Skip to content

Dataset · VLM2Vec

Video Understanding Benchmark for Multimodal AI

VLM2Vec/MVBench

An English set of 1k–10k examples for evaluating models on video question-answering tasks.

What's inside

This is a reproducible copy of the MVBench benchmark in English, including images, text, and tabular data ranging from 1k to 10k records. The set is designed to evaluate temporal dynamics understanding in videos using visual-question-answering and video classification tasks. Part of the 320 clips from the NTU RGB+D set requires manual downloading due to licensing restrictions.

What you can do with it

  • Evaluating multimodal language models (VLMs) on video dynamics and event understanding tasks
  • Benchmarking VideoQA systems
Tasks
Visual question answering, Video classification
Languages
English
Size
1K–10K rows
Downloads this month
87K
Likes
0

How to load it

from datasets import load_dataset

ds = load_dataset("VLM2Vec/MVBench")