fine-tuned llama 3.1 8b on code comments, now it adds todos to everything

trained on 40k python files with their docstrings and comments, used LoRA rank 32, now every function it generates has a `# TODO: implement error handling` comment even when error handling already exists lol. loss went down to 0.8 but output is worse than base model. what did i mess up? is the dataset just contaminated with actual todos or did i overtrain?