China’s Ant Group launched an AI model with hundreds of billions of token datasets and 600,000 instructions sourced from 300 financial industry use cases.
Heaptalk, Jakarta — Hangzhou-based financial technology company Ant Group officially introduced its financial large language model (LLM) at the Inclusion Conference on the Bund in Shanghai, China (09/08).
According to the Vice President of Ant Group and Head of Financial LLM Wang Xiaohang, the Ant Group’s financial LLM has been trained with more than hundreds of billions of token datasets including general corpus and Chinese financial documents. In addition, the AI model also has a dataset of at least 600,000 instructions sourced from 300 financial industry use cases to advance its capability for financial-specific tasks.
“Based on our extensive experiences from financial services businesses, Ant Group has summarized a set of best practices for the financial LLM. It takes the LLM as the center of cognition and interaction to understand financial intentions bring in domain knowledge and professional services and form an architecture integrating LLM technologies with domain knowledge and professional services. It provides support for consumer products and industry use cases in terms of language, knowledge, and expertise,” Xiaohang explained in an Ant Group official interview.
For this AI model, the company stated that it has developed a full technical and application stack, spanning the optimization of training and inference efficiency, improvement of financial domain knowledge and instruction, as well as implementation in large-scale consumer products and industrial applications. Xiaohang said that this end-to-end method is considered more reliable and sophisticated.
Releasing B2B & B2C AI-powered financial apps
Along with the launch, the affiliate company of Alibaba Group also released two applications armed with the AI model, called Zhixiaobao 2.0 and Zhixiaozhu 1.0. Zhixiaobao 2.0 is designed for consumers (2C) providing financial services such as market analysis, asset allocation suggestions, and investor education. Xiaohang said, “Our financial intent recognition is 95% accurate. When it comes to advanced capabilities such as expressing financial opinions and events reasoning, it reaches the average of professionals.”
Meanwhile, Zhixiaozhu 1.0 aims to serve industry professionals in the business segment (2B). This app offers 6 customized versions, including service representatives, investment advisors, and insurance claim specialists. According to Xiaohang, the AI model can extract and analyze up to 100 research reports and financial events per day based on the company’s development and test. Additionally, the AI model can effectively expand the service capabilities of wealth management consultants and insurance agents by more than 70% on average.
Currently, the LLM is still in the further development stage by Ant Group and will be released to the public when the technology is deemed mature. “Together with financial institutions in China, we are improving the capabilities of the financial LLM. When the financial LLM and its industry applications have reached a sufficient level of maturity, Ant Group will make it available to the public to help accelerate the process of intelligent industrial transformation,” Xioahang voiced.
In addition, Ant Group has made its financial-specific AI task benchmark named Fin-Eval available to the public. Fin-Eval masters 28 categories of specific financial tasks in five main areas, namely cognition, generation, domain knowledge, professional thinking, and compliance.
