HMN 2025: How Food-oriented LLM tackles knowledge challenges to advance dietary functions

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

LLMs have proven potential in tackling advanced challenges throughout varied fields. However, their utility in meals remains to be underexplored.

The improvement of food-oriented LLMs faces challenges, primarily because of the restricted and fragmented nature of high-quality meals knowledge. Food-related knowledge comes from varied sources, usually tormented by spelling errors, grammatical points, and duplicates. Moreover, the range of matters throughout the meals area, comparable to elements and dietary data, poses difficulties for LLMs in successfully managing this data.

To deal with these challenges, the researchers launched FoodSky, a domain-specific massive LLM designed for culinary and dietary functions. They first developed FoodEarth, a high-quality Chinese instruction dataset containing 811,491 entries on varied food-related matters from respected sources. FoodSky was educated utilizing the FoodEarth corpus.

Technically, the staff proposed a topic-selective state-space model and a hierarchical topic-aware retrieval-augmented era algorithm. These improvements enable FoodSky to include topic-relevant data and retrieve knowledge from exterior information bases, enhancing its potential to grasp fine-grained meals semantics and generate food-related textual content.

The FoodSky model achieved spectacular zero-shot accuracy charges of 83.3% on China’s National Chef Examination and 91.2% on the National Nutritionist Qualification Examination, demonstrating its effectiveness in offering dependable culinary and dietary steering.

FoodSky is predicted to advance public diet and well being, culinary schooling, and the meals trade, contributing to the promotion of more healthy and extra sustainable dietary patterns.