GokulRajaR/invoice-ocr-json

Invoice OCR Dataset

This dataset contains annotated invoice images and their corresponding OCR-extracted text in structured JSON format. The data was originally sourced from an open-source invoice dataset and processed using the GPT-4o mini model to extract relevant fields such as invoice number, date, total amount, vendor, and line items.

Dataset Details

Dataset Description

This dataset is designed to support training and evaluation of document understanding models, particularly for invoice OCR and information extraction tasks. Each record includes the original invoice image (or path reference) and the associated JSON output containing extracted fields. The extraction was performed automatically using GPT-4o mini, OpenAI’s optimized vision-language model.

• Curated by: Gokul Raja (with contributions from the open-source community)
• Shared by [optional]: https://github.com/gokulrajar15
• Language(s) (NLP): English (primarily)
• License: CC BY-SA 4.0 (or appropriate open-source license used by the base dataset)

Dataset Sources

• Original Dataset: https://www.kaggle.com/datasets/osamahosamabdellatif/high-quality-invoice-images-for-ocr

Uses

Direct Use

This dataset is suitable for:

• Fine-tuning document OCR or layout-aware models (e.g., LayoutLM, Donut, or Pix2Struct)
• Building invoice parsing systems
• Benchmarking structured data extraction performance
• Pre-training or few-shot prompting with vision-language models

Out-of-Scope Use

• Use in contexts requiring certified or production-grade financial records
• Any use involving sensitive data inference or extrapolation, as the data is synthetic/open-source
• Commercial deployment without validating the OCR output quality

Dataset Structure

Each entry in the dataset consists of:

• image_path: path or URL to the invoice image
• extracted_data: structured JSON containing fields such as:
  • invoice_number
  • invoice_date
  • vendor_name
  • total_amount
  • line_items: list of items with description, quantity, and price

Data was processed into a single JSONL or JSON format for ease of integration.

Dataset Creation

Curation Rationale

This dataset was created to support research and development of robust OCR pipelines for document understanding, with a focus on financial documents such as invoices.

Source Data

Data Collection and Processing

The source invoices were collected from an open-source dataset. GPT-4o mini was used to extract structured information from the raw invoice images using a vision-language prompting pipeline.

Processing included:

• Image-to-text extraction with GPT-4o mini
• Post-processing and normalization of field names
• Export to JSON format

Who are the source data producers?

The original invoices were created by synthetic or anonymized data generators as part of the open-source dataset. The GPT-4o extraction was performed by the author(s) of this dataset card.

Annotations

Annotation process

All annotations were automatically generated by GPT-4o mini using a consistent prompt pattern and verified through spot-checking.

Who are the annotators?

GPT-4o mini served as the automated annotator. Human review was limited to qualitative validation of extracted outputs.

Personal and Sensitive Information

This dataset does not contain real user or business data. All invoices are either synthetic or anonymized. No sensitive or private personal data is included.

Bias, Risks, and Limitations

• GPT-4o may hallucinate or omit fields, particularly in low-quality images.
• The data reflects the formatting of synthetic invoices and may not generalize to real-world invoice templates.
• Potential regional and language bias toward English formats and vendors.

Recommendations

Users should validate the extracted fields for correctness before use in production systems. Fine-tuning on domain-specific invoices is recommended for improved accuracy.

Citation

BibTeX:

@misc{invoiceocr2025,
  title={Invoice OCR Dataset},
  author={Gokul Raja},
  year={2025},
  url={https://github.com/your-repo/invoice-ocr-dataset}
}