Fin-R1: Financial Reasoning LLM

Fin-R1 is a 7.6 billion parameter large language model developed by SUFE-AIFLM-Lab and Caiyue Xingchen, designed for complex financial reasoning. Built on the Qwen2.5-7B-Instruct base, it undergoes a two-stage training process involving supervised fine-tuning (SFT) and reinforcement learning (RL) using a meticulously curated dataset of verifiable financial problems.

Key Capabilities

• Financial Reasoning: Achieves state-of-the-art performance on various financial benchmarks, including FinQA and ConvFinQA, demonstrating strong capabilities in numerical and multi-turn reasoning.
• Specialized Training Data: Utilizes a 60k entry, high-quality Chain-of-Thought (COT) dataset, Fin-R1-Data, covering diverse financial domains like code, professional knowledge, and quantitative investment.
• Reinforcement Learning Optimization: Employs GRPO (Group Relative Policy Optimization) with a dual reward mechanism (format and accuracy) and a Model-Based Verifier (using Qwen2.5-Max) to enhance accuracy and generalization.
• Lightweight Architecture: At 7B parameters, it offers significant performance advantages over other models in its size class, making it efficient for deployment.

Good For

• Financial Code Generation: Generating code for financial models and algorithms.
• Financial Calculations: Performing quantitative analysis and calculations for financial problems.
• Financial Security & Compliance: Assisting with regulatory adherence and financial crime prevention.
• Intelligent Risk Control: Identifying and managing financial risks through data analysis.
• ESG Analysis: Evaluating environmental, social, and governance performance for sustainable finance.