It seems like a really easy way to overfit your model to your data, even while using LLMs.