You need to agree to the conditions to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Multi-AI2D

Multi-AI2D is the multilingual version of AI2D, a multiple-choice benchmark on the understanding of science diagrams. It is released in English, Italian, French, German and Spanish, and contains 15,440 entries (3,088 per language).

The dataset is part of MEVBench (Multimodal European Vision Benchmark) and was used as an evaluation benchmark in the paper Multilingual Training and Evaluation Resources for Vision-Language Models.

Source dataset

Dataset AI2D
Hugging Face lmms-lab/ai2d
Paper A Diagram Is Worth A Dozen Images (Kembhavi et al., ECCV 2016)

AI2D is a benchmark for diagram interpretation and reasoning, built on over 5,000 grade-school science diagrams annotated with their constituents and relationships as Diagram Parse Graphs (DPG), for a total of 15,000 questions and answers. The Hugging Face release used here (lmms-lab/ai2d) provides 3,088 multiple-choice test items, each with an image, a question and a list of answer options.

How the dataset was built

The English test split was extended to four additional languages through an automated translation pipeline, keeping the benchmark structure intact:

  • Translated fields: question and the answer options.
  • Translation model: gemini-2.5-flash.
  • Unchanged: images and the gold answer — the index of the correct option — so that the original evaluation scripts and metrics can be applied unchanged to every language.
  • English configuration: the en configuration contains the original English test split and is kept as a reference for cross-lingual comparison.
  • No regeneration or re-annotation: only the textual content was translated; images were not modified or regenerated.

The translation protocol is described in Section 4 of the accompanying paper.

Dataset structure

Configurations and splits

Each language is a separate configuration, and all configurations share the same split name (test).

Config Language Split Entries
en English test 3,088
fr French test 3,088
de German test 3,088
it Italian test 3,088
es Spanish test 3,088
Total 15,440

Fields

Field Description
question Question about the diagram (translated)
options List of candidate answers (translated)
answer Gold answer: zero-based index (as a string) of the correct element in options, i.e. "0" = first option (A)
image The science diagram

Usage

from datasets import load_dataset

ds = load_dataset("VillanovaAI/multi-ai2d", "it", split="test")
print(ds[0]["question"])
print(ds[0]["options"], ds[0]["answer"])

Limitations

  • The translations are produced automatically and inherit any error of the source benchmark as well as any error of the translation model. They are intended for evaluation, not for training.
  • The benchmark is limited to grade-school science diagrams and does not cover other visual domains.

Licensing and attribution

This multilingual release is distributed under CC BY-SA 4.0, coherently with the license declared for the Hugging Face release used as source. Because the license is share-alike, derivative works must be distributed under the same terms. The upstream benchmark is AI2D, introduced by the Allen Institute for AI (Kembhavi et al., ECCV 2016); images are redistributed unchanged while the question and the answer options are translated.

Citation

If you use this dataset, please cite the paper describing the multilingual benchmark suite and the original AI2D paper.

@article{baiamonte2026multilingual,
  title={Multilingual Training and Evaluation Resources for Vision-Language Models},
  author={Baiamonte, Daniela and Fano, Elena and Gabburo, Matteo and Simonazzi, Stefano and Rigutini, Leonardo and Zugarini, Andrea},
  journal={arXiv preprint arXiv:2604.18347},
  year={2026}
}

@inproceedings{kembhavi2016diagram,
  title={A Diagram Is Worth A Dozen Images},
  author={Kembhavi, Aniruddha and Salvato, Mike and Kolve, Eric and Seo, Minjoon and Hajishirzi, Hannaneh and Farhadi, Ali},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2016}
}
Downloads last month
108

Papers for VillanovaAI/multi-ai2d