---
license: other
task_categories: [image-to-text, table-structure-recognition]
language:
- en
- zh
tags:
- table
- html
- image-to-html
- table-structure
- TABLET
---

# TABLET

18642 table image/HTML pairs (17733 train, 909 test) synthesised from 9 table-question-answering source sets.

## What is in here

`TABLET_html_image_pairs.zip` contains 18642 image/HTML pairs under `train/` and
`test/`, one `437b41c90b28ac41e73230c221a05c1e.png` and one `437b41c90b28ac41e73230c221a05c1e.html` each. The HTML is the table
annotation; **cells carry values**, so this is a content-and-structure
annotation, not a structure-only one.

Source sets, and how many samples each contributed:

- TabFact: 12231
- TABMWP: 2381
- ToTTo: 1575
- WikiTableQuestions: 1146
- TAT-QA: 475
- WikiBIO: 363
- StatCan: 258
- InfoTabs: 187
- NSF: 26

PubTabNet is excluded on purpose: several table-recognition training mixes
already contain it, and a benchmark that overlaps the training data does not
measure generalisation.

## Files

| file | what |
|---|---|
| `TABLET_html_image_pairs.zip` | the pairs (1.72 GB) |
| `TABLET_meta.jsonl` | one line per sample: `split`, `name`, `seed`, `has_cell_values` |
| `TABLET_stats.json` | counts and the seed breakdown |
| `TABLET_样本预览.html` | sample preview |

## Splits

| split | n |
|---|---|
| train | 17733 |
| test | 909 |

## Intended use

Table structure recognition and image-to-HTML generation. Score with TEDS when
the annotation carries cell text; a structure-only metric is not meaningful here
because the cells are not empty.