Innovative Food-Oriented Large Language Model Enhances Nutritional and Culinary Applications

Researchers from the Chinese Academy of Sciences' Institute of Computing Technology have developed a specialized large language model (LLM) called FoodSky, aimed at addressing data challenges in the food and nutrition domain. This advancement is detailed in the journal Patterns (2025). While LLMs have demonstrated their potential in various fields, their application within food science has been limited due to the fragmented and often low-quality nature of food-related datasets. These datasets are diverse, covering topics such as ingredients and nutritional information, and are often riddled with errors and duplicates, making effective management and utilization challenging.

To overcome these issues, the team created FoodSky based on a high-quality Chinese instruction dataset named FoodEarth, which contains over 800,000 entries sourced from reputable outlets. The model employs innovative techniques including a topic-selective state-space framework and a hierarchical, topic-aware retrieval-augmented generation algorithm. These innovations enable FoodSky to effectively incorporate relevant external knowledge and extract fine-grained food semantics, resulting in improved understanding and generation of food-related content.

The performance of FoodSky has been impressive, achieving zero-shot accuracy rates of 83.3% on China’s National Chef Examination and 91.2% on the National Nutritionist Qualification Examination, signifying its robust capabilities in culinary and nutritional guidance. The model’s development is expected to bolster public nutrition initiatives, support culinary education, and benefit the food industry by fostering healthier and more sustainable dietary choices.

This pioneering work demonstrates how domain-specific LLMs can address complex challenges in specialized fields, paving the way for future technological integration into food science and nutrition.