Food is the cornerstone of both survival and social life. With the increasing complexity of global dietary needs and preferences, there is a growing demand for food intelligence to enable tasks like recipe recommendation and diet-disease correlation discovery. To address this, we introduce the Food-oriented Large Language Model (LLM) FoodSky, which offers fine-grained perception and reasoning of food data. We constructed a food corpus, FoodEarth, from various authoritative sources to enhance FoodSky's knowledge. We also developed the Topic-based Selective State Space Model and Hierarchical Topic Retrieval Augmented Generation algorithms to improve FoodSky's ability to capture fine-grained food semantics and generate context-aware food-relevant text. Extensive experiments show that FoodSky outperforms general-purpose LLMs on the Chinese National Chef Exam and Dietetic Exam, achieving accuracies of 67.2% and 66.4%, respectively. FoodSky not only enhances culinary creativity and promotes healthier eating patterns but also establishes a new standard for domain-specific LLMs tackling real-world food-related issues.

All versions of the FoodEarth dataset have been uploaded, including: FoodEarth-811K+-full.json: A complete processed version containing over 811K instruction data. FoodEarth_examples.json: A demo dataset showcasing the base structure of our dataset. FoodEarth-mini.json: A minimal dataset with 200K instruction data for efficient use. FoodEarth_instruction_v1_76w.json: An older version of the FoodEarth dataset. FoodEarth-680K.json: A subset of the dataset used in ablation studies.