On September 8, at the Bund Conference in Shanghai, Ant Group officially released a large financial model. It is understood that the Ant Financial large
model is based on Ant's self-developed basic large model and is deeply customized for the financial industry. The underlying computing power cluster reaches the scale of 10,000 cards. This large
model focuses on the needs of real financial scenarios, and performs outstandingly in 28 types of financial exclusive tasks in the five dimensions of "cognition, generation, professional knowledge, professional logic, and compliance", and
in "extraction of research and judgment opinions" and "understanding of financial intentions" He has reached the level of industry experts in many fields such as "Financial Event Reasoning"
.
Currently, the Ant Financial large model has been fully tested on Ant Group’s wealth and insurance platforms .
On the same day, Ant Group also released two products based on the capabilities of financial large models: the intelligent financial assistant "Zhi Xiaobao 2.0" and the
intelligent business assistant "Zhi Xiaozhu" that serves financial industry experts, demonstrating Ant's transformation from basic large models to industry Full-stack layout and progress of large models and
industrial applications.
Among them, “Zhi Xiaobao 2.0” has been in internal testing for nearly half a year and will be launched after completing relevant filing work . "Zhixiaozhu" is being co-constructed in internal testing with Ant Platform partner institutions to
create a full-chain AI business assistant for financial consultants, insurance agents, investment research, financial marketing, insurance claims and other financial industry experts.
At this press conference, Ant also opened the financial-specific task evaluation set "Fin-Eval". This test set
evaluates the capabilities of large financial models from five dimensions and 28 categories, filling the industry gap in high-quality comprehensive evaluation sets.
Inject financial knowledge of hundreds of billions of tokens, focusing on large-scale industrial applications.
"General-purpose large models cannot be directly commercialized in professional and rigorous fields. In particular, financial services have a low tolerance for errors. Financial
large models must ensure domain knowledge and professional logic." Only with the rigor of implementation can we truly bring industrial value. Guaranteeing the four major capabilities of knowledge, professionalism,
language and security is a prerequisite, and it is also the true proposition of the industry to be solved by the financial model." Vice President of Ant Group,
Finance Wang Xiaohang, the person in charge of the big model, said that based on a large number of practices in financial scenarios, the Ant Financial Big Model
has formed a "big model + knowledge + service" driven architecture. This architecture has been tested internally on Ant's internal financial intelligence scenarios.
In terms of knowledge, it is reported that the Ant Financial large model injects financial knowledge of hundreds of billions of Tokens based on the universal corpus of trillions of Tokens,
and extracts a total of 600,000+ high-quality instruction data from 300+ real industry scenarios. , forming
advantageous data assets for performance optimization of financial-specific tasks.
In terms of professional capabilities, thanks to Ant's ten years of accumulation, there is a complete matrix of digital financial tools on the platform. The Ant Financial
large model can accurately call these professional tools within the Ant system by understanding the user's language, providing users with corresponding professional capabilities. Services
, the financial side includes six major categories of services including financial product selection, product evaluation, market interpretation, and asset allocation, while the insurance side includes more than 10
smart services such as product interpretation, home configuration, smart underwriting, and smart claims.
To address the security and controllability issues of content generation, Ant Financial's large model uses a combination of intent recognition and fact verification to
effectively improve the compliance, security and authenticity of generated content.
In order to systematically evaluate the performance of AI in the financial field, Ant has defined the financial AI task evaluation set
"Fin-Eval" from real financial scenarios. Fin-Eval represents the needs of real industry scenarios and is currently the most extensive and professional evaluation set in the field of financial intelligence. It is composed of a total of 28
categories from five dimensions: "cognition, generation, domain knowledge, financial logic, and security compliance".
. At this Bund Conference, Ant Group also officially opened Fin-Eval to the outside world, hoping to promote the common progress of industry technology
.
Wang Xiaohang believes that large models are bringing experience changes to the financial industry: more natural interactions, richer supplies, more effective
expressions, more personalized service customization, and more efficient services. “Every key function in the financial business chain is worth
redoing using large model technology.”
It is understood that at the end of August, Ant Financial’s large model has passed the qualifications of securities practitioners, insurance practitioners, practicing physicians, and
practicing pharmacists. Professional test questions.
At present, Ant Financial's large model has taken the lead in application testing in the fields of financial management and insurance. In the future, all digital financial businesses that Ant Group
cooperates with financial institutions will be fully integrated into this large model to help cooperative institutions digitally upgrade and intelligently transform.
On the day of the release of two major products: 2C's "Zhi Xiaobao 2.0" and 2B's "Zhi Xiaozhu 1.0"
, Ant Group also released its first application product based on a large financial model - the intelligent financial assistant "Zhi Xiaobao 2.0",
and "Zhixiaozhu 1.0", an intelligent business assistant for financial industry experts, demonstrates Ant's
full-stack layout and progress in the field of large models from technology to industry applications.
According to reports, "Zhi Xiaobao 2.0" has outstanding performance in four aspects: language power, knowledge power, professional power, and security power. It can provide users with high-quality
market analysis, position diagnosis, asset allocation, investment education and companionship and other professional services. “Zhi Xiaobao 2.0” has high
-precision intention understanding and personalized communication style: the financial intention recognition accuracy reaches 95%, its financial event analysis and reasoning ability is no less than that of
real industry experts, and it can conduct multiple rounds of high-quality conversations.
Version 1.0 of the intelligent business assistant "Zhixiaozhu" includes
six versions: "Service Expert Edition", "Investment Research Expert Edition", "Claims Expert Edition", and "Insurance Research Expert Edition", providing comprehensive services for different financial services Practitioners in this scenario can
provide in-depth intelligent services in investment research analysis, information extraction, professional creation, business opportunity insights, and use of financial tools.
Taking the "Investment Research Support Assistant" as an example, actual measured data shows that the Investment Research Analyst can assist each investment research analyst to complete the financial logic and viewpoint extraction of
more than , and 40+ The reasoning and attribution of financial events doubles the efficiency of analysis.
At the same time, Zhixiaozhu can basically replace basic financial engineering code writing, greatly improving the efficiency of quantitative research. With the help of "service
support", the effective account management radius of financial consultants and insurance agents can be expanded by more than 70% per person.
It is understood that Zhi Xiaobao 2.0 has been in internal testing for nearly half a year and will be launched after the registration work is completed. Zhixiaozhu is being
built in internal testing with cooperative institutions on the Ant platform, and will be officially opened to cooperative institutions on the Ant platform when it matures.
Continue to tackle artificial intelligence and explore the five major capabilities of large models
. According to reports, the Ant Financial large model released today is based on the Ant Basic Large Model and is deeply customized for the financial industry. The Ant Basic
Large Model Platform has a 10,000-ka heterogeneous cluster, in which the 1,000-ka scale training MFU can reach 40%, and the effective training time of the cluster accounts for 40% of the total.
Compared with more than 90%, the training throughput performance of RLHF training is 3.59 times higher than the industry solution under the same model effect, and the inference performance is
about 2 times higher than the industry solution. It is at the advanced level in the industry and provides strong support for the industrial application of large models. .
Ant Group stated that it will continue to explore and improve the five major capabilities of large models in the future. The first is to build a high-quality data annotation
team and precipitate a high-quality data system; the second is to tackle basic large model algorithms and efficient green engineering capabilities to improve
model logical reasoning and other capabilities; the third is to move from a general-purpose language large model to a general-purpose Multi-modal large models, from general knowledge to comprehensive
professionalism; fourth, build efficient large model evaluation standards and evaluation systems to speed up the iteration of large models; fifth, build large model
safety capabilities to ensure the healthy and sustainable development of large models .
Public information shows that Ant Group has continued to invest in artificial intelligence based on the needs of rich business scenarios, and has laid out
AI technology fields including knowledge graphs, operations optimization, graph learning, trusted AI, large models, etc. Ant Group’s artificial intelligence
technology not only supports the comprehensive intelligent upgrade of its own business, but is also being open sourced to the society to promote the digital
transformation of the entire society.