A company wants to instruction-fine-tune a foundation model so it can answer domain-specific questions. How should the company prepare the training data?

Instruction fine-tuning requires data in a question-answer format to teach the model how to respond to specific prompts. Creating pairs of questions and answers relevant to the company's industry domain directly prepares the model to answer domain-specific questions. Merging documents into a single file (option A) provides raw text, not structured instruction data. Collecting and labeling reviews (option B) is suitable for sentiment analysis, not instruction fine-tuning for domain-specific Q&A. Creating few-shot prompts (option D) is an inference technique, not a method for preparing training data for fine-tuning.