mohamedwasef/egyptian-official-documents-ocr

Egyptian Tax Legislation OCR

Page-level OCR dataset of scanned Egyptian tax legislation (laws and ministerial decisions published in the Official Gazette, الجريدة الرسمية). Each row is one full page image with a structured transcription: typed layout elements, article (section) links, document metadata, and the full page text.

Built for fine-tuning vision-language models on full-page Arabic document OCR, and for structure-aware RAG over legal text.

Statistics

Split	Pages	Documents
train	336	19
test	30	10

• Documents: 29 (laws from 2005 to 2026 and Ministry of Finance decisions)
• Low-resolution pages (scan width < 1200px): 145
• Pages containing tables: 20
• Pages containing illegible markers [؟]: 1
• Splits are by document, so no document appears in both splits.
• The test split contains only real scans with quality == "ok".

Fields

Field	Description
image	Full page image at native scan resolution
image_name, doc_id, page_index, doc_pages	Page identity
doc_type, issuer, doc_number, doc_title, date	Document metadata, copied to every page of the document
content	Full page text, one element per line; tables as one-line HTML with rowspan/colspan
elements	JSON string: list of typed elements in reading order
quality	ok or low (scan width below 1200px)
origin	scan (real scan) or rendered_digital (rendered from a digital PDF, cleaner than real scans)
has_illegible	Page contains [؟] for unreadable text
review_notes	Spots the annotator flagged as uncertain
status	Review status

Element types: running_header, page_number, title, subtitle, article_heading, paragraph, table, field, signature, date, footnote. Each element has a section path linking it to its legal article, e.g. المادة الأولى > مادة (٤٢).

Document metadata

• doc_type: قانون or قرار. Presidential decrees with force of law (قرار رئيس الجمهورية بالقانون رقم ...) are typed قانون, since they are numbered and cited as laws.
• date: Gregorian issue date (YYYY-MM-DD) from the printed الموافق line; null when not printed (343 of 366 pages have one).
• Source of the metadata: transcribed from page 1 for 14 documents, extracted by rule from the transcribed page-1 text for 14 documents, and filled manually for 1 document. Every extracted or manual value was verified by script to appear verbatim in the page-1 text.

Transcription conventions

• Diplomatic spelling: printed forms are kept (e.g. فى, الآتى), never corrected.
• Digits are kept as printed (Arabic-Indic stay Arabic-Indic).
• Printed diacritics are kept; kashida/tatweel is removed.
• Printed line breaks inside a paragraph are joined; one element per paragraph or heading.
• Spacing is normalized: no space inside parentheses or before ، . : ؛.
• Tables: cells are ordered right to left (first cell = rightmost column). In elements, merged cells are objects with rowspan/colspan; in content, tables are rendered as one-line HTML.
• Page furniture: the running header always comes before the page number, wherever the number is printed on the page.
• Unreadable text is marked [؟] instead of guessed.

How it was made

1. PDFs were rendered to page images at their native scan resolution.
2. Pages were transcribed into structured JSON by an LLM (Claude, Anthropic) following fixed written guidelines, with uncertain spots recorded in review_notes.
3. A human spot-checked random pages from every document (status = "spot_checked"); not every page was fully reviewed.
4. All transcriptions were normalized to one convention by script, and missing document metadata was filled as described above.

Limitations

• Labels are LLM-generated with spot checks, not fully human-verified. Numbers are the most error-prone content.
• Low-resolution pages are harder to read; filter with quality == "ok" if needed.
• For OCR training, consider excluding pages where has_illegible is true.

Usage

import json
from datasets import load_dataset

ds = load_dataset("mohamedwasef/egyptian-official-documents-ocr")
page = ds["test"][0]
page["image"]                       # PIL image
print(page["content"])              # full page text
elements = json.loads(page["elements"])

train = ds["train"].filter(lambda x: not x["has_illegible"])

License

Annotations are released under CC BY 4.0. The source documents are official Egyptian government publications.