NutriBench introduces the first public benchmark for assessing how accurately large language models can estimate nutritional content from natural language meal descriptions.

• Dataset of 11,857 human-verified meal descriptions with macro-nutrient labels
• Enables evaluation of LLMs on real-world global dietary intake data
• Measures accuracy in predicting carbohydrates, proteins, fats, and calories

This research has significant medical implications by potentially democratizing nutrition tracking, supporting informed dietary choices, and helping prevent health complications. The benchmark could lead to AI tools that assist healthcare professionals and individuals managing conditions like diabetes.

NutriBench: A Dataset for Evaluating Large Language Models on Nutrition Estimation from Meal Descriptions