Dataset · nyu-visionx
Video visual-spatial VQA benchmark
nyu-visionx/VSI-Bench
Video question-answering dataset for testing visual-spatial understanding in multimodal models.
What's inside
A dataset for visual-question-answering consisting of videos and English text questions, with 10k to 100k examples. The card mentions a VSI-Bench-Debiased subset—a manual pilot filtered by question types rather than the result of automatic Iterative Bias Pruning. Links to arXiv, website, and code are available.
What you can do with it
- Evaluating video multimodal models on spatial questions
- Fine-tuning VLMs to answer based on video
- Testing robustness to biases on the Debiased subset
- Tasks
- Visual question answering
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 28.8K
- Likes
- 78
Models trained on it
How to load it
from datasets import load_dataset
ds = load_dataset("nyu-visionx/VSI-Bench")