Skip to content

Dataset · YqjMartin

Long video and audio moment retrieval benchmark

YqjMartin/FLARE

Set of 87.7k short video clips with three descriptions each for evaluating multimodal retrieval in long videos.

What's inside

A benchmark accepted at the NeurIPS 2026 E&D Track, collected from 399 long videos (10 to 60 minutes, totaling 225.4 hours) from Video-MME. The clips are sliced into 87,697 short fragments, each augmented with three textual descriptions accounting for visuals and audio. Repository size is 100,000 to 1 million entries.

What you can do with it

  • Evaluating multimodal video and audio retrieval models
  • Testing video-text-to-text models on long-context understanding
Tasks
Video text → text
Size
100K–1M rows
Downloads this month
792
Likes
2

How to load it

from datasets import load_dataset

ds = load_dataset("YqjMartin/FLARE")