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-ScienceQA

Multi-ScienceQA is the multilingual version of ScienceQA, a multimodal multiple-choice benchmark on school-level science questions. It is released in English, Italian, French, German and Spanish, and contains 21,205 entries (4,241 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

ScienceQA is a benchmark of 21,208 multimodal multiple-choice questions on elementary and high-school science (natural, language and social science), where most items are annotated with a lecture and a solution/explanation that provide the reasoning behind the correct answer. The Hugging Face release used here (derek-thomas/ScienceQA) contains train (12,726), validation (4,241) and test (4,241); Multi-ScienceQA is built from the test split.

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, choices, hint, lecture and solution.
  • Translation model: gpt-4.1-mini.
  • Unchanged: images, the gold answer (answer, i.e. the index of the correct choice) and all the metadata used for stratified analysis (task, grade, subject, topic, category, skill).
  • 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 4,241
fr French test 4,241
de German test 4,241
it Italian test 4,241
es Spanish test 4,241
Total 21,205

Fields

Field Description
image Image context of the question (optional in ScienceQA)
question Question text (translated)
choices List of candidate answers (translated)
answer Gold answer: zero-based index of the correct element in choices
hint Optional hint for the question (translated)
lecture Background knowledge needed to answer (translated; may be empty)
solution Explanation of the correct answer (translated; may be empty)
task High-level task group (unchanged, in English)
grade School grade (unchanged)
subject Subject, e.g. natural science, language science, social science (unchanged)
topic Topic within the subject (unchanged)
category Category within the topic (unchanged)
skill Skill targeted by the question (unchanged)

The multimodal nature of the dataset is inherited: some questions come with an image, others are text-only. The image field is therefore empty for items whose context is purely textual.

Usage

from datasets import load_dataset

ds = load_dataset("VillanovaAI/Multi-ScienceQA", "it", split="test")
print(ds[0]["question"])
print(ds[0]["choices"], ds[0]["answer"])
print(ds[0]["lecture"])
print(ds[0]["solution"])

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.
  • Metadata fields are kept in English on purpose, to preserve comparability with the original benchmark.

Licensing and attribution

This multilingual release is distributed under CC BY-SA 4.0, coherently with the license of the upstream ScienceQA benchmark (Lu et al., NeurIPS 2022). Because the license is share-alike, derivative works must be distributed under the same terms. Images are redistributed unchanged; only the textual fields are translated.

Citation

If you use this dataset, please cite the paper describing the multilingual benchmark suite and the original ScienceQA 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{lu2022learn,
  title={Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering},
  author={Lu, Pan and Mishra, Swaroop and Xia, Tony and Qiu, Liang and Chang, Kai-Wei and Zhu, Song-Chun and Tafjord, Oyvind and Clark, Peter and Kalyan, Ashwin},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2022}
}
Downloads last month
96

Papers for VillanovaAI/Multi-ScienceQA