Digitized Receipts Dataset — OCR Documents
Dataset composed of scanned receipts from business trips, available in PDF format. Primarily used to train OCR systems or test the extraction of textual data on real documents.
1255 PDF files, receipts scanned with partially OCrised text
CC0: Public Domain
Description
This dataset contains 1255 scanned receipts in PDF format, collected during business trips. The majority of the documents were scanned using an OCR tool, allowing to have files with embedded or recognizable text. This corpus is useful for experiments around the automatic extraction of text from real documents (OCR).
What is this dataset for?
- Train OCR models to detect and read commercial receipts
- Test entity extraction algorithms (amount, date, supplier)
- Develop solutions for processing semi-structured PDF documents
Can it be enriched or improved?
Yes, it is possible to segment documents according to their type or language, or to annotate them manually (for example, amount fields, taxes, etc.). Adding raw OCR versions (extracted TXT files) would improve their value for supervised training.
🔎 In summary
🧠 Recommended for
- OCR developers
- Documentary NLP researchers
- Automated accounting tools
🔧 Compatible tools
- Tesseract
- LayoutLM
- DocTR
- PaddleOCR
💡 Tip
Combine PDFs with an automatic OCR step (Tesseract or other) to create text-annotated corpora.
Frequently Asked Questions
Are the documents already OCrised?
Partly yes, some receipts contain OCrised text. For others, an external OCR can be applied.
Can this dataset be used to train a document classification model?
Yes, by adding labels (type of receipt, country, supplier), it is possible to make it a structured training game.
What output format is recommended after processing?
It is recommended to extract the OCrisé text in JSON or CSV format, with standardized fields such as date, total, VAT, etc.




