Fino1

On the Transferability of Reasoning-Enhanced LLMs and Reinforcement Learning to Finance

Paper · Data · Models · Leaderboard

Overview

Fino1 studies how well reasoning-enhanced LLMs and reinforcement learning transfer to financial tasks. We build financial reasoning paths (FinCoT), train Fino1 models with a two-stage recipe (supervised fine-tuning followed by GRPO reinforcement learning), and evaluate them against a broad set of general-purpose and reasoning LLMs on FinQA, DocMath-Eval and XBRL-Math.

News

• [2025/03/30] We've trained the Fino1-14B model and evaluated its performance.
• [2025/02/12] We've trained the Fino1-8B model and evaluated its performance.

Datasets

We use three evaluation datasets to assess the performance of our Fino1 model.

Dataset	Description
FinQA	FinQA is a large-scale dataset for numerical reasoning in finance, featuring expert-annotated QA pairs that require integrating structured and unstructured data from financial reports while handling complex domain-specific terminology.
DocMath	DocMath-Eval (Simplong; Complong) is a benchmark for evaluating LLMs' numerical reasoning over long specialized documents and tables, with the Complong subset focusing on reasoning across multi-tiered financial or specialized tables within extended contexts.
XBRL-Math	The XBRL-Math dataset evaluates LLMs' numerical reasoning in XBRL filings, requiring models to interpret structured financial data, US GAAP XBRL tags, equations, and hierarchical numerical relationships for accurate financial analysis.

Models Evaluated

We compared our Fino1 model against 16 state-of-the-art large language models (LLMs).

Show the full list of evaluated models

Model	Description
GPT-4o	GPT-4o is OpenAI's versatile, high-intelligence flagship model. It accepts text and image inputs and produces text outputs (including Structured Outputs).
GPT-o1	The o1 series of models are trained with reinforcement learning to perform complex reasoning. o1 models think before they answer, producing a long internal chain of thought before responding to the user.
GPT-o3-mini	o3-mini is OpenAI's most recent small reasoning model, providing high intelligence at the same cost and latency targets of o1-mini. o3-mini also supports key developer features, like Structured Outputs, function calling, Batch API, and more.
GPT-4.5	GPT-4.5 is the largest and most capable GPT model yet. Its deep world knowledge and better understanding of user intent makes it good at creative tasks and agentic planning. GPT-4.5 excels at tasks that benefit from creative, open-ended thinking and conversation, such as writing, learning, or exploring new ideas.
DeepSeek-V3	DeepSeek-V3 is a 671B Mixture-of-Experts (MoE) model with 37B active parameters per token, leveraging Multi-head Latent Attention (MLA) and DeepSeekMoE for efficient training and inference, achieving state-of-the-art performance comparable to closed-source models with stable and cost-effective training.
DeepSeek-R1	DeepSeek-R1-Zero and DeepSeek-R1 are first-generation reasoning models, with DeepSeek-R1 incorporating cold-start data before RL to improve readability and performance, achieving results comparable to OpenAI-o1 across reasoning tasks, while open-sourced distilled models set new benchmarks for dense models.
Qwen2.5-7B-Instruct	Qwen2.5 is the latest series of Qwen LLMs, offering models from 0.5B to 72B parameters with improved knowledge, coding, math, instruction following, structured data handling, long-context support (up to 128K tokens), and multilingual capabilities across 29+ languages.
Qwen2.5-14B-Instruct	Qwen2.5 is the latest series of Qwen LLMs, offering models from 0.5B to 72B parameters with improved knowledge, coding, math, instruction following, structured data handling, long-context support (up to 128K tokens), and multilingual capabilities across 29+ languages.
Qwen2.5-32B-Instruct	Qwen2.5 is the latest series of Qwen LLMs, offering models from 0.5B to 72B parameters with improved knowledge, coding, math, instruction following, structured data handling, long-context support (up to 128K tokens), and multilingual capabilities across 29+ languages.
Qwen2.5-72B-Instruct	Qwen2.5 is the latest series of Qwen LLMs, offering models from 0.5B to 72B parameters with improved knowledge, coding, math, instruction following, structured data handling, long-context support (up to 128K tokens), and multilingual capabilities across 29+ languages.
Qwen/QwQ-32B	QwQ is the reasoning model of the Qwen series. Compared with conventional instruction-tuned models, QwQ, which is capable of thinking and reasoning, can achieve significantly enhanced performance in downstream tasks, especially hard problems.
Qwen2.5-Math-72B-Instruct	Qwen2.5-Math-72B-Instruct is an upgraded open-source mathematical LLM supporting both Chain-of-Thought (CoT) and Tool-integrated Reasoning (TIR) for solving math problems in Chinese and English, offering significant performance improvements over Qwen2-Math.
DeepSeek-R1-Distill-Llama-70B	DeepSeek-R1-Zero and DeepSeek-R1 are first-generation reasoning models, with DeepSeek-R1 incorporating cold-start data before RL to improve readability and performance, achieving results comparable to OpenAI-o1 across reasoning tasks, while open-sourced distilled models set new benchmarks for dense models.
Llama3-70B-Instruct	Meta released the Llama 3 family of 8B and 70B LLMs, optimized for dialogue, outperforming many open-source chat models while prioritizing helpfulness and safety.
Llama3.1-70B-Instruct	The Meta Llama 3.1 collection includes multilingual LLMs (8B, 70B, 405B) optimized for multilingual dialogue, outperforming many open-source and closed chat models on industry benchmarks.
Llama3.3-70B-Instruct	The Meta Llama 3.3 is a 70B instruction-tuned multilingual LLM optimized for dialogue, outperforming many open-source and closed chat models on industry benchmarks.
DeepSeek-R1-Distill-Qwen-32B	DeepSeek-R1-Zero and DeepSeek-R1 are first-generation reasoning models, with DeepSeek-R1 incorporating cold-start data before RL to improve readability and performance, achieving results comparable to OpenAI-o1 across reasoning tasks, while open-sourced distilled models set new benchmarks for dense models.
DeepSeek-R1-Distill-Qwen-14B	DeepSeek-R1-Zero and DeepSeek-R1 are first-generation reasoning models, with DeepSeek-R1 incorporating cold-start data before RL to improve readability and performance, achieving results comparable to OpenAI-o1 across reasoning tasks, while open-sourced distilled models set new benchmarks for dense models.
DeepSeek-R1-Distill-Llama-8B	DeepSeek-R1-Zero and DeepSeek-R1 are first-generation reasoning models, with DeepSeek-R1 incorporating cold-start data before RL to improve readability and performance, achieving results comparable to OpenAI-o1 across reasoning tasks, while open-sourced distilled models set new benchmarks for dense models.
Llama3-8B-Instruct	Meta released the Llama 3 family of 8B and 70B LLMs, optimized for dialogue, outperforming many open-source chat models while prioritizing helpfulness and safety.
Llama3.1-8B-Instruct	The Meta Llama 3.1 collection includes multilingual LLMs (8B, 70B, 405B) optimized for multilingual dialogue, outperforming many open-source and closed chat models on industry benchmarks.
LIMO	LIMO challenges the conventional wisdom in mathematical reasoning by demonstrating that models can achieve superior performance with significantly less but higher quality training data.
s1-32B	s1 is a reasoning model finetuned from Qwen2.5-32B-Instruct on just 1,000 examples. It matches o1-preview & exhibits test-time scaling via budget forcing.
FinR1-7B	Fin-R1 is a large language model designed for complex reasoning tasks in the financial domain. It was developed and open-sourced jointly by the Financial Large Language Model Research Group (SUFE-AIFLM-Lab) at the School of Statistics and Data Science of Shanghai University of Finance and Economics, in collaboration with Fintopia. The model is based on Qwen2.5-7B-Instruct and fine-tuned using high-quality, verifiable financial reasoning questions. It has achieved state-of-the-art performance across multiple financial benchmarks.

Reasoning Path Building

For the reasoning path building and training part, we were inspired by HuatuoGPT-o1. See CoT_build.py and process_data.ipynb.

We release the reasoning paths here: FinCoT.

Training

Following HuatuoGPT-o1, we train Fino1 in two stages:

• Stage 1: Supervised Fine-Tuning (SFT). We use HuatuoGPT-o1 for SFT; please check its code for more training details.
• Stage 2: Reinforcement Learning (RL). We use the code from open-r1 for GRPO; please check its code for more training details.

Evaluation

Inference: local models

Model inference for local models is conducted using FinBen with the vLLM framework.

Inference: API models

For API-based models, evaluation is performed using the query_llm.py script.

Scoring

For the final evaluation, we used DocMath-Eval to first use GPT to extract final answers from the results and then evaluate the correctness of the answers.

Key Results

Performance of different LLMs on financial datasets

Models	FinQA	DocMath-Simplong	XBRL-Math	DocMath-Complong	Average
GPT-4o	72.49	60.00	72.22	39.33	61.01
GPT-o1-preview	49.07	56.00	74.44	36.67	54.05
GPT-o3-mini	60.87	59.00	76.67	35.00	57.89
DeepSeek-V3	73.20	53.00	76.67	42.33	61.30
DeepSeek-R1	65.13	53.00	86.67	38.67	60.87
GPT-4.5	68.94	59.00	74.44	39.33	60.43
Meta-Llama-4-Scout	70.45	52.00	88.89	0.67	53.00
Meta-Llama-3-70B-Instruct	58.92	41.00	56.67	13.67	42.57
Llama-3.1-70B-Instruct	63.18	48.00	63.33	34.33	52.21
Llama-3.3-70B-Instruct	68.15	54.00	70.00	32.00	56.04
Qwen2.5-72B-Instruct	73.38	59.00	67.78	14.67	53.71
Qwen2.5-Math-72B-Instruct	69.74	42.00	83.33	5.00	50.02
DeepSeek-R1-Distill-Llama-70B	66.73	53.00	86.67	30.67	59.27
Qwen2.5-32B-Instruct	73.11	56.00	65.56	30.00	56.17
Qwen/QwQ-32B	61.22	46.00	84.44	20.00	52.92
DeepSeek-R1-Distill-Qwen-32B	65.48	55.00	84.44	24.67	57.40
Limo	63.44	45.00	61.11	15.33	46.22
S1-32B	66.81	53.00	84.44	24.00	57.06
Qwen2.5-14B-Instruct	67.44	59.00	57.78	26.67	52.72
DeepSeek-R1-Distill-Qwen-14B	63.27	44.00	84.44	21.00	53.18
DeepSeek-R1-Distill-Llama-8B	45.96	33.00	81.11	15.67	43.94
Meta-Llama-3-8B-Instruct	41.97	29.00	48.89	6.00	31.47
Llama-3.1-8B-Instruct	54.13	34.00	62.22	14.30	41.16
Qwen2.5-7B-Instruct	55.37	41.00	42.22	17.67	39.07
FinR1-7B	58.74	37.00	30.00	13.67	34.85
Fino1-8B	73.03	56.00	84.44	26.33	59.95
Fino1-14B	74.18	55.00	87.78	27.33	61.07

Resources on Hugging Face

Resource	Type	Link
Fino1 / Fin-o1 financial reasoning	Collection	TheFinAI/fino1-fin-o1-financial-reasoning
FinCoT (reasoning paths / training data)	Dataset	TheFinAI/FinCoT
FinQA (Fino1 reasoning-eval format)	Dataset	TheFinAI/en-finqa-reasoning
Fino1-8B	Model	TheFinAI/Fino1-8B
Fino1-14B	Model	TheFinAI/Fino1-14B
Fin-o1-8B	Model	TheFinAI/Fin-o1-8B
Fin-o1-14B	Model	TheFinAI/Fin-o1-14B
Open FinLLM Reasoning Leaderboard (FinReason)	Space	TheFinAI/open-finllm-reasoning-leaderboard

Citation

If you find our work useful, please cite our paper:

@misc{qian2025fino1transferabilityreasoningenhanced,
      title={Fino1: On the Transferability of Reasoning Enhanced LLMs to Finance}, 
      author={Lingfei Qian and Weipeng Zhou and Yan Wang and Xueqing Peng and Jimin Huang and Qianqian Xie},
      year={2025},
      eprint={2502.08127},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.08127}, 
}

License

The code in this repository is released under the MIT License. Datasets and models on Hugging Face keep their own licenses, stated on each card.

---

Built by The Fin AI · Hugging Face · GitHub