RePOPE (HuggingFace Version)

This dataset is a Hugging Face formatted version of the RePOPE benchmark, uploaded to make it easier to use in evaluation pipelines and multimodal research.

RePOPE is a corrected relabeling of the POPE benchmark, which is commonly used to evaluate object hallucination in Vision-Language Models (VLMs). The dataset fixes annotation errors in the original POPE benchmark and removes ambiguous examples.

Original paper: RePOPE: Impact of Annotation Errors on the POPE Benchmark Yannic Neuhaus, Matthias Hein (2025)

---

Dataset Overview

RePOPE evaluates whether a model incorrectly claims that an object exists in an image.

Example question format:

Question: Is there a car in the image?
Answer: Yes / No

The model must answer correctly based on the image.

Incorrect answers may indicate object hallucination, where the model claims objects exist that are not present.

---

Dataset Statistics

• Total examples: 8185 🔄 Changed 494 answers. ❌ Removed 815 ambigious questions in POPE.

• Image source: COCO 2014 validation set

• Categories:

  • random
  • popular
  • adversarial

These splits follow the structure of the original POPE benchmark.

---

Dataset Features

Feature	Type	Description
id	string	Unique identifier for the sample
question_id	string	Identifier linking to the question instance
question	string	yes/no question about objects in the image
answer	string	Correct label after RePOPE relabeling
pope_old_answer	string	Original answer from the POPE benchmark
image_source	string	COCO image identifier
image	image	The corresponding COCO image
category	string	Split category (random, popular, adversarial)

---

Example

from datasets import load_dataset

dataset = load_dataset("SushantGautam/RePOPE")

print(dataset["test"][0])

Example output:

{
  'id': '...',
  'question_id': '...',
  'question': 'Is there a car in the image?',
  'answer': 'no',
  'pope_old_answer': 'yes',
  'image_source': 'COCO_val2014_000000310196',
  'category': 'adversarial',
  'image': <PIL.Image>
}

---

Why RePOPE?

The original POPE dataset contains annotation errors and ambiguous samples. These issues can significantly impact evaluation metrics such as F1 score.

Key findings from the RePOPE paper:

• Incorrect "yes" labels: 9.3%
• Incorrect "no" labels: 1.7%

This imbalance can distort hallucination evaluation results.

RePOPE fixes these issues by:

• correcting incorrect labels
• removing ambiguous samples
• preserving compatibility with POPE evaluation pipelines

---

Intended Use

This dataset is intended for:

• Evaluating vision-language models
• Studying object hallucination
• Benchmarking multimodal systems
• Comparing hallucination mitigation techniques

Example models evaluated with POPE/RePOPE:

• LLaVA
• Qwen-VL
• BLIP
• GPT-4V

---

Dataset Source

Original annotations from the official RePOPE repository:

https://github.com/YanNeu/RePOPE

Images come from:

MS COCO 2014 validation set

---

Citation

If you use this dataset, please cite the original paper:

@article{neuhaus2025repope,
  title={RePOPE: Impact of Annotation Errors on the POPE Benchmark},
  author={Neuhaus, Yannic and Hein, Matthias},
  journal={arXiv preprint arXiv:2504.15707},
  year={2025}
}

---

Acknowledgements

Thanks to the authors of:

• POPE – Evaluating object hallucination in large vision-language models
• RePOPE – Correcting annotation errors in the POPE benchmark