license: cc-by-4.0
language:
- en
size_categories:
- 10M<n<100M
pretty_name: Scientific Openly-Licensed Publications - Caption Images
Scientific Openly-Licensed Publications
This repository contains companion material for the following publication:
Tim Tarsi, Heike Adel, Jan Hendrik Metzen, Dan Zhang, Matteo Finco, Annemarie Friedrich. SciOL and MuLMS-Img: Introducing A Large-Scale Multimodal Scientific Dataset and Models for Image-Text Tasks in the Scientific Domain. WACV 2024.
Please cite this paper if using the dataset, and direct any questions regarding the dataset to Tim Tarsi
Summary
Scientific Openly-Licensed Publications (SciOL) is the largest openly-licensed pre-training corpus for multimodal models in the scientific domain, covering multiple sciences including materials science, physics, and computer science. It consists of over 2.7M scientific scientific publications converted into semi-structured data. SciOL contains over 18 Million figure-caption pairs.
Note: This repository only contains the figures and captions of SciOL. For the textual data see: SciOL-text
Data Format
We provide the data in the webdataset format, e.g., captions in plain text files and group and compress them together with the images. Each tar file contains 1000 images and captions. Corresponding figures and captions have the same filename (excluding extention). We split the data into a train, test and dev set.
Citation
If you use our dataset in your work, please cite our paper:
TBD, please use informal citation above.
License
The SciOL corpus is released under the CC BY 4.0 license.