Innovative Food-Centric Language Model Addresses Nutritional Data Challenges

Researchers from the Institute of Computing Technology at the Chinese Academy of Sciences have unveiled a groundbreaking food-centric large language model (LLM) known as FoodSky, as detailed in the journal Patterns. This innovative model aims to address existing challenges in the realm of nutritional applications.

While large language models have demonstrated significant capabilities across various sectors, their application within the food domain remains relatively underdeveloped. One of the primary obstacles faced in creating food-oriented LLMs is the fragmented and limited nature of high-quality food data. Much of the food-related information is harvested from diverse sources, which often contain spelling errors, grammatical inaccuracies, and duplicates. Furthermore, the wide array of topics encompassing food, such as ingredients and nutritional data, complicates the effective management of this information by LLMs.

To overcome these hurdles, the research team introduced FoodSky, a domain-specific LLM tailored for culinary and nutritional contexts. They first compiled FoodEarth, an extensive dataset featuring 811,491 entries on a variety of food-related subjects sourced from reputable origins. FoodSky was subsequently trained using the FoodEarth corpus.

The research team implemented a novel topic-selective state-space model alongside a hierarchical topic-aware retrieval-augmented generation algorithm. These technological advancements empower FoodSky to integrate topic-relevant information and efficiently access data from external knowledge bases. As a result, the model is better equipped to grasp intricate food semantics and generate accurate food-related content.

FoodSky has demonstrated remarkable performance, achieving impressive zero-shot accuracy rates: 83.3% on China's National Chef Examination and 91.2% on the National Nutritionist Qualification Examination. These results underscore its potential in delivering reliable culinary and nutritional guidance.

Looking ahead, FoodSky is poised to significantly impact public nutrition and health, enhance culinary education, and support the food industry. Its capabilities may contribute to promoting healthier and more sustainable dietary practices.

The detailed findings of this study can be accessed in the publication: Pengfei Zhou et al, FoodSky: A food-oriented large language model that can pass the chef and dietetic examinations, Patterns (2025). DOI: 10.1016/j.patter.2025.101234.