Datasetpublic

Srijan-Upadhyay/arxiv-page-ocr-bench

arXiv Page OCR Benchmark Corpus Dataset Summary arXiv Page OCR Benchmark Corpus is an evaluation benchmark dataset containing high-resolution rendered document page images from scientific arXiv papers (covering STEM disciplines including physics, computer science, mathematics, and quantitative finance) paired with ground-truth Markdown / LaTeX text representations. This dataset is designed to evaluate document parsing, optical character recognition (OCR), and… See the full description on the dataset page: https://huggingface.co/datasets/Srijan-Upadhyay/arxiv-page-ocr-bench.

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes71downloads

Dataset Card

arXiv Page OCR Benchmark Corpus

Dataset Summary

arXiv Page OCR Benchmark Corpus is an evaluation benchmark dataset containing high-resolution rendered document page images from scientific arXiv papers (covering STEM disciplines including physics, computer science, mathematics, and quantitative finance) paired with ground-truth Markdown / LaTeX text representations.

This dataset is designed to evaluate document parsing, optical character recognition (OCR), and layout reconstruction accuracy for vision-language models like Nougat, Swin-Transformer, Qwen2-VL, and GOT-OCR.

Dataset Structure

—Split: train (2,511 rows)
—Total File Size: ~1.34 GB
—Data Fields:
—image: Rendered full-page document image (1.7k x 1.7k resolution).
—label / text: Ground truth Markdown / LaTeX source code of the arXiv document page.
—paper_id: arXiv identifier.

Usage

To load the dataset using the Hugging Face datasets library:

python

from datasets import load_dataset

dataset = load_dataset("Srijan-Upadhyay/arxiv-page-ocr-bench")
print(dataset)

Supported Tasks

—Image-to-Text: Full page LaTeX and Markdown reconstruction from document page images.
—OCR Benchmarking: Measuring Character Error Rate (CER), BLEU score, and Edit Distance on scientific papers.