Skip to content

Dataset · nyu-visionx

CV-Bench visual benchmark for multimodal LLMs

nyu-visionx/CV-Bench

1K to 10K test examples with images and questions to evaluate vision capabilities in multimodal models.

What's inside

Introduced in the Cambrian-1 paper, the CV-Bench benchmark evaluates the visual capabilities of multimodal LLMs in English. It contains 1K to 10K examples for visual question answering (VQA). Images and annotations are pre-packaged in parquet files for use with the HF Datasets library.

What you can do with it

  • Evaluating visual perception and spatial understanding in multimodal LLMs
  • Comparing visual encoders when training VLMs
Tasks
Visual question answering
Languages
English
Size
1K–10K rows
Downloads this month
12.2K
Likes
49

How to load it

from datasets import load_dataset

ds = load_dataset("nyu-visionx/CV-Bench")