A company has fine-tuned a large language model (LLM) for a help desk question-answering system. How should the company measure whether the fine-tuning improved the model's accuracy?