| | --- |
| | base_model: |
| | - Qwen/Qwen2.5-Coder-3B-Instruct |
| | datasets: |
| | - TIGER-Lab/VisCode-200K |
| | language: |
| | - en |
| | license: apache-2.0 |
| | tags: |
| | - code |
| | library_name: transformers |
| | pipeline_tag: text-generation |
| | --- |
| | |
| | # VisCoder-3B |
| |
|
| | [🏠 Project Page](https://tiger-ai-lab.github.io/VisCoder) | [📖 Paper](https://arxiv.org/abs/2506.03930) | [💻 GitHub](https://github.com/TIGER-AI-Lab/VisCoder) | [🤗 VisCode-200K](https://huggingface.co/datasets/TIGER-Lab/VisCode-200K) | [🤗 VisCoder-7B](https://huggingface.co/TIGER-Lab/VisCoder-7B) |
| |
|
| | **VisCoder-3B** is a lightweight language model fine-tuned for **Python visualization code generation and iterative correction**. It is trained on **VisCode-200K**, a large-scale instruction-tuning dataset that integrates natural language instructions, validated Python code, and execution-guided revision supervision. |
| |
|
| | ## 🧠 Model Description |
| |
|
| | **VisCoder-3B** is trained on **VisCode-200K**, a large-scale instruction-tuning dataset tailored for executable Python visualization tasks. It addresses a core challenge in data analysis: generating Python code that not only executes successfully but also produces **semantically meaningful plots** by aligning **natural language instructions**, **data structures**, and **visual outputs**. |
| |
|
| | We propose a **self-debug evaluation protocol** that simulates real-world developer workflows. In this setting, models are allowed to revise previously failed generations over multiple rounds with guidance from **execution feedback**. |
| |
|
| | ## 📊 Main Results on PandasPlotBench |
| |
|
| | We evaluate VisCoder-3B on [**PandasPlotBench**](https://github.com/TIGER-AI-Lab/VisCoder/tree/main/eval), which tests executable visualization code generation across **Matplotlib**, **Seaborn**, and **Plotly**. Evaluation includes both standard generation and **multi-turn self-debugging** |
| |
|
| |  |
| |
|
| | > VisCoder-3B outperforms existing open-source baselines on multiple libraries and shows consistent recovery improvements under the self-debug protocol. |
| |
|
| | ## 📁 Training Details |
| |
|
| | - **Base model**: Qwen2.5-Coder-3B-Instruct |
| | - **Framework**: [ms-swift](https://github.com/modelscope/swift) |
| | - **Tuning method**: Full-parameter supervised fine-tuning (SFT) |
| | - **Dataset**: [VisCode-200K](https://huggingface.co/datasets/TIGER-Lab/VisCode-200K), which includes: |
| | - 150K+ validated Python visualization samples with corresponding images |
| | - 45K+ multi-turn correction dialogues guided by execution results |
| |
|
| | ## 📖 Citation |
| |
|
| | If you use VisCoder-3B or VisCode-200K in your research, please cite: |
| |
|
| | ```bibtex |
| | @misc{ni2025viscoderfinetuningllmsexecutable, |
| | title={VisCoder: Fine-Tuning LLMs for Executable Python Visualization Code Generation}, |
| | author={Yuansheng Ni and Ping Nie and Kai Zou and Xiang Yue and Wenhu Chen}, |
| | year={2025}, |
| | eprint={2506.03930}, |
| | archivePrefix={arXiv}, |
| | primaryClass={cs.SE}, |
| | url={https://arxiv.org/abs/2506.03930} |
| | } |
| | |
| | For evaluation scripts and more information, see our [GitHub repository](https://github.com/TIGER-AI-Lab/VisCoder). |
