🤗 Training Data | 🤗 Training Data version 2 | 📄 Arxiv | 🤖 8B-Model 🤖 14B-Model 🏆 Leaderboard

Here, we utilized three evaluation datasets to assess the performance of our Fino1 model.

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

For the reasoning path building and training part, we were inspired by HuatuoGPT-o1

We release the reasoning path here: https://huggingface.co/datasets/TheFinAI/Fino1_Reasoning_Path_FinQA

Refer to HuatuoGPT-o1, we applied two-stage way to train our Fino1 model

• Stage 1: Supervised Fine-Tuning (SFT)

• Stage 2: Reinforcement Learning (RL)

We provide a simple PPO script using the trl library. Below is an example for training an 8B model with PPO on an 8-GPU A100 machine. Ensure you first download medical verifier as the reward model.

Please check HuatuoGPT-o1 for more training details.

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

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

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

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

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

BibTeX:

@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}, 
}