Lukaszl/clearocr-invoice-document-ai

clearOCR Invoice Document AI Dataset

This dataset shows a complete invoice document AI workflow built around clearOCR.

It contains 423 high-confidence invoice examples with:

• original invoice images,
• OCR text generated by clearOCR,
• Markdown reconstruction of the document,
• structured invoice JSON generated by a local fine-tuned extraction model,
• visual verification metadata.

The dataset demonstrates how clearOCR can serve as the OCR layer in an invoice automation pipeline where document images are converted into text, document-like representations and structured data.

clearOCR is an OCR API for PDF, JPG and PNG documents, focused on business documents in Polish and English.

Dataset Overview

For each invoice, the release provides aligned representations of the same document:

invoice image
  -> clearOCR text
  -> Markdown reconstruction
  -> structured invoice JSON
  -> visual verification metadata

The public files are organized so that metadata.jsonl is only a compact file map. OCR text, Markdown, extracted JSON and judge payloads are stored as separate files and referenced by path.

What Is Included

Each record contains:

• image: original invoice image loaded by Hugging Face imagefolder,
• clearocr_text_path: path to clearOCR text output,
• markdown_path: path to Markdown reconstruction,
• invoice_json_path: path to extracted invoice JSON,
• judges_path: path to visual verification payloads.

The JSON extraction model is described only at a high level as a local fine-tuned extraction model. The focus of this release is the document pipeline and the clearOCR OCR output.

Selection Criteria

This public release contains only high-confidence examples selected for publication.

An invoice was included when:

• clearOCR output was available,
• Markdown reconstruction was available,
• extracted JSON was available,
• Qwen returned pass for image vs Markdown,
• Qwen returned pass for image vs JSON,
• Gemini returned pass for image vs Markdown,
• Gemini returned pass for image vs JSON.

This keeps the public dataset focused on clean, usable examples selected for publication.

Splits

The dataset contains 423 invoices:

split	records
train	359
validation	47
test	17

Example Record

A typical metadata row is compact and points to the larger artifacts:

{
  "file_name": "images/00000.jpg",
  "source_id": "test:0",
  "clearocr_text_path": "ocr/00000.ocr",
  "markdown_path": "markdown/00000.mdoc",
  "invoice_json_path": "json/00000.invoice.payload",
  "judges_path": "judges/00000.judges.payload"
}

For a fuller example showing one invoice as clearOCR text, Markdown and extracted JSON, see:

• docs/EXAMPLE_RECORD.md

Loading

import json
from pathlib import Path
from datasets import load_dataset

ds = load_dataset("Lukaszl/clearocr-invoice-document-ai")
row = ds["train"][0]

image = row["image"]
source_id = row["source_id"]

When working from a local clone of the repository, use the path fields to load the companion artifacts:

root = Path("clearocr-invoice-document-ai")
split = "train"

ocr_text = (root / split / row["clearocr_text_path"]).read_text(encoding="utf-8")
markdown = (root / split / row["markdown_path"]).read_text(encoding="utf-8")
invoice_json = json.loads((root / split / row["invoice_json_path"]).read_text(encoding="utf-8"))

Dataset Structure

The repository uses a readable split layout:

train/
  images/
  ocr/
  markdown/
  json/
  judges/
  metadata.jsonl
validation/
  images/
  ocr/
  markdown/
  json/
  judges/
  metadata.jsonl
test/
  images/
  ocr/
  markdown/
  json/
  judges/
  metadata.jsonl
docs/
  ATTRIBUTION.md
  SCHEMA.md
  EXAMPLE_RECORD.md

The split-level metadata.jsonl files are intentionally compact. Large text and JSON artifacts are stored as separate files to make the dataset easier to inspect in the Hugging Face UI and in a local checkout.

Invoice JSON Shape

The extracted JSON follows this general structure:

{
  "header": {
    "invoice_no": "...",
    "invoice_date": "...",
    "seller": "...",
    "client": "...",
    "seller_tax_id": "...",
    "client_tax_id": "...",
    "iban": "..."
  },
  "items": [
    {
      "item_desc": "...",
      "item_qty": "...",
      "item_net_price": "...",
      "item_net_worth": "...",
      "item_vat": "...",
      "item_gross_worth": "..."
    }
  ],
  "summary": {
    "total_net_worth": "...",
    "total_vat": "...",
    "total_gross_worth": "..."
  }
}

Source And License

The invoice images originate from:

Kozlowski, Marek; Weichbroth, Pawel (2021), "Samples of electronic invoices", Mendeley Data, V2, doi: 10.17632/tnj49gpmtz.2

Original source:

• https://data.mendeley.com/datasets/tnj49gpmtz/2

Source license: CC BY 4.0.
This derived dataset is also released under CC BY 4.0.

Recommended Use

This dataset is useful for:

• evaluating invoice OCR workflows,
• testing OCR-to-structured-data pipelines,
• comparing document representations such as OCR text, Markdown and JSON,
• building demos around invoice document AI,
• inspecting how OCR output can be used by downstream extraction systems.

Notes

The extracted JSON and visual verification metadata are model-generated artifacts. They are included to demonstrate and inspect a practical document AI pipeline, not as legal or accounting certification.

If you use this dataset, please cite the original Mendeley dataset and mention that the derived OCR artifacts were prepared with clearOCR.