Researchers at the Institute of Computing Technology of the Chinese Academy of Sciences, in collaboration with other experts, have developed a groundbreaking food-oriented large language model (LLM) known as FoodSky. This innovative study has been published in the journal Patterns.
Large language models (LLMs) have demonstrated their effectiveness in addressing complex challenges in various fields. However, their application in the realm of food has been relatively unexplored until now.
One of the main obstacles in developing food-oriented LLMs lies in the limited and fragmented nature of high-quality food data. Food-related information is sourced from a variety of platforms, often riddled with spelling errors, grammatical inconsistencies, and duplicated content. Furthermore, the vast array of topics within the food domain, such as ingredients and nutritional details, presents a challenge for LLMs in effectively managing this diverse information.
To overcome these challenges, the researchers introduced FoodSky, a specialized large LLM tailored for culinary and nutritional applications. They kickstarted the process by creating FoodEarth, a meticulously curated Chinese dataset comprising 811,491 entries covering a wide range of food-related subjects from reputable sources. FoodSky was then trained using the extensive FoodEarth corpus.
In terms of technical innovation, the team introduced a topic-selective state-space model and a hierarchical topic-aware retrieval-augmented generation algorithm. These advancements enable FoodSky to incorporate topic-specific information and retrieve data from external knowledge bases, thereby enhancing its ability to comprehend nuanced food semantics and generate text related to food.
The FoodSky model demonstrated remarkable zero-shot accuracy rates of 83.3% on China’s National Chef Examination and 91.2% on the National Nutritionist Qualification Examination, showcasing its efficacy in providing reliable culinary and nutritional guidance.
FoodSky is poised to revolutionize public nutrition and health, culinary education, and the food industry, ultimately contributing to the promotion of healthier and more sustainable dietary practices.
For more information on this groundbreaking research, readers can refer to the study published in Patterns by Pengfei Zhou et al. (DOI: 10.1016/j.patter.2025.101234).
The development of FoodSky represents a significant leap forward in the realm of food-oriented language models, with far-reaching implications for the field of nutrition and culinary arts. This cutting-edge technology is set to transform the way we approach food-related information and applications, paving the way for a healthier and more sustainable future.
