Bayanat Labs Arabic Form Grounding

Arabic visual grounding synthetic forms section-aware extraction semantic field areas layout holdout controlled abstention

Key numbers

Counted by script from the released files. The SHA-256 identifies the reviewed canonical file manifest for the frozen dataset payload. Public SHA256SUMS.txt lists file checksums. Key numbers are recomputed from the exact copied summary JSON, with its own SHA-256.

Downloads and identifiers

Get started

Load the frozen dataset

from datasets import load_dataset
import json

forms = load_dataset(
    'Bayanat-Labs/arabic-form-grounding', 'forms',
    split='test', revision='v0.1.0',
)
questions = load_dataset(
    'Bayanat-Labs/arabic-form-grounding', 'questions',
    split='test', revision='v0.1.0',
)
form = forms[0]
image = form['image']  # Embedded PNG decoded as a PIL image.
elements = json.loads(form['elements_json'])
# Join questions by form_id. Use code/contract.py to exclude gold metadata.

What is released

Attribution: Bayanat Labs, Arabic Form Grounding, v0.1.0. https://www.bayanatlabs.com/open/arabic-form-grounding

Intended uses

• Test field association when the same label appears under different visible section headings.
• Evaluate complete semantic field areas, center-based relations and missing/ambiguous response handling.
• Reproduce matched digit and heading-counterfactual checks with a public scorer and transparent given-layout baselines.

Out of scope

• Treating the generated forms as real customer records or independently human-annotated ground truth.
• Using semantic field rectangles as tight word, token, line or glyph localization annotations.
• Inferring OCR accuracy, natural-form realism, general document reliability or training benefit from these fixtures.

Rights at a glance

Training, display, redistribution and synthetic-voice use are separate permissions.

Limitations

• All content uses controlled synthetic language, recurring field labels and a shared section schema. The six domains vary titles, headings, option labels and values, rather than providing six unrelated workflow schemas.
• The 1,152 images include repeated Eastern-digit and section-heading counterfactual views. Counts of distinct base contexts do not imply statistical independence; grammar, schema and field-layout components recur.
• Complete layout families are held out for validation and test. This is a scoped section-arrangement holdout; domains and field-arrangement patterns recur across splits.
• Missing and global repeated-label ambiguity controls are regular by construction. All gold is public, and the generator is available. There is no handwriting, scan degradation, native-speaker review or expert workflow validation.
• A rules baseline that uses section containment outperforms label matching on the held-out test. Both receive exact text, rectangles and checkbox states, and neither solves the spatial relations. These are not OCR or vision-model scores.

Sources and licences

Bayanat Labs authored the fictional text, values, queries and schemas; no Atlas images, customer forms or third-party benchmark examples are included. Pillow with RAQM/HarfBuzz renders the semantic areas. Unmodified Noto font files retain their original copyright, OFL notices and hashes. Software validation checks visible question/layout consistency, font coverage, image hashes and paired relations; it is not human annotation certification. English documentation and interface are the default, with an explicit Arabic question toggle and optional Arabic company page. Source details

How to cite

Cite as:

Bayanat Labs (2026). Arabic Form Grounding, v0.1.0. Bayanat Labs. https://www.bayanatlabs.com/open/arabic-form-grounding

• CITATION.cff

No DOI. Cite the URL in the citation above and the date you read it.

@misc{bayanatlabs2026arabicformgrounding,
  author       = {{Bayanat Labs}},
  title        = {Arabic Form Grounding, v0.1.0},
  year         = {2026},
  month        = oct,
  howpublished = {\url{https://www.bayanatlabs.com/open/arabic-form-grounding}}
}

How we handle citations

Changelog

Corrections and removal requests

If something in this release should not be there, or is wrong, email hello@bayanatlabs.com. We handle removal requests within 30 days, and platform requests within 24 hours. How removal requests work

Related product: Bayanat Documents