Development and Application of 100B-level Parameter Financial LLM
To tackle the difficulties applying LLMs in financial industry, including lack of experience in building LLMs, shortage of business-oriented solutions for LLM implementations, and insufficiencies on methodological guidance for large-scale LLM applications, ICBC carries out the duty of the leading bank to establish the first enterprise-level 100B-level parameter financial LLM technology system in the banking industry. This LLM technology system integrates infrastructure, computing power, algorithms, data, tools, solutions, security, applications, and ecosystems, thus forms a LLM-centered new-generation enterprise-level financial business empowerment method, and promotes the large-scale, end-to-end innovative applications of LLMs in more than 20 business areas, such as remote banking and financial markets. ICBC also cultivates a infrastructure ecosystem in the industry, and becomes the first to deliver it to peer banks, so to provide powerful propulsion for developing new quality productive forces in the "AI + Finance" industry.
Artificial Intelligence And Digital Services
Artificial intelligence
The Project began its research on LLM-related technologies in June 2022, commenced implementation developments in January 2023, went into production in April 2023, and was comprehensively applied to various business areas from June 2023 to June 2024.
49 countries and regions
The construction and application of financial domain LLM have no precedents to follow, neither on the industry side nor on the enterprise side, making the task extremely difficult. This project addresses such pain point by achieving the following innovations: (Ⅰ)To address the insufficiencies on methodological guidance for the large-scale application of financial LLM, we have pioneered the development methodology for LLM application in the financial industry. By means of front-line investigations, full-lifecycle business mapping, and other approaches, we effectively promoted end-to-end sorting of scenarios, and empowered more than 20 business domains by engineering practice via the process of ”3 benchmark scenarios - 1 standard technical solution - large-scale promotion” for various scenarios. Since its birth, a total of 500+ million LLM services calls have been made. (Ⅱ)To address the lack of practice in the construction of financial LLM, we have pioneered through the construction of a 100-billion-level financial LLM technology system in the industry. Firstly, we have built a AI computing cloud of thousand-level GPUs, with training and inference performance and stability reaching an international leading level. Secondly, we have created a "1+X" financial LLM solution with intelligent command center at its core, allowing zero-code minute-level construction of financial intelligent agents. Thirdly, we have built a 100-billion-scale financial LLM algorithm matrix based on collaboration of large and small models, and passed multiple financial capability certification tests. Fourthly, we have built a low-threshold, intelligent, and end-to-end financial knowledge engineering streamline, creating a high-quality multi-dimensional and trillion-token financial dataset. Fifthly, based on integration of data, AI and application, we formed an agile R&D platform together with a universally safeguarding security platform, with high development efficiency and reliable application security. (Ⅲ)To address the shortage of industry-level general technical capabilities for the empowerment of "AI + Finance” ecosystem, we have built financial LLM and its solutions for the financial industry, and have pioneered the delivery of these solutions to small and medium banks and financial institutions.
（Ⅰ）[
(Ⅰ)Financial Research Report Generation Methods, Apparatus, Devices, Medium, and Program Products; (Ⅱ) Intelligent Document Q&A Methods, Apparatus, Devices, Medium, and Program Products; (Ⅲ)Audit Report Generation Methods, Apparatus, Devices, Medium, and Program Products; (Ⅳ)A Merchant Qualification Audit Method and Apparatus based on Multi-Modal Data Processing
