DianJin-R1 - A large-scale financial reasoning model jointly launched by Alibaba Cloud Tongyi Dianjin and Soochow University

What is DianJin-R1?

DianJin-R1 is a large-scale enhanced inference model for the financial domain, jointly developed by Alibaba Cloud's Tongyi Dianjin team and Soochow University. Designed specifically for financial tasks, it integrates advanced technology and comprehensive data support. The model improves performance on financial inference tasks through enhanced supervision and reinforcement learning. Its core is the DianJin-R1-Data dataset, which integrates CFLUE, FinQA, and China Compliance Inspection (CCC) datasets, covering diverse financial inference scenarios. DianJin-R1 includes two versions, DianJin-R1-7B and DianJin-R1-32B, both optimized through a two-stage process of supervised fine-tuning (SFT) and reinforcement learning (RL), employing the Group Relative Policy Optimization (GRPO) method and combining dual reward signals to optimize inference quality. In benchmark tests such as CFLUE, FinQA, and CCC in the financial domain, DianJin-R1 significantly outperforms non-inference models. On the CCC dataset, the performance of the inference model in a single invocation surpasses that of a multi-agent system.

Main functions of DianJin-R1

• Enhanced Financial ReasoningThis approach enhances the reasoning capabilities for financial tasks by incorporating supervision and reinforcement learning. On the financial test sets CFLUE, FinQA, and CCC, DianJin-R1 outperforms the base models.
• High-quality dataset supportBuilt on the DianJin-R1-Data dataset, it integrates the CFLUE, FinQA, and CCC datasets, covering a variety of financial inference scenarios.
• Structured reasoning outputThe model can generate reasoning steps and final answers in a structured format, making it easy to understand and apply.
• Reinforcement learning optimizationThe Group Relative Policy Optimization (GRPO) algorithm is used, which combines format rewards and accuracy rewards to optimize inference quality.
• High-efficiency reasoning abilityIn practical applications, DianJin-R1's single-invocation inference model performs comparable to or even better than multi-agent systems, with lower computational costs.

Technical Principles of DianJin-R1

• Building high-quality datasetsAt the heart of DianJin-R1 is the DianJin-R1-Data dataset, which integrates CFLUE, FinQA, and the proprietary China Compliance Inspection (CCC) dataset. It covers diverse financial inference scenarios, with validated annotations ensuring high data quality and domain relevance.
• Structured Supervised Fine-Tuning (SFT)The model is trained through a supervised fine-tuning (SFT) phase, generating inference steps and the final answer using a structured format. Specifically, each training instance includes a question, an inference path (formatted as...) <think>...</think>) and answers (formatted as <answer>...</answer>The model learns from these structured inputs and outputs to generate coherent reasoning steps and correct answers.
• Reinforcement learning optimizationBuilding upon supervised fine-tuning, DianJin-R1 further applies the Group Relative Policy Optimization (GRPO) algorithm for reinforcement learning. This algorithm introduces dual reward signals: a format reward (encouraging the generation of structured outputs) and an accuracy reward (rewarding correct answers), further improving inference quality.

DianJin-R1's project address

• Github repository:https://github.com/aliyun/qwen-dianjin
• HuggingFace model library:https://huggingface.co/DianJin
• arXiv technical paper:https://arxiv.org/pdf/2504.15716

Application scenarios of DianJin-R1

• Financial risk assessment and compliance inspectionThe DianJin-R1 can efficiently handle complex financial compliance tasks. For example, in applications on the CCC (China Compliance Inspection) dataset, the model can achieve performance comparable to or even better than multi-agent systems with a single call, significantly reducing computational costs.
• Financial Q&A and Customer ServiceIn tests on the FinQA dataset, DianJin-R1 demonstrated powerful financial question-answering capabilities, providing customers with accurate financial information and solutions. This helps improve the quality and efficiency of customer service for financial institutions.
• Financial Examinations and EducationThe DianJin-R1 model performs particularly well on the CFLUE dataset, improving accuracy from 77.95% to 86.74%. This demonstrates that the model can be used for supplementary training in financial exams and educational scenarios, helping test takers better understand and master financial knowledge.
• Reasoning and decision-making for complex financial tasksDianJin-R1, optimized through structured reasoning and reinforcement learning, can handle complex financial tasks such as financial statement analysis and investment decisions.