|
|
| --- |
| language: multilingual |
| license: apache-2.0 |
| tags: |
| - invoice |
| - key-information-extraction |
| - kie |
| - document-understanding |
| - ocr |
| - visual-annotation |
| - checkmarks |
| - layout-analysis |
| - ukrainian |
| - chinese |
| - swedish |
| size_categories: |
| - 600 images |
| --- |
| |
| # Invoice Checkmark Annotations |
|
|
| **Multilingual dataset of real invoices with human-drawn visual checkmarks/circles indicating verified key fields.** |
|
|
| This dataset contains ~600 annotated invoice images (≈200 per language) in **Ukrainian**, **Chinese**, and **Swedish**. Each image shows **real-world invoices** where a human has manually added **checkmarks (✓)** or **circles** to highlight correctly extracted or verified fields (e.g. invoice number, buyer name, line totals, tax rate). |
|
|
| Every sample includes: |
| - The original scanned/photographed invoice image (with visible pen/circle markings) |
| - A JSON annotation file with: |
| - `file_name`: path to the image |
| - `data`: list of extracted fields, each with: |
| - `field`: field name (e.g. "Unique Invoice Identifier", "Vendor Business Address", "Customer/Buyer Name", "Invoice Table Row 1: Line Total Amount", "Applied Tax Percentage") |
| - `checked`: boolean (`true` if the field was marked) |
| - `text`: the extracted text string |
|
|
| Example annotation snippet: |
| ```json |
| { |
| "file_name": "UK/187.jpeg", |
| "data": [ |
| {"field": "Unique Invoice Identifier", "checked": true, "text": "#213253"}, |
| {"field": "Vendor Business Address", "checked": true, "text": "Аллея Беринга 494"}, |
| {"field": "Customer/Buyer Name", "checked": true, "text": "Владилена Басок"}, |
| {"field": "Invoice Table Row 1: Line Total Amount", "checked": true, "text": "1,314.17 грн"}, |
| {"field": "Applied Tax Percentage", "checked": true, "text": "15"} |
| ] |
| } |
| ``` |
|
|
| Why this dataset? |
| Current public invoice datasets (e.g. FATURA, SROIE, CORD, etc.) focus mainly on clean text extraction or layout parsing. |
| This is (to our knowledge) the first public dataset that includes explicit visual human verification signals — checkmarks and circles drawn directly on the invoice images. |
| These visual cues are extremely valuable for training next-generation Document AI / VLM / KIE models that need to: |
|
|
| Understand human feedback/confirmation signals |
| Detect visual annotations (underlines, circles, ticks) |
| Improve reliability in high-stakes invoice processing (finance, logistics, auditing) |
|
|
| The idea was inspired by discussions on visual marking detection in complex documents (see the [Hacker News thread on GLM-OCR](https://news.ycombinator.com/item?id=46924075), where users highlighted the need for better handling of pen/pencil marks like checkmarks in contract/invoice analysis pipelines). |
|
|
| Languages & Size |
|
|
| ```json |
| Ukrainian: ~200 images (UAH currency, Cyrillic addresses, typical UA invoice layouts) |
| Chinese: ~200 images |
| Swedish: ~200 images |
| |
| Total: ≈600 images + corresponding JSON annotations. |
| Structure |
| textinvoice-checkmark-annotations/ |
| ├── Ukrainian/ |
| │ ├── 001.jpeg |
| │ ├── 002.jpeg |
| │ ├── ... |
| │ └── label.txt |
| ├── Chinese/ |
| │ └── ... |
| ├── Swedish/ |
| │ └── ... |
| └── README.md |
| |
| ``` |
| (You can load it easily with datasets.load_dataset("AlroWilde/invoice-checkmark-annotations") — split by language subfolders or add a language column if you prefer a flat/parquet structure later.) |
| License |
| Apache License 2.0 — feel free to use, modify, and build commercial models on top of this dataset. Attribution is appreciated but not required. |
| |
| ## Related Project |
| This dataset pairs well with [ocr-producer](https://github.com/alrowilde/ocr-producer) - a synthetic generator focused on documents. |
| |
| Use real + synthetic data together with this checkmark-annotated set to train more robust KIE / visual-verification models. |
| |
| Contact |
| Questions, collaborations, other language support, or bug reports? |
| Reach out at hi@support.alrowilde.com |
| |
| Happy training! 🚀 |
