What is DianJin-R1
DianJin-R1 is a financial-domain reasoning-enhanced large model jointly launched by the Alibaba CloudTongyi Dianjinteam and Soochow University. It is designed specifically for financial tasks and integrates advanced technology and comprehensive data support. The model improves performance on financial reasoning tasks through reasoning-enhanced supervision and reinforcement learning. At its core is the DianJin-R1-Data dataset, which integrates the CFLUE, FinQA, and China Compliance Check (CCC) datasets, covering diverse financial reasoning scenarios. DianJin-R1 includes two versions, DianJin-R1-7B and DianJin-R1-32B, both optimized through two stages: supervised fine-tuning (SFT) and reinforcement learning (RL). It adopts the Group Relative Policy Optimization (GRPO) method and combines dual reward signals to optimize reasoning quality. On financial benchmarks such as CFLUE, FinQA, and CCC, DianJin-R1 significantly outperforms non-reasoning models. On the CCC dataset, the single-call reasoning model outperforms multi-agent systems.
Main features of DianJin-R1
- Financial reasoning enhancement: Through reasoning-enhanced supervision and reinforcement learning, it improves reasoning capabilities on financial tasks. On the financial test sets CFLUE, FinQA, and CCC, DianJin-R1 outperforms the base model.
- High-quality dataset support: Built on the DianJin-R1-Data dataset, it integrates the CFLUE, FinQA, and CCC datasets, covering a variety of financial reasoning scenarios.
- Structured reasoning output: The model can generate reasoning steps and final answers in a structured format, making them easy to understand and apply.
- Reinforcement learning optimization: It adopts the Group Relative Policy Optimization (GRPO) algorithm and combines format rewards and accuracy rewards to optimize reasoning quality.
- Efficient reasoning capability: In practical applications, the single-call reasoning model performance of DianJin-R1 is comparable to or even better than multi-agent systems, with lower computational cost.
Technical principles of DianJin-R1
- High-quality dataset construction: The core of DianJin-R1 is the DianJin-R1-Data dataset, which integrates the CFLUE, FinQA, and proprietary China Compliance Check (CCC) datasets. It covers diverse financial reasoning scenarios, and the validated annotations ensure high data quality and domain relevance.
- Structured supervised fine-tuning (SFT): The model is trained through a supervised fine-tuning (SFT) stage, using a structured format to generate reasoning steps and final answers. Specifically, each training instance includes a question, a reasoning path (formatted as
<think>...</think>) and an answer (formatted as<answer>...</answer>). By learning these structured inputs and outputs, the model generates coherent reasoning steps and correct answers. - Reinforcement learning optimization: On the basis of 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 structured output) and an accuracy reward (rewarding correct answers), further improving reasoning quality.
Project address of DianJin-R1
- 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 checking: DianJin-R1 can efficiently handle complex financial compliance tasks. For example, in its application on the CCC (China Compliance Check) dataset, the model can achieve performance comparable to or even better than a multi-agent system with a single call, significantly reducing computational costs.
- Financial Q&A and Customer Service: In tests on the FinQA dataset, DianJin-R1 demonstrated strong financial Q&A capabilities and can provide customers with accurate financial information and solutions. This helps improve the quality and efficiency of customer service at financial institutions.
- Financial Examinations and Education: DianJin-R1's performance on the CFLUE dataset is particularly outstanding, with accuracy improving from 77.95% to 86.74%. This indicates that the model can be used in auxiliary training and educational scenarios for financial examinations, helping candidates better understand and master financial knowledge.
- Reasoning and Decision-Making for Complex Financial Tasks: Through structured reasoning and reinforcement learning optimization, DianJin-R1 can handle complex financial tasks, such as financial statement analysis and investment decision-making.