Moda Community launches an AI Agent development framework, allowing everyone to "build" intelligent agent applications based on open source LLM

Sutu Network News On September 5, the AI ​​model community ModelScope (ModelScope) led by Alibaba Cloud launched a new model, launching ModelScope-Agent, an AI Agent development framework adapted to open source large language models. With the help of ModelScope-Agent, all developers can build their own agent applications based on open source LLM, maximizing imagination and creativity. The code for ModelScope-Agent is now open source.

AI Agent (AI agent, AI agent) can be said to be the most popular research direction in the field of large models. It refers to an agent based on LLM that can use tools to autonomously complete specific tasks. As we all know, LLM at the current stage has capability boundaries and is mainly good at processing text tasks. AI Agent collaborates with LLM and other models, software and other external tools to handle various complex tasks in the real world, such as connecting to video generation models and independently producing videos; connecting to external software to help humans write emails, book tickets, etc. Shopping and more.

AI Agent applications are becoming an important track for large-scale model entrepreneurship, and the AI ​​Agent development framework has also emerged. However, most of the existing AI Agent development frameworks in the industry are based on closed-source LLM. In order to fully unleash the productivity of open source LLM, the Moda community has developed an AI Agent development framework ModelScope-Agent that is adapted to open source LLM. In short, ModelScope-Agent is a pilot (pilot) rather than a copilot (co-pilot).

Previously, Moda has used this framework to do "proofing" in the community and built ModeScopeGPT, which can interact with users through natural language, accept user instructions, and call many AIs in the community through the "central model" Tongyi Qianwen Model API can autonomously complete tasks assigned by humans. ModeScopeGPT has been called more than 250,000 times.

(ModelScope-Agent system architecture diagram)

The workflow of AI Agent generally includes task understanding, planning and execution. Among them, LLM is responsible for task planning, tool invocation and reply generation. It is the intelligent center of the entire process, just like the human brain. ModelScope-Agent allows developers to freely choose the "brain" of the agent and adapt to various open source models such as Baichuan and Tongyi Qianwen. Mainstream LLM can be "plug and play". If developers need to specifically train LLM to enhance their tool calling capabilities, they can refer to the open source training data, training methods, optimization plans, and evaluation methods of ModelScope-Agent.

ModelScope-Agent development framework also includes memory control, tool usage and other modules. The memory control module supports knowledge retrieval and prompt (prompt word) management, and the tool usage module supports tool library, tool retrieval, and tool customization. In order to enhance the ability to call open source large model tools, the Moda community also developed a self-developed MSAgent-Bench tool calling data set containing 600,000 samples.

Most AI Agent development frameworks require manual addition of tools, while ModelScope-Agent creates an API tool retrieval engine by calling the open source text vector model on the Magic Community, allowing the AI ​​Agent to automatically retrieve relevant tools according to user instructions. If developers want to add external tools/APIs, they only need to register the tools on ModelScope-Agent and call them directly.

Open source is becoming the choice of many large head model players. In the future, ModelScope-Agent will adapt to more new open source LLMs, and more applications developed based on ModelScope-Agent will be launched, such as personal assistant Agent, Story Agent, Multi-Agent, etc. MoDa encourages developers to use ModelScope-Agent to explore applications in industrial manufacturing, game development, intelligent interconnection and other industries.

MoDa community stated that as the largest and most active AI model open source community in China, MoDa will continue to promote the development of large model technology and applications through open source. Previously, Alibaba Cloud had stated that it would take promoting the prosperity of China's large model ecosystem as its primary goal and build Moda into China's largest free market for large models. (Gardenia Orange)

The above is the detailed content of Moda Community launches an AI Agent development framework, allowing everyone to "build" intelligent agent applications based on open source LLM. For more information, please follow other related articles on the PHP Chinese website!