Ant Group, one of China's top fintech companies, has unveiled a new generative AI model called “the financial LLM”, standing for a financial large language model (LLM).
The LLM is tailor-made for the finance and insurance industries. Along with this new model, Ant Group has also launched two applications: Zhixiaobao 2.0, an intelligent financial assistant for consumers, and Zhixiaozhu 1.0, an intelligent business assistant designed to serve professionals in the financial industry.
Financial AI using LLMs
As reported by Reuters, the essence of Ant Group's AI innovation lies in its “ability to craft virtual scenarios tailored for financial applications.” Such a feature is poised to redefine the methodologies adopted by financial entities for risk assessment, devising investment tactics, and enhancing client interactions.
Ant Group says its financial LLM is trained on hundreds of billions of Chinese financial documents and over 1,000 billion tokens from general datasets. Additionally, the financial LLM uses a dataset of more than 600,000 instructions derived from over 300 real-world industry use cases, significantly enhancing its ability to perform financial-specific tasks.
Ant Group's self-developed general-purpose LLM served as the basis for the optimization of the financial LLM. The general-purpose model is designed to make efficient use of computing power. It features a heterogeneous hardware cluster that can support up to ten thousand GPUs of varying types. The model FLOPs utilization (MFU) for a thousand GPUs can reach 40%. Furthermore, the Reinforcement Learning from Human Feedback (RLHF) training throughput of the model is improved by a factor of 3.59 while maintaining consistent performance. The inference performance is also enhanced by approximately two times compared to the industry benchmark.
The financial LLM can support a diverse range of professional services, including wealth management services such as financial product evaluation, market analysis, and investor education, as well as insurance services such as explaining insurance products, creating family insurance plans, and verifying insurance claims.
