albertosei/Merged_Invoice_11k

Merged Invoice NER Dataset

Dataset Description

This dataset consists of processed invoice documents intended for Document Layout Analysis and Named Entity Recognition (NER) tasks (e.g., training LayoutLM, LayoutLMv2, LayoutLMv3, or LiLT).

The data was created by merging multiple invoice datasets. It includes the document images, OCR-extracted words, bounding boxes, and NER tags.

Supported Tasks

• Token Classification: Extracting entities like INVOICE_NUMBER, TOTAL_GROSS_WORTH, SELLER_NAME, etc.
• Key Information Extraction (KIE)

Data Structure

Each example in the dataset contains the following fields:

• id: A unique identifier for the document.
• image: The PIL Image of the invoice.
• words: List of strings (tokens) obtained via OCR.
• boxes: List of bounding boxes corresponding to the words.
  • Format: [x_min, y_min, x_max, y_max]
  • Note: These boxes are likely in absolute pixel coordinates (based on the image size) and NOT normalized to 0-1000. You must normalize them before inputting them into models like LayoutLM.
• ner_tags: List of class IDs (integers) in BIO (Beginning, Inside, Outside) format.

Important Usage Warnings

1. Bounding Box Normalization Required

The boxes in this dataset are likely stored in absolute pixel values (e.g., 0-2500). Models like LayoutLM expect boxes normalized to a 0-1000 scale.

How to handle this: When loading the dataset, you should normalize the boxes relative to the image size:

def normalize_box(box, width, height):
    return [
        int(1000 * (box[0] / width)),
        int(1000 * (box[1] / height)),
        int(1000 * (box[2] / width)),
        int(1000 * (box[3] / height)),
    ]

def preprocess(example):
    w, h = example['image'].size
    # Normalize boxes to 0-1000
    example['boxes'] = [normalize_box(box, w, h) for box in example['boxes']]
    return example

dataset = load_dataset("your-username/dataset-name")
dataset = dataset.map(preprocess)

2. Mixed OCR Sources
This dataset was created by merging sources that may have used different OCR engines (e.g., Tesseract, EasyOCR, or commercial APIs).
Some documents may have slightly different bounding box tightness or word-splitting logic.
Ensure your training pipeline is robust to slight variations in OCR quality.
3. Missing I-Tags for some entities
The label map includes B- (Beginning) tags for all entities, but some entities (like BALANCE_DUE, TOTAL_VAT, IBAN) do not have a corresponding I- (Inside) tag in the schema.
If a value spans multiple words (e.g., "Total VAT"), the model may only be able to label the first word, or you may need to map the subsequent words to O or a generic tag during training.
4. Privacy / PII
This dataset contains invoice data. While processed, users should handle the data with care regarding Personally Identifiable Information (PII) such as names, addresses, and financial details.