---
license: other
license_name: openxvqa-research-only
license_link: https://arxiv.org/abs/2310.08864
task_categories:
- visual-question-answering
- multiple-choice
language:
- en
size_categories:
- 1K<n<10K
pretty_name: OpenXVQA
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*.parquet
---

# OpenXVQA

VQA benchmark derived from [Open X-Embodiment](https://arxiv.org/abs/2310.08864)
robot manipulation data. Tests visual understanding of diverse robotic scenes
and actions.

## Format

Each row contains:

```json
{
  "id": "<unique id>",
  "question": "<question text>",
  "choices": "['option1', 'option2', ...]",
  "correct_answer": <int index>,
  "image": <PNG bytes>
}
```

`choices` is a Python-literal string of a list of answer options.
`correct_answer` is the integer index into `choices` of the correct option.
`image` is embedded as binary in the parquet (HF `Image()` decodes on load).

## Citation

```
@article{padalkar2023open,
  title={Open X-Embodiment: Robotic Learning Datasets and RT-X Models},
  author={Padalkar, Abhishek and ...},
  journal={arXiv:2310.08864},
  year={2023}
}
```

## License

Annotations are derived from Open X-Embodiment, which is openly published.
The MCQ formatting here is a research-only artifact.