VisCoder2: Building Multi-Language Visualization Coding Agents

🌐 Project Page | 📖 arXiv | 🤗 VisCode-Multi-679K | 🤗 VisPlotBench | 🤗 VisCoder2

---

🔔 News

• 🔥 [2025-10-25] VisCode-Multi-679K, VisPlotBench and VisCode2 models are now publicly released! Check out our paper and Huggingface collections.

🧠 Introduction

VisCoder2 is an open-source family of multi-language visualization coding agents capable of iteratively generating, executing, rendering, and self-debugging visualization code.

This work addresses core challenges where existing models fail:

• Limited language coverage
• Unreliable code execution
• Lack of iterative correction mechanisms

Unlike general code generation, visualization requires grounding across natural language, code, and rendered visual outputs.

To enable this, we introduce three complementary resources:

1. VisCode-Multi-679K:
   A large-scale supervised dataset with 679K executable visualization samples and multi-turn correction dialogues across 12 programming languages, including Python, Vega-Lite, LaTeX, Mermaid, LilyPond, and more.
2. VisPlotBench:
   A new benchmark spanning 8 languages and 13 visual categories, designed to systematically evaluate both initial code generation and multi-round self-debug capabilities.
3. VisCoder2:
   The family of multi-language visualization models trained on VisCode-Multi-679K.

---

📊 Main Results on VisPlotBench

We evaluate VisCoder2 on VisPlotBench, our new benchmark for executable visualization code generation across 8 diverse languages.
The primary metric is Execution Pass Rate, which measures whether the code runs without error and produces a valid visual.

With iterative self-debug, VisCoder2-32B achieves an 82.4% overall execution pass rate, matching the performance of GPT-4.1 and significantly outperforming all open-source baselines.

---

🛠️ Training & Evaluation

We provide both the training dataset and evaluation benchmark for VisCoder2.

• 📦 Training is performed using the ms-swift framework with full-parameter supervised fine-tuning on our new VisCode-Multi-679K dataset.
• 📊 Evaluation is based on VisPlotBench, using a standardized execute–render–score pipeline that assesses models across 8 languages.
  This includes a self-debug evaluation mode that allows models to revise failed generations over multiple rounds.

See the following folders for details:

• train/ : Training scripts and configurations based on ms-swift
• VisPlotBench/ : Evaluation framework for VisPlotBench

---

📬 Contact

• Yuansheng Ni: [email protected]
• Wenhu Chen: [email protected]

---

📖 Citation

BibTeX:

@article{ni2025viscoder2,
  title={VisCoder2: Building Multi-Language Visualization Coding Agents},
  author={Ni, Yuansheng and Cai, Songcheng and Chen, Xiangchao and Liang, Jiarong and Lyu, Zhiheng and Deng, Jiaqi and Zou, Kai and Nie, Ping and Yuan, Fei and Yue, Xiang and others},
  journal={arXiv preprint arXiv:2510.23642},
  year={2025}
}

@article{ni2025viscoder,
  title={VisCoder: Fine-Tuning LLMs for Executable Python Visualization Code Generation},
  author={Ni, Yuansheng and Nie, Ping and Zou, Kai and Yue, Xiang and Chen, Wenhu},
  journal={arXiv preprint arXiv:2506.03930},
  year={2025}
}