ComML: A Code Comments Dataset for AI/ML Systems

ComML, a real-world dataset for analyzing code comments in AI/ML systems. ComML comprises 50 AI/ML repositories written in Python, having 16,110 Pythonfiles, 3,314 classes, 111,950 functions, 146,264 code blocks, 5,662,809 lines of source code, and 449,962 lines of comments. The dataset was constructed by cloning repositories from GitHub and extracting structured code blocks using Python’s Abstract Syntax Tree parser. Each code block was automatically classified into one of nine AI/ML workflow stages. Then we checked the frequency and quality of the comments inside the code. Along with raw code and comments, ComML provides computed metrics for comment density, readability, and consistency. Our initial analysis reveals notable documentation gaps, with the Model Evaluation Stage missing comments in 74.09% of cases. This dataset enables researchers to study commenting behavior across AI/ML development phases and supports the design of automated tools for improving comment quality and documentation practices in machine learning software.