What are the most significant failure modes when fine-tuning LLMs for structured, non-textual outputs, and how do you mitigate them at Mirage?
