Fin-R1 - A large-scale financial reasoning model jointly launched by Shanghai Finance and Economics and Caiyue Xingchen.

What is Fin-R1?

Fin-R1 is the first large-scale R1-class inference model in the financial field, jointly launched by Shanghai University of Finance and Economics and Caiyue Xingchen. Based on the Qwen2.5-7B-Instruct architecture with 7B parameters, it effectively improves complex financial reasoning capabilities through two-stage training using SFT and RL on high-quality thought chain data in financial reasoning scenarios. In authoritative evaluations, Fin-R1 achieved an average score of 75.2, only 3 points behind the industry benchmark DeepSeek-R1, ranking second on the list. The data construction integrates high-quality datasets from multiple financial fields, constructing approximately 60k high-quality COT datasets through data distillation.

Fin-R1's main functions

• Financial Reasoning and Decision MakingIt can handle complex financial reasoning tasks, such as numerical reasoning of financial data, sentiment classification of financial news, and causal relationship extraction, providing accurate and interpretable basis for financial decision-making.
• Automated financial business processesIt performs exceptionally well in practical applications such as financial compliance checks and robo-advisors, automating financial business processes, improving efficiency and reducing labor costs.
• Multilingual supportIt supports financial reasoning in both Chinese and English, covering a variety of financial business scenarios and meeting the financial reasoning needs in different language environments.
• Efficient resource utilizationAchieving high performance with a lightweight structure of 700 million parameters significantly reduces deployment costs and is more suitable for use in resource-constrained environments.
• Financial code generationSupports the generation of programming code for various financial models and algorithms.
• Financial computingTo conduct quantitative analysis and calculation of complex financial problems.
• English Financial CalculationSupports the use of English to build and write financial models.
• Financial security and compliance: Help businesses ensure their operations comply with relevant regulations.
• Intelligent risk controlUtilizing AI technology to identify and manage financial risks, thereby improving decision-making efficiency.
• ESG analysisTo assess a company's sustainable development capabilities and promote the fulfillment of its social responsibilities.

Fin-R1's technical principles

• Model ArchitectureFin-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, making it more suitable for use in resource-constrained environments.
• Data buildingFin-R1 addresses the fragmentation of financial data by constructing the high-quality financial inference dataset Fin-R1-Data. The dataset contains approximately 60,000 high-quality COT (Consumer-Oriented) data entries tailored to professional financial inference scenarios. The dataset construction process involved domain knowledge distillation and screening from multiple authoritative data sources, employing a dual-round quality scoring method of "answer + reasoning" to ensure data accuracy and reliability.
• Phase One – Injection of Reasoning AbilitySupervised fine-tuning (SFT) of Qwen2.5-7B-Instruct was performed using the ConvFinQA and FinQA financial datasets to help the model initially improve its financial inference capabilities.
• Phase Two – Reinforcement Learning OptimizationAfter mastering complex reasoning skills, the GRPO (Group Relative Policy Optimization) algorithm is adopted as the core framework, combining format rewards and accuracy rewards for reinforcement learning. Simultaneously, a model-based verifier is introduced, using Qwen2.5-Max for answer evaluation, generating more accurate and reliable reward signals, thus improving the effectiveness and stability of reinforcement learning.

Fin-R1 project address

• HuggingFace model library:https://huggingface.co/SUFE-AIFLM-Lab/Fin-R1
• arXiv technical paper:https://arxiv.org/pdf/2503.16252

Application scenarios of Fin-R1

• Intelligent risk controlIn the field of intelligent risk control, Fin-R1's dynamic credit scoring model makes risk assessment more accurate, can monitor transaction anomalies in real time, and effectively prevent financial risks.
• Investment Decision SupportIn fund investment, it can assist investment advisors in asset allocation and help users make more informed decisions.
• Quantitative TradingIn securities trading, Fin-R1 can be used to write quantitative trading code, improving the coding efficiency of practitioners and helping to develop quantitative trading strategies.
• ESG analysisIt can assist in generating ESG reports that comply with GRI standards, helping companies achieve green transformation and meet market requirements for sustainable development.
• Market trend forecastIn the insurance industry, Fin-R1 can efficiently assess policy returns and predict market trends.