Datasets:

IslamMesabah
/

AraReceipt

license: mit
language:
  - ar
task_categories:
  - object-detection
  - image-to-text
tags:
  - document-understanding
  - key-information-extraction
  - receipts
  - ocr
  - arabic
size_categories:
  - n<1K

AraReceipt

AraReceipt is a manually annotated dataset of 100 Arabic retail receipt images labeled with 25 key-information classes plus an Ignore category, in WildReceipt style. It contains 4,609 annotated regions (46.1 ± 21.6 per image), each with a box, a transcription, and a semantic class.

The dataset was built with GAIDA, a human-in-the-loop annotation system that combines OCR-based region extraction, LLM-based semantic pre-annotation, and human validation in Label Studio. Producing AraReceipt required no changes to GAIDA's architecture — only OCR settings for Arabic and mixed-script text, prompt revisions for right-to-left writing, and language-specific field expressions. What's released here is the final result of that process: reviewed, human-corrected annotations only — no model predictions, scores, or annotation-session metadata.

Images are derived from ReceiptSense; AraReceipt is released under the MIT license, matching the terms of the source data.

Loading

from datasets import load_dataset

ds = load_dataset("IslamMesabah/AraReceipt", split="train")
print(ds[0]["id"], len(ds[0]["annotations"]))
print(ds[0]["annotations"][0])

Fields

Each row is one receipt image, named images/0001.jpg through images/0100.jpg.

Field	Description
`image`	The receipt image.
`file_name`	Path of the image inside the dataset.
`id`	Sequential dataset id (1-100), matching the image file name.
`width`, `height`	Image size in pixels.
`annotations`	The annotated regions (see below).

Each entry of annotations has:

Field	Description
`box`	Quadrilateral as`[x1, y1, x2, y2, x3, y3, x4, y4]` in absolute pixels, corners ordered top-left, top-right, bottom-right, bottom-left.
`text`	Transcription of the region.
`label`	Numeric class id; see`class_list.txt` for the id-to-name mapping (same 26 classes as WildReceipt: 25 key-information classes plus `Ignore`).

Examples

Each region is boxed and labeled with its label class name, colored per class. These are the final, human-reviewed annotations.

AraReceipt example

Class distribution

Class	Regions	Class	Regions
Others	1598	Store_addr_value	88
Prod_price_value	466	Date_key	61
Prod_quantity_value	461	Tax_value	45
Prod_item_value	402	Tel_key	40
Subtotal_value	225	Store_addr_key	39
Total_value	160	Time_key	30
Total_key	127	Tax_key	21
Date_value	113	Tips_key	10
Subtotal_key	110	Tips_value	9
Store_name_value	109	Ignore	7
Prod_price_key	101	Store_name_key	1
Time_value	100
Prod_item_key	99
Prod_quantity_key	97
Tel_value	90

Slightly over half of the regions contain Arabic script and 43.6% are purely numeric (prices, quantities, totals), reflecting the mixed-script character of Arabic retail receipts. Key-type classes are markedly rarer than their value counterparts, because Arabic receipts often print values without field labels.

Annotation process

One native Arabic speaker with professional English proficiency annotated all 100 images using GAIDA: PaddleOCR proposed regions and transcriptions, Llama 4 Maverick proposed a class for each region, and the annotator reviewed, corrected, and approved every region before export.

The end-to-end assisted annotation time averaged 5:51 per image (~10 hours in total).

Limitations

The dataset was annotated by a single annotator without an independent second pass, so no inter-annotator agreement is available. Region boundaries originate from OCR detections and were corrected rather than drawn from scratch, which may bias them toward what the detector proposed.

Citation

@inproceedings{mesabah2026gaida,
  title     = {GAIDA: Generative AI for Human-in-the-Loop Document Annotation},
  author    = {Mesabah, Islam and Elhosary, Ali and Sandulu, Priyabanta and
               Obradovic, Darko and Vollmer, Sebastian J.},
  year      = {2026}
}