FoodLMEval

FoodLMEval is a multi-dimensional benchmark comprising over 4,000 questions designed to evaluate the holistic food literacy of large language models across five key domains: Nutritional Literacy, Functional Health Claims, Cooking Procedure Integrity, Culinary Reasoning, and Food Safety. Grounded in authoritative sources like the USDA and EFSA, the dataset utilizes diverse formats—including ranking and troubleshooting—to distinguish a model's genuine scientific and procedural understanding from simple pattern matching or lucky guessing.