The error introduced by the data is expected and internalized, it's the error of LLMs on _top_ of that that's hard to.