---
license: cc-by-4.0
task_categories:
- image-to-text
language:
- en
tags:
- table-structure-recognition
- table-recognition
- html
- document-understanding
pretty_name: CompTSRBench
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: id
    dtype: string
  - name: split
    dtype: string
  - name: pattern
    dtype: int64
  - name: image
    dtype: image
  - name: html
    dtype: string
  splits:
  - name: train
    num_examples: 8000
  - name: test
    num_examples: 2000
---

# CompTSRBench

CompTSRBench is a benchmark of 10,000 structurally complex synthetic tables for table structure recognition (image → HTML).
Each table pairs a rendered image with its ground-truth HTML. The tables cover 25 canonical structural patterns
(hierarchical headers, row and column spans, gridless regions and more) across application domains such as finance,
science, education and healthcare.

| Split | Tables |
|---|---|
| train | 8,000 |
| test | 2,000 |

Every pattern appears in both splits, with at least 40 test tables per pattern.

## Fields

| Field | Type | Description |
|---|---|---|
| `id` | string | Table identifier |
| `split` | string | `train` or `test` |
| `pattern` | int | Canonical structural pattern (1–25) |
| `image` | image | Rendered table image |
| `html` | string | Ground-truth table HTML (minified, no styling); headers are marked with `<th>` |

## Structural statistics (all 10K tables)

- 81.4% contain column spans, 89.5% row spans, 58.4% multi-level headers
- On average 17.3 merged cells per table

## Usage

```python
from datasets import load_dataset

ds = load_dataset("researchgroup542/Table_Dataset")
example = ds["test"][0]
example["image"]   # PIL image
example["html"]    # ground-truth HTML
```

## Evaluation

We report TEDS, TEDS-Struct and the Triple Metric Score (TMS), a structure-first metric that checks each cell's position,
header context and content in that order. See the paper for details.

## License

CC BY 4.0. All table content is synthetically generated and does not refer to real individuals or organizations.