aliyun/qwen-dianjin

What it solves

Qwen DianJin is a comprehensive research hub for applying large language models (LLMs) to the financial sector. It addresses the chronic scarcity of high-quality financial data and the need for specialized reasoning, tool use, and multimodal understanding of financial documents to create reliable AI agents for banking, insurance, and asset management.

How it works

The project employs a "Data Flywheel" methodology: Evaluation (using industry-grade benchmarks) $\rightarrow$ Data Synthesis (creating controllable trajectory data) $\rightarrow$ Post-Training (using SFT and RL to convert data into capabilities). This cycle is applied across several research directions:

• Reasoning & Post-Training: Developing specialized models like DianJin-R1 and the first financial-domain process reward model (Fin-PRM).
• Agentic Capabilities: Creating skill libraries (DianJin-SKILLS) and autonomous planning frameworks (DianJin-TIR).
• Multimodal Understanding: Using reasoning-and-tool interleaved VLMs (DianJin-OCR-R1) for financial document analysis.
• Dialogue Systems: Implementing cognitive-emotion-strategy reasoning chains for customer and emotional support (CSC and CARE).
• Evaluation: Building tiered benchmarks such as CFLUE and FinMCP-Bench to measure foundational and agentic performance.

Who it’s for

This project is designed for AI researchers, financial institutions, and developers building specialized financial AI agents and reasoning models.

Highlights

• DianJin-R1: A series of financial reasoning models (7B, 13B, 32B) and associated datasets.
• DianJin-SKILLS: A plug-and-play AI Agent skill library covering 10 professional roles and 130+ standardized skills.
• Fin-PRM: A domain-specialized process reward model for financial reasoning.
• DianJin-RED: An action-grounded red-teaming benchmark for agent systems with 1,661 executable cases.
• DianJin-OCR-R1: A vision-language model enhancing OCR capabilities for financial documents.