“LLMs prefer training data with no noise, HTML tags, or styling, seeking only the pure signal of the content.”