Dataset Card for SynthTabNet_OTSL

Dataset Description

Overview

SynthTabNet_OTSL is a converted version of the original SynthTabNet dataset, using the OTSL format proposed in our paper. It includes the original annotations plus newly added content. SynthTabNet is divided into four parts, each containing 150 k tables (600 k total). Each part is categorized by table size, structure, style, and content, and split into train, test, and validation sets.

Structure

• cells: Ground‑truth cell content from the original dataset.
• otsl: New simplified table‑structure token format.
• html: Ground‑truth HTML structure from the original dataset.
• html_restored: HTML generated from OTSL.
• cols: Number of columns.
• rows: Number of rows.
• image: PIL image.

OTSL Vocabulary

OTSL: New simplified table‑structure token format. More information can be found in our paper. The dataset extends the work presented in the paper with minor modifications:

• "fcel" – cell with content
• "ecel" – empty cell
• "lcel" – cell looking left (handles horizontal merges)
• "ucel" – cell looking up (handles vertical merges)
• "xcel" – 2‑D spanning cell, covering the whole merged area in this dataset
• "nl" – new line token

Data Splits

The dataset provides three splits:

• train
• val
• test

Additional Information

Curators

The dataset was converted by IBM Research's Deep Search team.

Curators:

• Maksym Lysak, @maxmnemonic
• Ahmed Nassar, @nassarofficial
• Christoph Auer, @cau‑git
• Nikos Livathinos, @nikos‑livathinos
• Peter Staar, @PeterStaar‑IBM

Citation

@misc{lysak2023optimized,
      title={Optimized Table Tokenization for Table Structure Recognition}, 
      author={Maksym Lysak and Ahmed Nassar and Nikolaos Livathinos and Christoph Auer and Peter Staar},
      year={2023},
      eprint={2305.03393},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}