README.md

---
license: other
license_name: research-only
license_link: LICENSE
task_categories:
  - image-to-text
language:
  - zh
  - en
tags:
  - Table-Parsing
  - Table-Recognition
  - OCR
pretty_name: TableVerse-5K
size_categories:
  - 1K<n<10K
---

# TableVerse-5K <!-- omit in toc -->

**A Table-Parsing Benchmark for the StrucTab Framework**

<p align="center">
  <a href="https://github.com/VirtualLUOUCAS/StrucTab">GitHub Repo</a> •
  <a href="https://modelscope.cn/datasets/pspdada/TableVerse-5K">ModelScope Dataset</a> •
  <a href="https://arxiv.org/abs/2606.29905">Paper</a>
</p>

## Overview <!-- omit in toc -->

**TableVerse-5K** is the evaluation benchmark for **StrucTab**, a structured optimization framework for **table parsing**, the task of converting a table image into structured HTML. Each sample pairs a table image with an instruction prompt and a ground-truth HTML table, and models are scored with the TEDS / TEDS-S metrics.

<table align="center">
    <p align="center">
      <img src="docs/figures/introduction.jpg" width="80%" />
    </p>
</table>

The benchmark pipeline is illustrated below:

<table align="center">
    <p align="center">
      <img src="docs/figures/benchmark_pipeline.jpg" width="80%" />
    </p>
</table>

## Contents <!-- omit in toc -->

- [Statistics](#statistics)
- [Dataset Structure](#dataset-structure)
- [Data Format](#data-format)
- [Usage](#usage)
- [Citation](#citation)
- [License](#license)

## Statistics

| Item            | Details                                       |
| --------------- | --------------------------------------------- |
| Samples         | 5K table images                               |
| Task            | Table parsing (image → HTML table)            |
| Languages       | Bilingual (Chinese and English table content) |
| Output format   | HTML (`<table>...</table>`)                   |
| Scoring metrics | TEDS, TEDS-S                                  |

## Dataset Structure

```
data/
├── TableVerse_5K.jsonl     # annotations for all samples
└── images/                 # table images (*.jpg)
```

## Data Format

Each line of `TableVerse_5K.jsonl` is a JSON object:

```json
{
  "image_path": "images/xxx.jpg",
  "question": "You are an AI specialized in recognizing and extracting table from images...",
  "ref_answer": "<table>...</table>"
}
```

| Field        | Type   | Description                                                       |
| ------------ | ------ | ----------------------------------------------------------------- |
| `image_path` | string | Relative path from `data/`; also serves as the unique sample key  |
| `question`   | string | The instruction / prompt fed to the model together with the image |
| `ref_answer` | string | Ground-truth table in HTML (`<table>...</table>`)                 |

## Usage

Please refer to the [GitHub repository](https://github.com/VirtualLUOUCAS/StrucTab) for
the full inference and evaluation scripts.

```bash
# 1. Clone the code repository
git clone https://github.com/VirtualLUOUCAS/StrucTab
cd StrucTab/benchmark
pip install -r requirements.txt

# 2. Clone this dataset and place its contents under benchmark/data/
#    so that you have benchmark/data/TableVerse_5K.jsonl and benchmark/data/images/

# 3. Inference
python infer.py --api_type openai_compat --model_name <model> --base_url <url>

# 4. Score (requires the TEDS judging service, see the repo README)
python judge.py
```

## Citation

If you find TableVerse-5K useful, please consider citing (placeholder; to be updated):

```bibtex
@article{StrucTab_2026,
  title   = {{StrucTab}: A Structured Optimization Framework for Table Parsing},
  author  = {Li, Gengluo and Peng, Shangpin and Zhang, Chengquan and Wu, Binghong and Feng, Hao and Wang, Weinong and Lyu, Pengyuan and Shen, Huawen and Wan, Xingyu and Tian, Zhuotao and Hu, Han and Ma, Can and Zhou, Yu},
  journal = {arXiv preprint arXiv:2606.29905},
  year    = {2026}
}
```

## License

This dataset is released for **research purposes only**.