What is Fin-R1
Fin-R1 is the first R1-class reasoning large model in the financial field, jointly launched by Shanghai University of Finance and Economics and Caiyue Xingchen. Based on the 7B-parameter Qwen2.5-7B-Instruct architecture, it effectively improves complex financial reasoning capabilities through two-stage training, SFT and RL, on high-quality chain-of-thought data in financial reasoning scenarios. In authoritative evaluations, Fin-R1 scored an average of 75.2 points, only 3 points behind the industry benchmarkDeepSeek-R1, ranking second on the leaderboard. The data construction integrates multiple high-quality datasets in the financial field and builds approximately 60k high-quality COT data entries through data distillation.
Main functions of Fin-R1
- Financial reasoning and decision-making: Capable of handling complex financial reasoning tasks, such as numerical reasoning on financial data, sentiment classification of financial news, and causal relationship extraction, providing accurate and interpretable support for financial decisions.
- Automating financial business processes: Performs excellently in practical applications such as financial compliance checks and robo-advisors, can automate the execution of financial business processes, improve efficiency, and reduce labor costs.
- Multilingual support: Supports financial reasoning in Chinese and English, covers a variety of financial business scenarios, and meets financial reasoning needs in different language environments.
- Efficient resource utilization: Achieves high performance with a lightweight structure of 700 million parameters, significantly reducing deployment costs and making it more suitable for use in resource-constrained environments.
- Financial code generation: Supports programming code generation for various financial models and algorithms.
- Financial computation: Performs quantitative analysis and computation of complex financial problems.
- English financial computation: Supports building and writing financial models in English.
- Financial security and compliance: Helps enterprises ensure that business operations comply with relevant regulations.
- Intelligent risk control: Uses AI technology to identify and manage financial risks and improve decision-making efficiency.
- ESG analysis: Evaluates enterprises' sustainable development capabilities and promotes the fulfillment of social responsibility.
Technical principles of Fin-R1
- Model architecture: Fin-R1 is based on the Qwen2.5-7B-Instruct architecture and adopts a lightweight 7B-parameter design. While ensuring model performance, the architecture significantly reduces deployment costs and is more suitable for use in resource-constrained environments.
- Data construction: Fin-R1 addresses the problem of fragmented financial data by building a high-quality financial reasoning dataset, Fin-R1-Data. The dataset contains approximately 60k high-quality COT data entries oriented toward professional financial reasoning scenarios. The dataset construction process includes domain knowledge distillation and screening from multiple authoritative data sources, and adopts an "answer + reasoning" two-round quality scoring and screening method to ensure the accuracy and reliability of the data.
- Phase one - reasoning capability injection: Uses the ConvFinQA and FinQA financial datasets to perform supervised fine-tuning (SFT) on Qwen2.5-7B-Instruct, helping the model initially improve its financial reasoning capability.
- Phase two - reinforcement learning optimization: After mastering complex reasoning skills, the GRPO (Group Relative Policy Optimization) algorithm is used as the core framework, combined with format rewards and accuracy rewards for reinforcement learning. At the same time, a Model-Based Verifier is introduced, using Qwen2.5-Max for answer evaluation to generate more precise and reliable reward signals and improve the effectiveness and stability of reinforcement learning.
Fin-R1's project address
- HuggingFace model repository：https://huggingface.co/SUFE-AIFLM-Lab/Fin-R1
- arXiv technical paper：https://arxiv.org/pdf/2503.16252
Fin-R1's application scenarios
- Intelligent risk control: In the field of intelligent risk control, Fin-R1's dynamic credit scoring model makes risk assessment more precise, can monitor transaction anomalies in real time, and effectively prevents financial risks.
- Investment decision assistance: In fund investment, it can assist investment advisors in asset allocation and help users make wiser decisions.
- Quantitative trading: In securities trading, Fin-R1 can participate in writing quantitative trading code, improve practitioners' coding efficiency, and support the development of quantitative trading strategies.
- ESG analysis: It can assist in generating ESG reports that comply with GRI standards, support enterprises' green transformation, and meet the market's requirements for corporate sustainable development.
- Market trend forecasting: In the insurance industry, Fin-R1 can efficiently evaluate policy returns and forecast market trends.