Skip to content

Dataset · DatologyAI

Vision-language model evaluation benchmark

DatologyAI/DatBench-Full

A curated set of images and text for testing VLMs without multiple-choice inflation or noise.

What's inside

DatBench is a curated evaluation dataset for vision-language models. Designed to be faithful, discriminative, and efficient, it addresses common issues in modern VLM benchmarks such as multiple-choice inflation, language-only shortcuts, annotation noise, and redundant weak examples. Metadata indicates image and text modalities with a size between 100K and 1M.

What you can do with it

  • Evaluating VLM robustness against shortcuts
  • Comparing models without multiple-choice inflation
  • Selecting examples for an efficient benchmark
Size
100K–1M rows
Downloads this month
1.4K
Likes
77

How to load it

from datasets import load_dataset

ds = load_dataset("DatologyAI/DatBench-Full")