Fin-R1 is a large language model designed for complex reasoning in the financial domain. It was jointly developed and open-sourced by Professor Zhang Liwen of the School of Statistics and Data Science at Shanghai University of Finance and Economics and the Financial Large Language Model Research Group (SUFE-AIFLM-Lab) led by him, together with Caiyue Xingchen. The model is based on Qwen2.5-7B-Instruct and adopts a lightweight 7B parameter-scale architecture. While significantly reducing deployment costs, through fine-tuning training on high-quality verifiable financial problems, its final performance reaches SOTA level among participating models on multiple financial domain benchmarks.
[Purpose] Fin-R1 is specifically designed for the field of financial reasoning. By adopting two-stage training with SFT (supervised fine-tuning) and RL (reinforcement learning) on high-quality chain-of-thought data for financial reasoning scenarios, it provides theoretical support, business rules, decision logic, and technical implementation capabilities for the model's application in the financial domain, thereby effectively enhancing complex financial reasoning ability and providing support for core financial business scenarios such as banking, securities, insurance, and trusts.
[Main Capabilities] The model covers the following application scenarios: financial code, financial calculation, English financial calculation, financial security and compliance, intelligent risk control, and ESG analysis. Among them, financial code refers to computer programming code used in the financial domain to implement various financial models, algorithms, and analytical tasks, covering multiple aspects from simple financial calculations to complex financial derivative pricing, risk assessment, and portfolio optimization; financial calculation is the process of quantitative analysis and calculation of various problems in the financial domain, with the core being the establishment of mathematical models and the use of numerical methods to solve practical financial problems; English financial calculation emphasizes the use of English in cross-language environments for financial model construction and calculation, and the ability to write financial analysis reports in English and communicate with international peers; financial security and compliance focuses on preventing financial crime and complying with regulatory requirements, helping enterprises establish and improve compliance management systems; intelligent risk control uses AI and big data technology to identify and manage financial risks, with higher efficiency, accuracy, and real-time capability; ESG analysis measures a company's sustainable development capability by evaluating its performance in environmental, social, and governance aspects.
[Data Construction] The team built a data distillation framework based on DeepSeek-R1, strictly processed data according to official parameter settings, and adopted a two-stage data screening method to improve the quality of data in the financial domain, generating SFT datasets and RL datasets. Using DeepSeek-R1 (full version) for industry corpora (FinCorpus, Ant_Finance), professional cognition (FinPEE), business knowledge (FinCUGE, FinanceIQ, Finance-Instruct-500K), table parsing (FinQA), market insights (TFNS), multi-turn interaction (ConvFinQA), and quantitative investment (FinanceQT) for domain knowledge distillation and screening, constructing approximately 60k high-quality COT dataset Fin-R1-Data oriented to professional financial reasoning scenarios. This dataset covers multi-dimensional professional knowledge in the Chinese and English financial vertical domains, and is divided into four major modules: financial code, financial professional knowledge, financial non-reasoning business knowledge, and financial reasoning business knowledge.
For data screening, an innovative two-round quality scoring and screening method of "answer + reasoning" for the chain of thought is proposed: the first round scores answer accuracy based on rule matching and Qwen2.5-72B-Instruct; the second round performs in-depth verification of the reasoning logic of the reasoning chain, including logical consistency and terminology compliance. The data selected by each round of scoring is labeled as good or bad for distinction. The reasoning process scoring adopts indicators such as internal consistency, terminology overlap, number of reasoning steps, logical consistency, content diversity, relevance to the task domain, and consistency with task instructions. After two rounds of screening, data labeled as good in both rounds is used as high-quality COT data for SFT; data not screened and labeled as bad is used as reasoning QA data for reinforcement learning (RL).
[Training Method] For complex reasoning tasks in the financial domain, Qwen2.5-7B-Instruct is used for two-stage fine-tuning training: the first stage uses the ConvFinQA and FinQA financial datasets for supervised fine-tuning (SFT) to help the model initially improve its financial reasoning ability; the second stage, based on the GRPO (Group Relative Policy Optimization) algorithm, combines format rewards and accuracy rewards for reinforcement learning, and introduces a Model-Based Verifier, using Qwen2.5-Max for answer evaluation to improve the possible bias of regex-based rewards, thereby enhancing the effectiveness and stability of reinforcement learning.
[Model Evaluation Results] On benchmark tests covering multiple financial business scenarios, the model Fin-R1-SFT that only underwent instruction fine-tuning (SFT) has already achieved a certain performance improvement compared with the base model; Fin-R1, after further reinforcement learning training on the basis of Fin-R1-SFT, achieves an average score of 75.2 with only a lightweight parameter scale of 7B, ranking second, comprehensively surpassing participating models of the same scale, and with the industry benchmark DeepSeek-R1, the average score gap is only 3.0, and it surpasses DeepSeek-R1-Distill-Llama-70B (69.2) by 6.0 points. On the two key task tests, FinQA, which focuses on real financial table numerical reasoning tasks, and ConvFinQA, which involves multi-turn reasoning interaction scenarios, it ranks first among participating models with scores of 76.0 and 85.0, respectively.
[Usage] The model weights can be downloaded directly from huggingface. For the specific dependent environment configuration and installation commands and other usage instructions of the project, the materials only mention "prepare the dependent environment and use the following command to install vllm with one click," and do not provide the complete installation command and other step details.
[Applicable Scenarios] Applicable to core financial business scenarios such as banking, securities, insurance, and trusts, including financial code writing, financial calculations, English financial calculations, financial security and compliance, intelligent risk control, ESG analysis, etc.
[Limitations and Statement] As a reasoning-oriented large language model in the financial domain, although Fin-R1 can complete many financial tasks, at the current stage it still has technical bottlenecks and application limitations. The suggestions and analysis results it provides are for reference only and cannot be equated with the precise judgment of professional financial analysts or experts. Users should examine model outputs with critical thinking and make decisions in combination with their own professional knowledge and experience. The project was completed by the Financial Large Language Model Research Group of the School of Statistics and Data Science, Shanghai University of Finance and Economics (SUFE-AIFLM-Lab) together with Caiyue Xingchen, and can be contacted via zhang.liwen@shufe.edu.cn.
