Fin-R1 — Open-Source Financial Reasoning LLM

Fin-R1 is a 7B-parameter open-source financial reasoning model based on Qwen2.5-7B-Instruct, designed for complex reasoning tasks across banking, securities, insurance and trust businesses.

Approach

• Distilled 60K financial chain-of-thought (CoT) samples from DeepSeek-R1.
• Two-round quality filtering: rule-based + model scoring + logical-consistency verification.
• Two-stage training: SFT for behavior cloning, then GRPO for reasoning refinement.
• Verifier-based reward signal to improve reliability.

Results

• FinEval 75.2 — 2nd place overall.
• Near-SOTA on FinQA (76.0) and ConvFinQA (85.0).
• Outperforms all comparable open models at the 7B scale.
• One-click deployment via vLLM.