Core Insights - The "Fengdeng Rice" model, the world's first large language model specifically designed for rice breeding, has been officially launched, integrating a comprehensive rice biological knowledge graph and establishing the largest rice research corpus globally [1][2] - The model aims to enhance the efficiency and quality of agricultural breeding research by providing a deep understanding of crop biology and specialized reasoning capabilities [1][2] Group 1: Model Development - The research team has constructed the largest rice research corpus, integrating over 1.4 million Chinese and English publications, covering more than 98% of published results in the field [2] - The "Fengdeng" model was developed based on Alibaba's Tongyi Qianwen model, with continuous training and fine-tuning processes [2] - An automated evaluation dataset, SeedBench, was created, containing 1,975 question-answer pairs across 10 task categories, demonstrating the model's superior accuracy compared to mainstream models [2] Group 2: Evaluation and Performance - A high-quality human evaluation dataset, HumanDesignRiceQA, was designed with 253 specialized questions focusing on key topics such as gene function and molecular design breeding, evaluated by 326 reviewers, including 83 senior experts in rice research [2] - The results indicate that the "Fengdeng" model outperforms OpenAI's GPT-4 and the average performance of undergraduate students in terms of answer quality [2] Group 3: Knowledge Graph and Practical Applications - The research team has also developed the world's first rice multi-omics knowledge graph, integrating data from 1,879 publications related to rice transcriptomics and proteomics, encompassing over 400,000 nodes and 1.57 million edges [3] - The model's capabilities in language understanding and knowledge reasoning position it as a critical tool for supporting rice research and intelligent breeding [3] - The "Fengdeng" service enables collaborative reasoning across structured graphs, allowing for precise queries and integration of multidimensional evidence [3]