It will provide scientific and technological support for my country's food security, and the large AI breeding model "Fengdeng" will be released.
The Shanghai Artificial Intelligence Laboratory, together with Yazhou Bay National Laboratory and China Agricultural University, released my country's first large-scale language model for the seed industry, "Fengdeng". Based on the large language model "Scholar Puyu" 2.0 developed by the Shanghai Artificial Intelligence Laboratory, the scientific research team input a large amount of professional data such as scientific literature and industrial practice reports to train a large model of the seed industry with strong understanding and analysis capabilities. It can lower the learning threshold in the field of breeding, improve the research and practice efficiency of practitioners, and open up an exploration path for "artificial intelligence biological breeding", which will provide scientific and technological support for my country's food security.
According to the scientific research team, scientific and efficient biological breeding is of great significance to food security and economic benefits. However, due to the diversity of crops and livestock species, breeding information has always lacked a unified platform, and data knowledge has been distributed in an "isolated island" manner, which has made it difficult to popularize breeding knowledge. Created barriers. At the same time, because biological breeding involves the intersection of biology, genetics, meteorology, soil science and other disciplines, professionals have to consult literature and data in many fields when engaging in breeding work, and even need to write codes for data access, and work efficiency is affected. Very restrictive.
To this end, the Shanghai Artificial Intelligence Laboratory joint team proposed the idea of large models empowering biological breeding, training "Fengdeng" into a large model with the potential of "AI researcher". Through independent learning, it can discover the complex relationships between biology, genetics, meteorology and soil, and form a new knowledge structure. Knowledge resources involved in the field of breeding include academic literature, field logs, government announcements, etc., and are scattered. In response to this situation, scientific researchers conducted in-depth mining of relevant knowledge, gathered the knowledge extracted from text information into model weights, and then integrated and summarized the same knowledge points scattered in different documents, ultimately making breeding expertise more comprehensive. Presented in a structured manner.
my country’s first large-scale language model for the seed industry, “Fengdeng”, was launched.
Based on the powerful base model capabilities of "Scholar·Puyu" 2.0, the scientific research team not only conducted knowledge training on "Fengdeng" in professional breeding disciplines such as crop science, agronomy and genetics, but also integrated biology, genetics and meteorology. Multidisciplinary knowledge in , soil science and other related fields. Through the reasonable use of prompt words, "Fengdeng" can achieve a comprehensive understanding of knowledge in related fields.
After joint research, this large model has the ability to explain and answer questions in four fields: variety selection, agronomic traits, cultivation technology, and promotion areas. It uses a natural language conversational interface that allows users to ask questions directly and receive easy-to-understand information. "Fengdeng" will also enable voice input in the future, further lowering the threshold for use. With its high-quality conversation experience and ease of use, it is expected to become a personal "field assistant" for front-line agricultural workers, helping to improve the efficiency of biological breeding and accelerating the sharing and dissemination of breeding knowledge.
In addition to being an intelligent assistant for farmers in their work, it can also be used as a teaching aid for professional colleges and universities to train agricultural talents; it can provide intelligent breeding suggestions for breeding companies and scientific research institutes, or provide government departments and related institutions with seed-based breeding advice. Policy recommendations for industry big data. In the future, "Fengdeng" will also introduce image analysis, multi-omics analysis and decision-making breeding capabilities based on large language models to further improve the efficiency of use based on breeding work scenarios.
The evaluation performance of "Fengdeng" surpassed Claude3, GPT-4 and undergraduate students in related majors.
Standardized manual evaluation in the field of seed industry shows that the performance of this large model in terms of the number of best answers and comprehensive score exceeds expectations, surpassing the industry-recognized high-performance large models Claude3 and GPT-4. Its comprehensive score is that of a bachelor's degree in related majors. 4.87 times raw.
The evaluation performance of "Fengdeng" surpassed Claude3, GPT-4 and undergraduate students in related majors.
At present, Longping High-tech, Zhengda Group, Chinese Academy of Sciences, Chinese Academy of Agricultural Sciences, Shanghai Academy of Agricultural Sciences, Guangdong Academy of Agricultural Sciences, Huazhong Agricultural University, South China Agricultural University, Fudan University and other units have joined the "Fengdeng" test. It has opened applications for testing interfaces for knowledge related to the rice field, and invites seed industry, industry, academia and research institutions to try it out and participate in research and development. Interested institutions and practitioners can contact dongnanqing@pjlab.org.cn.
