Du Xiaoman Open Source 100 Billion Parameter Financial Model "Xuanyuan"

On May 26th, it was reported that recently, Du Xiaoman officially opened up the "Xuanyuan", a Chinese financial model worth hundreds of billions. The Xuan Yuan model is trained on the basis of the Bloom model with 176 billion parameters, and its performance in tasks such as understanding financial terms, commenting on financial markets, analyzing financial data, and understanding financial news has significantly improved compared to the general model

On May 26th, it was reported that recently, Du Xiaoman officially opened up the "Xuanyuan", a Chinese financial model worth hundreds of billions. The Xuan Yuan model is trained on the basis of the Bloom model with 176 billion parameters, and its performance in tasks such as understanding financial terms, commenting on financial markets, analyzing financial data, and understanding financial news has significantly improved compared to the general model.

According to Du Xiaoman, in the task evaluation of financial scenarios, Xuanyuan comprehensively surpassed the mainstream open source big models in the market, winning a 63.33% victory rate out of 150 responses. In the general ability assessment, 10.2% of tasks performed by Xuanyuan exceeded ChatGPT3.5, while 61.22% performed equally, involving 13 main dimensions such as mathematical calculation, scenario writing, logical reasoning, and text summarization.

In order to enhance the Xuanyuan Model's understanding of financial issues, Du Xiaoman used the Chinese pre training dataset of hundreds of billions of tokens accumulated in his own business in the financial field to train the model. This dataset covers professional knowledge in various fields such as financial research reports, stocks, funds, banks, insurance, etc. Du Xiaoman said that the cleaned and labeled high-quality data set not only made it possible to achieve the same level of universality as ChatGPT, but also improved the performance of the model in the financial vertical field.

BLOOM (Big Science Language Open science Open access Multilingual) was created by over 1000 volunteer researchers in a project called "Big Science" in 2021 and officially released on July 12, 2022. BLOOM has 176 billion parameters (variables that determine how input data is converted into output content), slightly more than GPT-3 with 175 billion parameters. BLOOM has 1.61TB of text, including 46 natural languages and 13 programming languages. Compared to the 13 billion parameter LLaMA (Large Language Model MetaAI) model released by Meta, Bloom has an advantage in terms of parameter quantity.

At present, the Xuanyuan model, which is worth billions, can be applied for download in Huggingface and is open to all financial institutions.

Xu Dongliang, CTO of Du Xiaoman, stated that the Xuanyuan Big Model is trained from financial data accumulated in the business scenarios of Du Xiaoman, and has an advantage in understanding financial related issues compared to the general big model. We open up the capability of big models to financial institutions, which is conducive to promoting the application of big models in the financial industry, reducing the application threshold of big models, and improving the intelligence level of the financial industry.

Xu Dongliang believes that generative large models have excellent abilities in content generation and creation, information summarization and summarization, knowledge understanding and Q&A, natural interaction and dialogue, and will be widely applied in financial scenarios. At the front desk, the generative model will significantly enhance the professional level and service capabilities of account managers, significantly reduce their operational costs, and make it possible for everyone to have a professional account manager who is available 24 hours a day. Excellent content generation capabilities will also lead to a significant increase in marketing content production capacity. In the middle stage, generative large models have the opportunity to change the way knowledge acquisition, content creation, meetings and communication, code development and testing are conducted within the enterprise, thereby significantly improving internal office efficiency and even triggering a change in research and development testing mode, comprehensively improving the internal operational efficiency of financial enterprises. In the backend, large models will become the standard configuration of intelligent technology bases, greatly reducing the threshold for intelligent technology applications. With only a small amount of annotated data or even no adjustment, intelligent technology can cover a wide range of scenarios. (One Orange)