Fin-R1
7B financial reasoning LLM from Shanghai University of Finance and Economics: SFT plus GRPO on 60K CoT samples, FinQA 76.0.
About
Fin-R1 is a financial reasoning model from the SUFE-AIFLM-Lab (Shanghai University of Finance and Economics, with Fudan and FinStep collaborators), released March 2025. It starts from Qwen2.5-7B-Instruct and trains in two stages: supervised fine-tuning on about 60,000 bilingual Chinese-English financial chain-of-thought samples, then GRPO reinforcement learning with rule-based rewards for format and answer accuracy. The paper reports 76.0 on FinQA and 85.0 on ConvFinQA, for a 75.2 average across five financial benchmarks, beating the 70B DeepSeek-R1 distillate by 8.7 points.
It is a clean example of the post-R1 playbook applied to a vertical: small base, reasoning-focused data, RL on top. The honesty footnote is that the official Fin-R1-Data training set was never publicly released despite the paper's promise, which limits reproducibility, and the Hugging Face model card declares no license. The weights are downloadable and the paper is thorough, but anyone building on it should know the data side is a black box.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Authors report FinQA 76.0 and ConvFinQA 85.0, 75.2 average across five financial benchmarks (provider claim, arXiv:2503.16252)
- Authors report beating DeepSeek-R1-Distill-Llama-70B by 8.7 points while trailing full DeepSeek-R1 (78.2) by 3.8 points (provider claim)
- Official Fin-R1-Data training set was never publicly released, limiting reproducibility (third-party replication notes)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 5/10
Scored 5/10 (metrics tier, metrics-v1). Evidence: paper arXiv:2503.16252 (Mar 2025, SUFE/Fudan/FinStep) with FinQA 76.0 and ConvFinQA 85.0 figures; HF repo SUFE-AIFLM-Lab/Fin-R1 created Mar 17, 2025 with 2,884 downloads and 230 likes (HF API, Oct 2026); third-party replication notes confirm the promised Fin-R1-Data release never happened. Weights downloadable. Strengths: strong reported numbers for a 7B model, thorough paper, real community interest. Docked for: (1) training data never released despite the promise, hurting reproducibility; (2) no license declared on the model card; (3) all benchmark figures are the authors' own. License: none declared.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 4.0/10
- Manual review · sanity read + red-flag check
- 5.0/10
- · Paper arXiv:2503.16252 (Mar 2025, SUFE/Fudan/FinStep); FinQA 76.0, ConvFinQA 85.0 reported
- · HF repo SUFE-AIFLM-Lab/Fin-R1 created 2025-03-17; 2,884 downloads, 230 likes (HF API Oct 2026)
- · Promised Fin-R1-Data training set never publicly released (third-party replication notes)
- · No license declared on model card; all benchmarks author-reported