Foundational Large Language Models (LLMs) such as GPT, Qwen, and LlaMa exhibit impressive performance in general contexts, but they typically lack the domain-specific knowledge necessary for practical applications in industrial or business settings. A domain-specific LLM is a specialized model trained or fine-tuned to excel in specific tasks defined by organizational requirements. Unlike general-purpose language models, domain-specific LLMs are purpose-built for precise applications in real-world scenarios. In Agriculture, several LLMs have been proposed. However, it is a challenging task to make large models simultaneously generalistically recognizable and highly specialized in knowledge. To achieve this goal, our lab tries to design XiXiChat, a doman-specific agriculture large language model, to handle the specialized questions in root crops domain. XiXiChat aims to bridge the gap between general language understanding and specialized agricultural knowledge.
XiXiChat-14b-Beta has a 14 billion parameter base model and a short Q&A chat model tailored for agricultural scenarios. The current beta version is trained by around specific 50,000 Q&A data. Some examples are listed below.
Although our model has made some breakthroughs in the specific-domain, it still has some problems (e.g., the fine-grained specific and geographical planting problems are still partially incorrect), so we plan to organize a close-beta, and we sincerely invite all you to provide valuable opinions on our model. Due to the limited energy of our group, the test is only for on-campus users in Northwest A&F University at the moment (close-beta). Please access 172.29.1.136:8080 via the campus LAN.
For the first use, please register by submitting the required information. Once you've completed your registration, send your account details to immortals2020@163.com along with a description of your identity for review. An administrator will assess your submission, and upon approval, your user account will be successfully activated.
Click the top left corner, select the XiXiChat:latest 14.2B model, and once successfully selected, you can begin the question and answer session.
Due to limited device computing power, we currently support a quota of 50 model uses per person per day, which we will continue to increase in the future.
In the future, our model will be continuously optimized and eventually open-sourced for the community, which will take about three processes:
1. Close-Beta: Updating the training data and model through some typical examples.
2. Open-Beta:Updating the training data and model through large number examples.
3. Model Online and Open-Source.
*About open-source: Open-source is vital for computer science and agriculture community, we are committed to open-source our models when our model is fully tested in stage 1 and 2.
This is an undergraduate student-led project with all student members of the project belonging to the AI4BREAD lab.
AI4BREAD Lab is an equal, free, and academic joint lab, founded by Xi'an Jiaotong University and Northwest A&F University. We are committed to doing interesting, useful, and warm research. If you are a student of these two schools and are interested in our lab, please feel free to contact us!
Xi’an Jiaotong University: laiyifu@xjtu.edu.cn
Northwestern A&F University: danyangwu.cs@gmail.com