LLMs are only as good as the average codebase they were trained on. What else did you expect?