with limited training data that llm generated code must be atrocious