README.md

metadata

license: other
task_categories:
  - image-to-text
  - table-to-text
language:
  - as
  - bn
  - gu
  - hi
  - kn
  - ml
  - or
  - pa
  - ta
  - te
  - ur
  - zh
  - en
tags:
  - table
  - html
  - image-to-html
  - table-structure
  - structure-only
  - SPRINT-MUSTARD

SPRINT-MUSTARD

1429 table image/HTML pairs. Test split only -- the source dataset publishes no train partition, so this is a benchmark, not a training set.

Ground truth is structure-only -- report TEDS-S, not TEDS

Every <td> in the HTML is empty. The annotation expresses the grid and the rowspan/colspan merge relationships and nothing else; the source carries no cell text.

TEDS compares cell text as well as structure. Against an all-empty ground truth it degenerates into measuring how well a model reproduces nothing, so TEDS is not meaningful on this dataset and only TEDS-S should be reported.

This is the opposite of TABLET, whose cells do carry values and where TEDS is the right metric.

What is in here

SPRINT-MUSTARD_html_image_pairs.zip contains a single test/ directory with 1429 pairs, one {language}_{n}.png and one {language}_{n}.html each. The language prefix prevents filename collisions across languages (hindi_1.png vs tamil_1.png).

Breakdown: 1215 pairs over 12 Indic languages, plus 214 scenetext and the chinese/english sets already present in the source.

Files

file	what
SPRINT-MUSTARD_html_image_pairs.zip	the 1429 image/HTML pairs (0.56 GB)
SPRINT-MUSTARD_html_image_pairs_说明.txt	the source dataset's own description, kept verbatim

Intended use

Evaluating table structure recognition on low-resource scripts, where cell text transcription is out of scope. Pair it with a text-bearing set such as TABLET if you need content accuracy.