Skip to content

Dataset · BLINK-Benchmark

Visual perception benchmark for multimodal LLMs

BLINK-Benchmark/BLINK

Image and text dataset for evaluating basic visual perception skills in multimodal models.

What's inside

The benchmark contains 1,000 to 10,000 examples combining images and text tasks. It was created by the authors of the paper "BLINK: Multimodal Large Language Models Can See but Not Perceive" to test visual perception abilities that humans solve instantly but other tests miss.

What you can do with it

  • Evaluate and compare multimodal language models (MLLMs)
  • Diagnose gaps in basic visual perception of neural networks
Size
1K–10K rows
Downloads this month
19.3K
Likes
50

How to load it

from datasets import load_dataset

ds = load_dataset("BLINK-Benchmark/BLINK")