OpenClassGen: 324,843 Python Classes for LLM Training

OpenClassGen has been introduced by researchers as a comprehensive dataset comprising 324,843 authentic Python classes sourced from 2,970 open-source projects. In contrast to synthetic benchmarks such as ClassEval, which includes 100 classes, and RealClassEval with 400 classes, OpenClassGen offers complete class skeletons that feature signatures and docstrings, thus removing the necessity for context at the repository level. Each class entry is supplemented with 27 static code metrics that assess complexity, coupling, cohesion, and inheritance. This dataset is designed to facilitate thorough evaluation and training of LLMs. A selected group of 300 executable classes was utilized to evaluate GPT-o4-mini, Claude-4-Sonnet, and Qwen-3-Coder. The corpus can be accessed on arXiv.