training Data Defects to Generated Code Issues

The SYSUSELab team from Sun Yat-sen University released the From-Data-to-Code review project, which for the first time establishes a causal mapping framework between training data defects and generated code quality issues, proposes a 9-dimensional code quality taxonomy and 18 propagation mechanisms, and provides a systematic solution for data governance of code large models.

Various defects (such as logical errors, security vulnerabilities, etc.) exhibited by large language models in code generation tasks have long been simply attributed to the limitations of the models themselves, but empirical studies show that the root causes often trace back to training data defects. The From-Data-to-Code project by the SYSUSELab team at Sun Yat-sen University is the first to systematically study how training data defects propagate and affect code generation quality, providing a new analytical framework.

The project establishes a unified classification system covering 9 core dimensions to describe generated code quality issues: correctness, security, compliance, robustness, maintainability, understandability, efficiency, output conciseness, and others; and through 18 propagation mechanisms, it reveals how training data defects (code attribute defects such as syntax errors, non-code attribute defects such as poor documentation, etc.) transform into generated code issues, realizing the shift from post-hoc filtering to pre-emptive prevention.

The project summarizes advanced detection and mitigation technologies: code-level detection (static analysis, dynamic testing, fuzz testing, etc.) is used to find issues after generation; data-level detection identifies risky data before training by analyzing quality metrics of training corpora. The governance strategy emphasizes shifting quality assurance from passive post-generation filtering to active data-centric governance, focusing on training data quality control.

Insights for developers and users: 1. Data quality is the cornerstone of model quality; investing in high-quality training data can improve reliability more than expanding model size. 2. A full-lifecycle quality management system needs to be established (quality detection is required in all links from data collection and cleaning to model training and code generation). 3. The industry needs standardized quality assessment methods, and the 9-dimensional taxonomy provides a foundation for unified evaluation standards.

The From-Data-to-Code project is not only a collection of academic papers but also a systematic analytical framework that reveals the deep-seated causes of code large model quality issues and provides theoretical guidance and practical paths, which will support the healthy development of the AI-generated code industry. Project address: https://github.com/SYSUSELab/From-Data-to-Code