Datasets:

tianyin
/

RotBench

license: apache-2.0
dataset_info:
  features:
    - name: image
      dtype: image
    - name: image_name
      dtype: string
  splits:
    - name: small
      num_bytes: 694935
      num_examples: 50
    - name: large
      num_bytes: 4651568
      num_examples: 300
  download_size: 10233742
  dataset_size: 5346503
configs:
  - config_name: default
    data_files:
      - split: large
        path: data/large-*
      - split: small
        path: data/small-*
task_categories:
  - image-text-to-text

RotBench

Data for RotBench: Evaluating Multimodal Large Language Models on Identifying Image Rotation.

arXiv | GitHub

Dataset Summary

RotBench is a benchmark for evaluating whether multimodal large language models (MLLMs) can identify image orientation. It contains 350 manually filtered images. The dataset includes two subsets:

Large: 300 images
Small: 50 images

All images were drawn from the Spatial-MM dataset and passed a two-stage human verification process to ensure rotations are distinguishable.

Dataset Download

from datasets import load_dataset

dataset = load_dataset("tianyin/RotBench")
data = dataset['large'] # or dataset['small']

for i, sample in enumerate(data):
    image = sample['image']  # PIL Image object
    image_name = sample['image_name']

Citation

If you find our data useful in your research, please cite the following paper:

@misc{niu2025rotbenchevaluatingmultimodallarge,
      title={RotBench: Evaluating Multimodal Large Language Models on Identifying Image Rotation}, 
      author={Tianyi Niu and Jaemin Cho and Elias Stengel-Eskin and Mohit Bansal},
      year={2025},
      eprint={2508.13968},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.13968}, 
}