metadata

dataset_info:
  features:
    - name: id
      dtype: string
    - name: image
      dtype: image
    - name: text_output
      dtype: string
  splits:
    - name: train
      num_bytes: 28126563
      num_examples: 149
    - name: test
      num_bytes: 10195323
      num_examples: 50
  download_size: 36491665
  dataset_size: 38321886
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
license: mit
task_categories:
  - question-answering
language:
  - en
size_categories:
  - n<1K

Dataset Card for FUNSD (JSON Format)

This dataset contains the preprocessed JSON format of the FUNSD dataset, designed for form understanding in noisy scanned documents. It includes the original structure with annotations in JSON format, as per the original FUNSD dataset. This dataset is intended for document understanding tasks such as OCR, layout parsing, and key-value extraction.

Dataset Details

Dataset Description

The FUNSD (Form Understanding in Noisy Scanned Documents) dataset is a collection of noisy scanned document images, designed for OCR and form recognition tasks. This version contains the original JSON annotations for training and testing models in form understanding tasks.

Curated by: David (Tan) LE
Language(s): English (labels, tags)
License: Apache License 2.0 (from the original FUNSD dataset)

Dataset Sources

Repository: Original FUNSD Dataset Repository
Paper: Jaume et al., 2019

Uses

Direct Use

This dataset can be used for:

Fine-tuning document understanding models such as Donut, LayoutLM, and other OCR-based models.
Training and evaluating models for form extraction (questions, answers, headers).
Research in key-value extraction and layout parsing.

Dataset Structure

id: Unique identifier for each document.
image: The document image (RGB format).
text_output: JSON annotations containing:
- "form": A list of labeled fields (e.g., "question", "answer", "header").
- "words": Word-level annotations with bounding boxes.
- "linking": Relationships (e.g., links between "question" and "answer").

Splits

Train: 149 examples
Test: 50 examples

Dataset Creation

Curation Rationale

This version retains the original JSON format of the FUNSD dataset, making it easy to use for models designed for form understanding tasks without the need to convert to XML or TEI formats.

Source Data

The dataset is sourced from the original FUNSD repository:

Images: Scanned document images.
Annotations: Original annotations in JSON format.

Data Collection and Processing

The dataset was not altered apart from formatting to maintain compatibility with the Hugging Face datasets library.

Annotations

Annotation Process

The annotations in JSON format contain:

Bounding boxes for each form element.
Labels for "question", "answer", "header", and "other".
Linking information to capture relationships between form elements.

Bias, Risks, and Limitations

Risks and Limitations

The dataset contains noisy, real-world scanned documents that may not generalize to synthetic or clean data.
Some form layouts may not follow typical Western document standards, which could bias model training.

Citation

If you use this dataset, please cite the original paper:

BibTeX:

@article{Jaume2019FUNSDAD,
  title={FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents},
  author={Guillaume Jaume and H. K. Ekenel and J. Thiran},
  journal={2019 International Conference on Document Analysis and Recognition Workshops (ICDARW)},
  year={2019},
  volume={2},
  pages={1-6}
}

APA: Jaume, G., Ekenel, H. K., & Thiran, J. (2019). FUNSD: A dataset for form understanding in noisy scanned documents. 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW), 1–6.

Dataset Card Authors

David (Tan) LE - GitHub Profile

Dataset Card Contact

For any issues or questions about this dataset, contact me through the HuggingFace message system.