Datasets:

Tex-TAR
/

MMTAD

dataset_info:
  features:
    - name: image
      dtype:
        image:
          decode: false
    - name: annotation_json
      dtype: string
  splits:
    - name: test
      num_bytes: 231043093
      num_examples: 481
  download_size: 191751258
  dataset_size: 231043093
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

We introduce TexTAR, a multi-task, context-aware Transformer for Textual Attribute Recognition (TAR), capable of handling both positional cues (bold, italic) and visual cues (underline, strikeout) in noisy, multilingual document images.

MMTAD Dataset

MMTAD (Multilingual Multi-domain Textual Attribute Dataset) comprises 1,623 real-world document images—from legislative records and notices to textbooks and notary documents—captured under diverse lighting, layout, and noise conditions. It delivers 1,117,716 word-level annotations for two attribute groups:

T1: Bold,Italic,Bold & Italic
T2: Underline,Strikeout,Underline & Strikeout

Language & Domain Coverage

English, Spanish, and six South Asian languages
Distribution: 67.4 % Hindi, 8.2 % Telugu, 8.0 % Marathi, 5.9 % Punjabi, 5.4 % Bengali, 5.2 % Gujarati/Tamil/others
300–500 annotated words per image on average

To address class imbalance (e.g., fewer italic or strikeout samples), we apply context-aware augmentations:

Shear transforms to generate additional italics
Realistic, noisy underline and strikeout overlays

These augmentations preserve document context and mimic real-world distortions, ensuring a rich, balanced benchmark for textual attribute recognition.

More Information
For detailed documentation and resources, visit our website: TexTAR

Downloading the Dataset

from datasets import load_dataset

ds = load_dataset("Tex-TAR/MMTAD")
print(ds)

Dataset contains

textar-testset: document images
testset_labels.json: a JSON array or dict where each key/entry is an image filename and the value is its annotated attribute labels (bold, italic, underline, strikeout, etc. for each word)

Viewer Format

To power the Hugging-Face Data Studio we convert the original testset_labels.json into a line-delimited JSONL (hierarchical.jsonl) of the form:

{"image":"textar-testset/ncert-page_25.png",
 "annotation_json":"[{"bb_dim":[73,176,157,213],"bb_ids":[{"id":71120,"ocrv":"huge","attb":{"bold":false,"italic":false,"b_i":false,"no_bi":true,…}}]},…]"}

Citation

Please use the following BibTeX entry for citation .

@article{Kumar2025TexTAR,
  title   = {TexTAR: Textual Attribute Recognition in Multi-domain and Multi-lingual Document Images},
  author  = {Rohan Kumar and Jyothi Swaroopa Jinka and Ravi Kiran Sarvadevabhatla},
  booktitle = {International Conference on Document Analysis and Recognition,
            {ICDAR}},
  year    = {2025}
}