---
dataset_info:
  features:
    - name: image
      dtype: string
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 11860671915
      num_examples: 13320
  download_size: 11661151472
  dataset_size: 11860671915
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: apache-2.0
language:
- es
tags:
- finance
pretty_name: SpanishOCR
size_categories:
- 10K<n<100K
task_categories:
- image-to-text
---

----------------------------------------------------------------
# Dataset Card for SpanishOCR Dataset


### Dataset Summary
The SpanishOCR dataset contains images derived from regulatory documents from Peru government in pdf format. This dataset is used for benchmarkingg and evaluating Large Language Models ability on converting unstructured dcuments, such as pdfs and images, into machine readable format, particularly in finance domain, where the conversion task is more complex and valuable.

### Supported Tasks

- **Task:** Image-to-Text
- **Evaluation Metrics:** ROUGE-1

### Languages

- Spanish

## Dataset Structure

### Data Instances
Each instance in the SpanishOCR dataset comprises 2 fields:
- **image** : image of regulatory document, each image represent one page in pdf
- **text**: ground truth of text extracted from regulatory document

### Data Fields
- **image** : string - Base64-encoded png
- **text**: extracted text from pdf files

## Dataset Creation

### Curation Rationale
The SpanishOCR dataset was curated to support research and development on information extraction techniques and layout retain ability for unstructured documents in Spanish. By providing real-world regulatory documents in unstructured format with ground truth, the dataset seeks to address challenges in extracting informat as well as layouts and convert into machine-readable format.

### Source Data
#### Initial Data Collection and Normalization
- The source data are regulatory documents for Securities Market from Peru government publically available.
- The pdf files of those documents are downloaded and split via API, split into page per file, and convert into images. 

#### Who are the Source Language Producers?
- The source data are regulatory documents for Securities Market from Peru government, and is collected to a Peruvian digital platform: https://www.gob.pe/institucion/smv/buscador?contenido=normas&sheet=1&sort_by=re

### Annotations

#### Annotation Process
- The dataset was prepared by collecting, spliting, and converting regulatory documents in Spanish
- The annotation of ground truth text is done by Python OCR package ```fitz```

#### Who are the Annotators?

- The dataset stems from publicly available regulatory documents.
- No external annotation team was involved beyond this.

### Personal and Sensitive Information

- The SpanishOCR dataset does not contain any personally identifiable information (PII) and is strictly focused on Spanish-language regulatory data. No personal or sensitive information is present in the dataset.

## Considerations for Using the Data

### Social Impact of Dataset
This dataset enables AI models to extract structured information from scanned financial documents in Spanish, supporting downstream applications in finance, regulation, and transparency initiatives across Spanish-speaking regions. By aligning page-level PDF images with accurate ground truth text, it supports the development of fairer, more inclusive models that work across diverse formats and languages. 

### Discussion of Biases
- The source data is limited to regulatory documents for Securities Markets, it may underrepresent other financial document types such as tax records, bank statements, or private company reports, potentially limiting model generalizability.

### Other Known Limitations
- The ground truth text is extracted using the Python package fitz (PyMuPDF), which may introduce inaccuracies in complex layouts, potentially affecting training quality and evaluation reliability.
- While the dataset covers regulatory documents, it may lack sufficient variety in layout styles (e.g., handwritten notes, non-standard financial forms, embedded charts), which could limit a model’s ability to generalize to less structured or unconventional financial documents.

## Additional Information

### Dataset Curators
- Yueru He
- Ruoyu Xiang
- The FinAI Team

### Licensing Information

- **License:** Apache License 2.0

### Citation Information

If you use this dataset, please cite:

```bibtex
@misc{peng2025multifinbenmultilingualmultimodaldifficultyaware,
      title={MultiFinBen: A Multilingual, Multimodal, and Difficulty-Aware Benchmark for Financial LLM Evaluation}, 
      author={Xueqing Peng and Lingfei Qian and Yan Wang and Ruoyu Xiang and Yueru He and Yang Ren and Mingyang Jiang and Jeff Zhao and Huan He and Yi Han and Yun Feng and Yuechen Jiang and Yupeng Cao and Haohang Li and Yangyang Yu and Xiaoyu Wang and Penglei Gao and Shengyuan Lin and Keyi Wang and Shanshan Yang and Yilun Zhao and Zhiwei Liu and Peng Lu and Jerry Huang and Suyuchen Wang and Triantafillos Papadopoulos and Polydoros Giannouris and Efstathia Soufleri and Nuo Chen and Guojun Xiong and Zhiyang Deng and Yijia Zhao and Mingquan Lin and Meikang Qiu and Kaleb E Smith and Arman Cohan and Xiao-Yang Liu and Jimin Huang and Alejandro Lopez-Lira and Xi Chen and Junichi Tsujii and Jian-Yun Nie and Sophia Ananiadou and Qianqian Xie},
      year={2025},
      eprint={2506.14028},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.14028}, 
}
```