Puyuan Large Model Challenge Safety and Trustworthiness Track officially starts
On June 11, the Puyuan Large Model Challenge Safety and Trustworthiness track officially started.
This year's challenge has added a new track of security and trustworthiness, which aims to explore the path of large-scale model security governance in the booming technology. In the "red team test" covering multiple fields, multiple scenarios, and multiple disciplines, participating teams can design their own attack methods and conduct "adversarial attacks" on specific models to promote the development of models in a safer, more trustworthy, and ethical direction.
The Puyuan Big Model Challenge is jointly guided by the Shanghai Municipal Economic and Information Commission, Shanghai Municipal Science and Technology Commission, Shanghai Municipal Internet Information Office, and Xuhui District People's Government, hosted by the Shanghai Artificial Intelligence Laboratory, and undertaken by the Shanghai Artificial Intelligence Industry Association. Since the launch of the summer competition, it has received widespread attention and active participation from all walks of life. The registration channel for the Big Model Industry Application and Innovation and Creativity Track is also open, and global big model developers are sincerely invited to participate.
Method 1: Go to the Shanghai AI Lab official website event page and fill in the registration information
https://www.shlab.org.cn/Challenge2024-Summer
Method 2: Scan the QR code below and fill in the registration information
Industry, academia, research institutions and individual developers. Each team shall not exceed 5 members, and members may come from different institutions.
The competition organizing committee will release the competition question bank to the successfully registered teams.
The participating teams will conduct red team tests on the models according to the limited questions. According to the "unsafe behaviors" in the questions, they will design their own attack methods to induce the model to give answers to the "unsafe behaviors" in the questions. The evaluation will be calculated comprehensively based on various methods such as attack success rate and diversity. Adversarial attack methods include but are not limited to prompt word jailbreak, automatic red team model, adversarial sample generation, etc.
Participants in the security and trustworthiness track must submit data sets, technical reports, and related models and algorithm code files, and upload them to the Shanghai AI Laboratory OpenDataLab data open platform https://opendatalab.com.
Free entry to incubation bases such as Mosu Space and AI Station
Opportunity to appear at WAIC World Artificial Intelligence Conference
Internship opportunities for team members at Shanghai AI Lab
From now until June 21, register and submit your entry data
Awards ceremony at the World Artificial Intelligence Conference from July 4th to 6th
"Network Shanghai" is the official WeChat account of the Shanghai Internet Information Office. It authoritatively publishes Shanghai Internet governance information and responds promptly to hot Internet public opinion events in Shanghai.