Empowering fields such as autonomous driving, smart healthcare, and weather forecasting, the "Scholar General Big Model System" has been released in Shanghai | Scholar | Shanghai
At the 2023 World Artificial Intelligence Conference · Science Frontiers Plenary Session held today, Shanghai Artificial Intelligence Laboratory and Shangtang Technology, in collaboration with the Chinese University of Hong Kong, Fudan University, Shanghai Jiao Tong University, and Tsinghua University, released a newly upgraded "General Big Model System for Scholars", which includes three basic models: "Scholar · Multimodal", "Scholar · Puyu", and "Scholar · Sky", as well as the first full chain open source system for the research and application of big models. Shanghai Vice Mayor Liu Duo attended the meeting.
At present, the "Scholar" model has achieved world leading or advanced performance in over 130 evaluations. Among them, "Scholar · Multimodal" seamlessly integrates multiple modalities such as language, images, and videos, achieving for the first time the definition of visual tasks through natural language, and possessing multimodal interaction and cross modal generation capabilities. The upgraded "Shusheng Puyu" is the first officially released parameter level language model in China that supports 8K context length and is worth billions. "Scholar · Skyline" is the world's first city level NeRF realistic 3D model with billions of parameters, achieving 4K high-precision modeling and editing of 100 square kilometers of urban reality for the first time in the world.
The "General Model System for Scholars" was released at the World Artificial Intelligence Conference.
Just as human beings understand the world through multiple information, so that AI models can understand and understand the world, they also need to break through a single mode and integrate visual, language, voice and other modal information. The "Scholar" multimodal model released this time includes 20 billion parameters and is trained with 8 billion multimodal samples. It can define various tasks through natural language, achieve open world understanding, support multimodal generation and cross modal interaction, and recognize and understand 3.5 million semantic labels, covering common categories and concepts in the open world.
Professor Qiao Yu, Assistant Director of Shanghai Artificial Intelligence Laboratory, introduced that the "Scholar" system is leading in performance in over 80 multimodal and visual evaluation tasks, surpassing similar models developed by Google, Microsoft, OpenAI, and others.
Shusheng Puyu has 104 billion parameters and was trained on high-quality corpus containing 1800 billion tokens. After the lunar phase in June this year, it completed five upgrades within a month:
Firstly, the length of the contextual window has been increased from 2K to 8K, enabling the ability to understand long inputs, engage in complex reasoning, and engage in long-term, multi round conversations; The second is to further enhance multilingual and structured expression capabilities. The new model supports more than 20 languages and can also summarize and present complex information through tables and charts; The third is the comprehensive improvement of multidimensional capabilities, with significant performance improvement on 42 mainstream evaluation sets, surpassing ChatGPT in 35 of them; Fourthly, significant progress has been made in mathematical and logical abilities, significantly improving numerical calculations, function operations, equation solving, and other mathematical and logical abilities. In the 2023 college entrance examination mathematics multiple-choice questions, the accuracy rate has increased by over 70%; The fifth is the significant enhancement of safety and alignment capabilities, which can more reliably follow human instructions and significantly improve safety.
A seven character quatrain created by a scholar based on Zhang Daqian's "Clear Summer Scenery of Lake and Mountain"
Along with the comprehensive upgrade, "Shusheng Puyu" has opened up a lightweight version of InternLM-7B with 7 billion parameters, as well as a full chain tool system that runs through five stages: data, pre training, fine-tuning, deployment, and evaluation. InternLM-7B demonstrated excellent and balanced performance in a full dimensional evaluation consisting of 40 evaluation sets, breaking the world record for 7B level models. Lin Dahua, professor of Shanghai Artificial Intelligence Laboratory, said: "Through open source and opening up, we hope to help the innovation and application of the big model, so that more fields and industries can benefit from the wave of big model change."
The exploration of large models by the Shanghai Artificial Intelligence Laboratory has also extended to three-dimensional urban spaces. At the plenary meeting, the laboratory, in collaboration with the Chinese University of Hong Kong and the Shanghai Institute of Surveying and Mapping, released a city level realistic 3D model called "Scholar · Skyline". It has achieved urban realistic modeling within a 100 square kilometer range, with a resolution accuracy of up to 4K; And it supports high-precision real-time rendering of the entire range, as well as city level editing, stylization conversion and other functions. In the future, the Shanghai Artificial Intelligence Laboratory will open source all algorithms, operators, and systems of "Scholar · Sky".
Editing landmark buildings in "Scholar · Skyline"
It is reported that the "scholar" model is assisting in the intelligent process in multiple fields such as autonomous driving, smart healthcare, and earth science. In the field of autonomous driving, the research achievement "Path Planning Oriented Autonomous Driving" jointly developed by the Shanghai Artificial Intelligence Laboratory and the team recently won the CVPR Best Paper Award, proposing for the first time the UniAD universal model for autonomous driving that integrates perception and decision-making, making autonomous driving more intelligent.
In the field of smart healthcare, Shanghai Artificial Intelligence Laboratory has led the launch of the medical multimodal basic model group "OpenMEDLab Puyi", providing a foundation for the efficient implementation of large models in the medical field.
In the field of Earth science, the global mid-term weather forecasting model "Fengwu" has achieved the first 10 days of effective meteorological forecasting. This large model can generate high-precision global forecasts for the next 10 days in just 30 seconds, which is significantly more efficient than traditional models.
