LightOnOCR-bbox-bench

Evaluation benchmark for assessing the ability of vision-language models (VLMs) to localize images within documents using bounding boxes.

Task Description

Given a document page (PDF), the model must predict bounding boxes around images (figures, charts, photographs, etc.) present in the document. This evaluates the model's spatial understanding and ability to distinguish visual content from text in complex document layouts.

Each sample contains 1-5 images to localize, with ground truth bounding boxes normalized to a 0-1000 coordinate space.

Dataset Structure

Splits:

• arxiv: 565 samples from scientific papers
• olmocr_bench: 290 samples from diverse document types

Columns:

• bboxes: List of [x1, y1, x2, y2] bounding boxes normalized to 0-1000 coordinate space
• pdf: Single-page PDF as bytes

Usage

from datasets import load_dataset

# Load dataset
dataset = load_dataset("lightonai/LightOnOCR-bbox-bench")

# Access a sample
sample = dataset['arxiv'][0]
gt_bboxes = sample['bboxes']  # [[x1, y1, x2, y2], ...] normalized to 0-1000
pdf_bytes = sample['pdf']      # Single-page PDF as bytes

# Render PDF to image using your preferred library
# Convert normalized bboxes (0-1000) to pixel coordinates based on rendered image dimensions

Dataset Composition

ArXiv (565 samples):

• Scientific papers with figures, charts, and diagrams
• Automatically annotated using nvpdftex toolkit
• Filtered to 1-5 images per page

OlmOCR (290 samples):

• Diverse document types: mathematical papers, tables, multi-column layouts, historical scans
• Images and annotations from allenai/olmOCR-bench
• Filtered to 1-5 images per page, excluding logo-only samples

Source Datasets

• ArXiv subset: Scientific papers from arXiv
• OlmOCR subset: Derived from allenai/olmOCR-bench

Citation

If you use this dataset, please cite:

@misc{lightonocr2_2026,
  title        = {LightOnOCR: End-to-End, Multilingual, Efficient, State-of-the-Art Vision-Language Model for OCR},
  author       = {Said Taghadouini and Adrien Cavaill\`{e}s and Baptiste Aubertin},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/blog/lightonai/lightonocr-2}}
}