| --- |
| license: mit |
| --- |
| # Dataset Card for Text2Vis |
|
|
| ## 📝 Dataset Summary |
| **Text2Vis** is a benchmark dataset designed to evaluate the ability of large language models (LLMs) to generate accurate and high-quality data visualizations from natural language queries. Each example includes a natural language question, an associated data table, the expected answer, and ground-truth visualization code, along with rich metadata such as chart type, axis labels, domain, complexity, and question types. The dataset enables fine-grained evaluation through both **answer matching** and **chart quality metrics**. |
|
|
| The dataset enables **multi-dimensional evaluation** via: |
| - **Answer Matching** |
| - **Chart Quality Scores** |
| - **Reasoning Difficulty Levels** |
| - **Real-world vs Synthetic Data** |
| - **Question Answering** over structured tables |
|
|
| 🔗 [View paper on arXiv](https://arxiv.org/abs/2507.19969) |
|
|
| ## 💡 Use Cases |
| - Benchmarking LLMs for data visualization generation |
| - Question answering over structured table data |
| - Training models for text-to-visualization tasks |
| - Evaluating interpretability, correctness, and fidelity of generated visualizations |
|
|
| ## 📁 Dataset Structure |
| Each record includes the following key fields: |
|
|
| - **ID**: Unique identifier |
| - **Question**: Natural language query |
| - **Table Data**: Flattened version of the input data table in string format |
| - **Answer**: Ground-truth answer |
| - **Visualization Code**: Ground-truth Python/matplotlib code |
| - **Prompt**: Full prompt given to the model |
|
|
| Other metadata fields include: |
| - **Chart Type**, **X-Axis Label**, **Y-Axis Label**, **Summary** |
| - **High-Level Category**, **Low-Level Category**, **Domain** |
| - **Complexity**: (low/medium/high) |
| - **Tags**: `real_data`, `web_data`, `conversational`, `singlechart`, `multichart`, `open_ended`, `closed_ended`, `unanswerable` |
| - **Set**: |
| - The **test** split includes two subsets: |
| - A primary test set (test1) for general evaluation. |
| - A curated set of **236 stratified samples** (test2) used for **detailed manual evaluation** across chart types, domains, and reasoning difficulties. |
| |
|
|
| ## ✅ Evaluation |
|
|
| - For **question answering**, you can compare the predicted answer with the **Answer** field (answer match). |
| - For **visualization quality**, use the scoring prompts and scripts available in the GitHub repo below. |
|
|
| ## 🔗 GitHub |
|
|
| **Text2Vis GitHub Repository**: [vis-nlp/Text2Vis](https://github.com/vis-nlp/Text2Vis/tree/main) |
|
|
|
|
| ## 📚 Citation |
| If you use this dataset, please cite: |
|
|
| ```bibtex |
| @misc{rahman2025text2vischallengingdiversebenchmark, |
| title={Text2Vis: A Challenging and Diverse Benchmark for Generating Multimodal Visualizations from Text}, |
| author={Mizanur Rahman and Md Tahmid Rahman Laskar and Shafiq Joty and Enamul Hoque}, |
| year={2025}, |
| eprint={2507.19969}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.CL}, |
| url={https://arxiv.org/abs/2507.19969}, |
| } |
| |
| |
| |
| |
