You need to agree to share your Hugging Face information 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-POPE

Multi-POPE is the multilingual version of POPE, a benchmark for measuring object hallucination in large vision-language models through binary yes/no questions. It is released in English, Italian, French, German and Spanish, and contains 45,000 entries (9,000 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 POPE
Hugging Face lmms-lab-encoder/POPE
Paper Evaluating Object Hallucination in Large Vision-Language Models (Li et al., EMNLP 2023)
Official repository RUCAIBox/POPE

POPE (Polling-based Object Probing Evaluation) measures object hallucination by asking binary questions of the form "Is there a {object} in the image?" on 500 COCO 2014 validation images, with three negative-sampling strategies (random, popular and adversarial) contributing 3,000 questions each, for 9,000 items in total. The Hugging Face release used here (lmms-lab/POPE) merges the three strategies in a single test split, with the strategy recorded in the category field.

How the dataset was built

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

  • Translated fields: question.
  • Translation model: gpt-4.1-mini.
  • Unchanged: images, object identities, identifiers (id, question_id), the image provenance (image_source), the gold answer (yes / no) and the sampling strategy (category), so that per-strategy metrics remain comparable with the English benchmark.
  • 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 question text 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 contains the three sampling strategies merged in a single test split. The original splits can be recovered by filtering on the category field.

Config Language Split Entries
en English test 9,000
fr French test 9,000
de German test 9,000
it Italian test 9,000
es Spanish test 9,000
Total 45,000

Each test split contains 3,000 random, 3,000 popular and 3,000 adversarial items.

Fields

Field Description
id Item identifier
question_id Identifier of the question inherited from the source release
question Polling question, e.g. "Is there a dog in the image?" (translated)
answer Gold answer: yes or no
image COCO 2014 validation image
image_source Provenance of the image in the source release
category Negative sampling strategy: random, popular or adversarial (unchanged, in English)

Usage

from datasets import load_dataset

ds = load_dataset("VillanovaAI/Multi-POPE", "it", split="test")

# Per-strategy evaluation, as in the original benchmark
random_items = ds.filter(lambda x: x["category"] == "random")
print(ds[0]["question"], 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 covers objects annotated in COCO's 80 categories and is therefore limited to that vocabulary.

Licensing and attribution

The upstream dataset POPE is released by the RUCAIBox group under the MIT license, and the images are the original COCO 2014 validation images, which remain subject to the terms of use of the COCO dataset. Images are redistributed unchanged; only the question text is translated.

Citation

If you use this dataset, please cite the paper describing the multilingual benchmark suite and the original POPE 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{li2023evaluating,
  title={Evaluating Object Hallucination in Large Vision-Language Models},
  author={Li, Yifan and Du, Yifan and Zhou, Kun and Wang, Jinpeng and Zhao, Wayne Xin and Wen, Ji-Rong},
  booktitle={Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year={2023}
}
Downloads last month
170

Papers for VillanovaAI/Multi-POPE