---
base_model: Qwen/Qwen2.5-VL-3B-Instruct
library_name: peft
pipeline_tag: image-text-to-text
license: apache-2.0
tags:
- finance
- lora
- vision-language
- qwen
---

# PyFi-QwenVL-3B-47K

This model is a fine-tuned LoRA adapter for [Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct), specialized for financial image understanding. It was introduced as part of the PyFi framework.

## Model Details

- **Developed by:** Yuqun Zhang, Yuxuan Zhao, Sijia Chen (AgenticFinLab)
- **Model type:** Vision-Language Model (LoRA Adapter)
- **Base Model:** [Qwen/Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct)
- **Language(s):** English
- **License:** Apache 2.0
- **Paper:** [PyFi: Toward Pyramid-like Financial Image Understanding for VLMs via Adversarial Agents](https://arxiv.org/abs/2512.14735)
- **Repository:** [https://github.com/AgenticFinLab/PyFi](https://github.com/AgenticFinLab/PyFi)

## Summary

PyFi (Pyramid-like Financial Image Understanding) is a framework designed to enhance Visual Language Models (VLMs) in understanding complex financial images through adversarial agents. The framework enables VLMs to reason through question chains in a progressive, simple-to-complex manner across six hierarchical capability levels:
1. **Perception**: Basic visual understanding
2. **Data Extraction**: Foundational information retrieval
3. **Calculation Analysis**: Numerical analysis tasks
4. **Pattern Recognition**: Identifying trends and patterns
5. **Logical Reasoning**: Complex logical analysis
6. **Decision Support**: Strategic decision-making assistance

This specific checkpoint was fine-tuned on approximately 47,000 question-answer chains from the **PyFi-600K** dataset. In this version ("w/o CoT"), only the question and answer from the final sample in each chain were used during training to target the ultimate reasoning goal.

## Training Details

The model was fine-tuned with the following configuration:
- **Optimizer**: AdamW
- **Learning Rate**: $1.0 \times 10^{-4}$
- **Learning Rate Schedule**: Cosine scheduling with warmup ratio of 0.1
- **Training Epochs**: 1
- **Batch Size**: Effective batch size of 8
- **PEFT**: LoRA with full-module adaptation
- **Hardware**: 4x NVIDIA RTX 5090 GPUs

## Citation

If you use PyFi in your research, please cite:

```bibtex
@article{pyfi2025,
  title={PyFi: Toward Pyramid-like Financial Image Understanding for VLMs via Adversarial Agents},
  author={Zhang, Yuqun and Zhao, Yuxuan and Chen, Sijia},
  journal={arXiv preprint arXiv:2512.14735},
  year={2025}
}
```

## Contact

For questions or inquiries, please contact:
- **Email**: Yuqun Zhang ([research@yuqunzhang.com](mailto:research@yuqunzhang.com)) or Yuxuan Zhao ([yx.zhao129@gmail.com](mailto:yx.zhao129@gmail.com))
- **GitHub Issues**: [https://github.com/AgenticFinLab/PyFi/issues](https://github.com/AgenticFinLab/PyFi/issues)