README.md

---
license: apache-2.0
---
---

# 📄 Invoice Image Dataset (N8N-Friendly Streaming Format)

This repository contains a large-scale collection of invoice data stored in a **simple, sequential, machine-friendly structure** designed specifically for workflow automation tools such as **n8n**, Airflow, Make.com, and other ETL/automation systems.

The dataset is formatted to allow **easy, reliable, and high-performance pulling of individual invoice files one-by-one**, without needing to download the entire dataset.

---

## 📁 Directory Structure

All invoice data are stored inside a single directory:

```
data_0001/
│
├── invoice_00000001.png
├── invoice_00000002.png
├── invoice_00000003.png
├── invoice_00000004.png
└── ...
data_0002/
│
├── invoice_00010001.png
├── invoice_00010002.png
├── invoice_00010003.png
├── invoice_00010004.png
└── ...
...
```

### ✅ Naming Format

Each invoice follows a **fixed-length 8-digit numbering scheme**:

```
invoice_00000000.format
```

* Zero-padded (`06d` or more) numbering supports **millions of invoices** without renaming.
* `format` can be `.png`, `.jpg`, `.jpeg`, `.webp`, depending on original source.
* Each file is **independent** and can be downloaded individually.

---

## 🎯 Purpose of This Dataset

This dataset acts as a **normalized, unified output repository** for invoice data collected and standardized from multiple sources:

* Other Hugging Face datasets
* GitHub repositories
* Parquet datasets containing embedded data
* Third-party OCR datasets
* Custom pipelines

Each invoice is extracted, normalized, renamed, and uploaded in a consistent structure.

---

## 🔌 Why This Format Is Perfect for n8n Workflows

Most automation tools (including **n8n**) struggle with:

* datasets stored in large compressed files (ZIPs, JSONL, Parquet)
* datasets where data must be decoded from arrays/blobs
* datasets requiring full-dataset downloads

This repo is intentionally built to support **file-by-file streaming**, so n8n can:

### ✔ Pull one file at a time

Example URL:

```
https://huggingface.co/datasets/<USER>/<REPO>/resolve/main/data_0001/invoice_00000042.png
```

### ✔ Process invoices sequentially

You can loop:

1 → 2 → 3 → 4 → ...

As long as the file exists, n8n continues; when it fails (404), it stops — perfect for paginated workflows.

### ✔ Avoid handling large downloads

Each invoice is an independent file.

### ✔ Use lightweight logic

You only need a simple incrementing counter in n8n.

---

## 🧩 Example n8n Workflow Logic

1. Start with variable:

   ```
   index = 1
   ```

2. Construct file URL:

   ```
   url = https://huggingface.co/datasets/<USER>/<REPO>/resolve/main/<folder>/invoice_<index padded>.png
   ```

3. HTTP GET → process invoice

4. Increment index

5. Repeat until HTTP GET returns 404

6. Stop workflow automatically

This allows n8n to **stream** invoices from Hugging Face like a controlled queue.

---

## 📦 What This Dataset Contains

* Only raw invoice data
* No labels
* No associated OCR text
* No metadata

This repository is intended specifically as an **output target** for standardized invoices — not as an annotated machine learning dataset.

---

## ⚙️ How It Was Generated

A controlled pipeline:

1. Loads invoices from multiple formats (HF dataset, Parquet, GitHub, OCR outputs, etc.)
2. Extracts image data
3. Converts and normalizes to `.png` or `.jpg`
4. Saves to `folder/invoice_XXXXXXXX.format`
5. Uploads each file individually using the Hugging Face API
6. Logs progress so the job can resume at any index

Designed with:

* Kaggle execution compatibility
* Resume-safe operation
* Support for millions of invoices
* No Git requirements (uploads via HF API)

---

## 📬 Contact / Usage Notes

This dataset is meant for:

* ETL automation
* Synthetic invoice processing pipelines
* AI model pre-processing
* OCR experiments
* Workflow system testing (n8n, make.com)

If you use this dataset in your workflow, feel free to reference or credit the repo.

---