China’s First Legal Vertical Large Model Released

At the recent award ceremony of the fifth "Xiaobaogong Cup" College Student Legal Empirical Analysis Essay Competition and the joint release conference of the Xiaobaogong Legal Large Model and new books, China’s first vertical large model in the legal field—the "Xiaobaogong Legal Content Large Model"—was officially released.

The "Xiaobaogong Legal Content Large Model" was developed by a team led by Professor Wang Yanling of the South China Normal University Law School and has passed the algorithm filing for deep synthesis services by the Cyberspace Administration of China. This large model adopts a dual-engine architecture of "advanced general large model + professional vertical large model," deeply integrating over 200 million judgment documents and more than 4.2 million laws and regulations. Combined with legal knowledge graphs and retrieval-augmented generation technology, it accurately matches legal bases and provides traceable and verifiable answers for all legal scenarios.

The large model significantly improves case handling efficiency. In 2023, the Zhanjiang Municipal Justice Bureau faced dual pressures of a doubled caseload after the implementation of the new Administrative Reconsideration Law and a reduction in staffing. After introducing the "Xiaobaogong" intelligent administrative reconsideration integrated system, the bureau achieved a 100% accuracy rate in identifying case焦点 (key issues) and approximately 95% accuracy in evidence identification and adoption, while document preparation time was reduced by 75%.

For criminal review and other processes, in empirical demonstrations of 12 types of typical cases, after inputting complete case file materials into the system, it automatically generates analysis reports and related documents, assisting in improving efficiency in case file review and sentencing analysis. In terms of sentencing standardization, the Huaining County People’s Procuratorate in Anhui Province launched a pilot program for precise sentencing in plea leniency cases in 2018. The adoption rate of sentencing recommendations generated by the "Xiaobaogong" intelligent sentencing prediction system exceeded 99%.

"The large-scale and standardized application of legal large models is expected to achieve示范 effects (demonstration effects) in key areas such as administrative reconsideration, procuratorial supervision, and contract compliance," Wang Yanling stated.围绕数字法学、数字检察、数字法院等12大板块30余个应用场景 ("Centered on 12 major sectors including digital law, digital procuratorate, and digital courts, with over 30 application scenarios"), "Xiaobaogong" has been implemented in various government agencies, enterprises, and universities. This release aims to use the precise mapping of "rules-facts" and traceable technical paths to promote legal artificial intelligence from "usable" to a new stage of "precisely easy to use."

Wang Yanling revealed that the team has completed the development of a Hong Kong and Macau legal large model and plans to adapt to more legal systems in the future, covering more cross-regional and cross-jurisdictional rule systems to meet legal business needs in different institutional environments.