Skip to content

Dataset · MCG-NJU

Temporal grounding in long videos (93k pairs)

MCG-NJU/TimeLens2-93K

Large dataset for text-based event retrieval in video clips up to 100 minutes long.

What's inside

The dataset contains 23,793 videos and 93,232 text-query-to-time-interval pairs, including 12,091 examples with multiple segments per query. Ranging from short clips to nearly 100 minutes, the videos cover diverse topics such as education, sports, news, gaming, science, travel, and daily life. Texts are available in English, Chinese, Japanese, and Korean.

What you can do with it

  • Training multimodal models (Video-LLM) for navigation and moment retrieval in long videos
  • Evaluating the quality of temporal grounding algorithms on multilingual material
Tasks
Video text → text
Languages
English, Chinese, Japanese, Korean
Size
10K–100K rows
Downloads this month
21.2K
Likes
13

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("MCG-NJU/TimeLens2-93K")