Skip to content

Dataset · princeton-nlp

Benchmark for analyzing complex scientific charts

princeton-nlp/CharXiv

2,323 scientific charts from arXiv papers to test multimodal model ability to understand complex visualizations.

What's inside

The dataset contains 2,323 high-resolution chart images manually selected by experts from arXiv preprints. The data is in English, presented in image and text modalities, and accompanied by questions about the visualizations and documents. The project was presented at NeurIPS 2024 to identify weaknesses in multimodal LLMs.

What you can do with it

  • Test VLMs on understanding complex scientific diagrams and charts
  • Evaluate visual document data extraction systems (Document VQA)
Tasks
Visual question answering, Document question answering, Question answering
Languages
English
Size
1K–10K rows
Downloads this month
19.2K
Likes
57

How to load it

from datasets import load_dataset

ds = load_dataset("princeton-nlp/CharXiv")