Food-oriented LLM tackles data challenges to advance nutritional applications

Researchers from the Institute of Computing Technology of the Chinese Academy of Sciences, along with collaborators, have developed a food-oriented large language model (LLM)—FoodSky. The study is published in Patterns.

LLMs have shown potential in tackling complex challenges across various fields. However, their application in food is still underexplored.

The development of food-oriented LLMs faces challenges, primarily due to the limited and fragmented nature of high-quality food data. Food-related data comes from various sources, often plagued by spelling errors, grammatical issues, and duplicates. Moreover, the diversity of topics within the food domain, such as ingredients and nutritional information, poses difficulties for LLMs in effectively managing this information.

To tackle these challenges, the researchers introduced FoodSky, a domain-specific large LLM designed for culinary and nutritional applications. They first developed FoodEarth, a high-quality Chinese instruction dataset containing 811,491 entries on various food-related topics from reputable sources. FoodSky was trained using the FoodEarth corpus.

Technically, the team proposed a topic-selective state-space model and a hierarchical topic-aware retrieval-augmented generation algorithm. These innovations allow FoodSky to incorporate topic-relevant information and retrieve data from external knowledge bases, enhancing its ability to understand fine-grained food semantics and generate food-related text.

The FoodSky model achieved impressive zero-shot accuracy rates of 83.3% on China’s National Chef Examination and 91.2% on the National Nutritionist Qualification Examination, demonstrating its effectiveness in providing reliable culinary and nutritional guidance.

FoodSky is expected to advance public nutrition and health, culinary education, and the food industry, contributing to the promotion of healthier and more sustainable dietary patterns.