mauroibz/facturas-argentinas

Benchmark Dataset

This directory contains a curated benchmark dataset of 500 factura documents, sampled from a full dataset with specific distribution criteria.

Contents

• facturas.json - JSON file containing the extracted factura data for all 500 documents
• jpg/ - Directory containing the corresponding JPG images (one per document)
• documento_ids.txt - Plain text file listing all selected documento_id values (one per line)

Sampling Criteria

The 500 documents were randomly sampled to achieve the following distributions:

1. Currency (Moneda): 90% ARS, 10% other currencies (EUR, USD)
2. Document Type (Tipo Comprobante): Uniform distribution across types A, B, and C
3. Template: Uniform distribution across all available templates (baseline, cards, columns, columns_notitles, sidebar)
4. Theme: Uniform distribution within each template across themes (code, sans, typewriter)

Dataset Statistics

• Total documents: 500
• Currency breakdown: ~450 ARS, ~50 other currencies
• Document types: Uniformly distributed across A, B, C
• Templates: Uniformly distributed across 5 template types
• Themes: Uniformly distributed across 3 theme types

Usage

The dataset can be used for:

• Model evaluation and benchmarking
• Testing vision-language models on structured document understanding
• Validating extraction accuracy across different document types and styles

Each document has a corresponding JPG image in the jpg/ directory, named as {documento_id}.jpg.

Generation

The code for this dataset generation will be made public within the upcoming week