README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: image
      dtype: image
    - name: UUID
      dtype: string
    - name: Image ID
      dtype: string
    - name: Paper DOI
      dtype: string
    - name: Paper Title
      dtype: string
    - name: Image Caption
      dtype: string
    - name: Image Authors
      dtype: string
    - name: References DOIs
      list: string
    - name: Citing DOIs
      list: string
  splits:
    - name: train
      num_bytes: 56943467858
      num_examples: 94233
  download_size: 72360746667
  dataset_size: 56943467858
license: mit
task_categories:
  - image-text-to-text
language:
  - en
tags:
  - astronomy
  - multimodal
size_categories:
  - 100K<n<1M

AstroCLIMB:

Astronomy Citation Linking from Illustrations: a Multimodal Benchmark

AstroCLIMB is the shared task for 4th WASP: Workshop on Artificial Intelligence for Scientific Publications.

Motivation

Scientists rely on figures to share their discoveries, but this makes the information contained in the figures hard to parse, archive, and search. Recent multimodal neural-network models promise to extract this information, but have not yet been widely tested and adopted by the astronomy community. In partnership with astroexplorer.org, we offer a novel dataset and an associated task as a benchmark to evaluate a model’s multimodal capabilities. The task is to partially reconstruct the citation graph of astronomy papers from their figures and captions.

Dataset Description

The dataset consists of over 100K figure+caption pairs from the astroexplorer.org from recent open access astronomy papers. More details to come.

Dataset({
    features: ['image', 'UUID', 'Image ID', 'Paper DOI', 'Paper Title', 'Image Caption', 'Image Authors', 'References DOIs', 'Citing DOIs'],
    num_rows: 94233
})

(test set to be released in November 2026)

• images are stored as PIL PNG objects.
• Image Caption are English language text.
• UUID, Image ID, Paper DOI, Paper Title are metadata.

A more complete descriptions of the dataset can be found on the WASP2026 website here.

Task Description

The goal of the task is single-class classification. The inputs are pairs of figures, a figure and a caption, or a pair of captions. The possible output classes are:

• same paper
• related papers (references or citing)
• unrelated papers (neither references nor citing)
• same figure (only for figure+caption pairs)

The challenge is being evaluated on Kaggle here.

Contact

For inquiries, contact Felix Grezes at felix.grezes@cfa.harvard.edu.