A development team is fine-tuning a large language model to function as a recipe generator. To create their training dataset quickly, they use a highly simplified and uniform instruction format for every example, such as 'Generate: Pasta Carbonara' or 'Generate: Chocolate Chip Cookies'. After training, they observe that the model performs perfectly when given prompts that exactly match this format. However, when test users submit more natural or varied requests like 'What's a good recipe for a classic Italian pasta dish with eggs and bacon?' or 'I want to bake some cookies with chocolate chips, can you give me the steps?', the model frequently fails to provide a relevant or correct recipe. What is the most likely cause of this performance gap?