cdek-ocr/receipt-ocr-ru

Russian Receipts OCR (Semantic)

Overview

This repository contains a dataset of Russian retail receipts with semantic OCR annotations.

⚠️ Important note
This repository is used primarily as a data storage and distribution location.
The dataset is consumed by external inference, evaluation and training pipelines and is not intended to be loaded directly via datasets.load_dataset() without a custom loader.

---

Description

Each sample consists of:

• a receipt image
• a corresponding JSON annotation with structured OCR information

The dataset is designed for:

• OCR and document understanding
• semantic information extraction
• evaluation of receipt parsing pipelines

---

Dataset structure

The data is organized as a simple and explicit file-based structure:

images/ ├── train/ ├── validation/ └── test/

annotations/ ├── train/ ├── validation/ └── test/

For each image file: images//.jpg

there exists a corresponding annotation file: annotations//.json

Annotations format

Each annotation JSON contains structured information extracted from a receipt, including:

• seller
• inn
• date
• total
• items (list of line items, if available)

Fields may be missing and are represented as null.

Bounding boxes are taken directly from OCR output and are not manually corrected.

---

Completeness rule

A receipt is considered complete if:

• total is present
• date is present
• either seller or inn is present

Usage notes

This dataset is intentionally stored in a file-based format and is consumed by custom pipelines that:

• explicitly load images from disk
• load annotation JSON files by filename
• perform OCR and semantic parsing separately

This design ensures:

• transparent data loading
• reproducible evaluation
• full control over preprocessing and inference steps

---

Related resources

• OCR model checkpoint:
  https://huggingface.co/SvetaLana25/dek-receipt-donut-stage1

• Evaluation and inference code:
  Provided as part of the project submission.

---

Disclaimer

This dataset is provided for educational and research purposes.

Sensitive information is anonymized or originates from user-provided examples.