YT-Temporal-1B
Video-Text Dataset
YT-Temporal-1B is a large-scale video-text dataset constructed by the Meta AI research team, containing approximately 18 million YouTube video clips and automatically generated subtitles, providing temporally aligned video language data, widely used for temporal video understanding, action recognition, and video language pre-training research.
Dataset Highlights
A large-scale temporal video language dataset that provides a solid foundation for video understanding research
Billion Frame Scale
As one of the largest temporally annotated video datasets, it contains over 1 billion video frames, providing ample data support for large-scale video language model pre-training.
YouTube Diversity
The data comes from a vast array of YouTube video content, covering various fields such as education, entertainment, lifestyle, and technology, ensuring broad and representative data distribution.
Temporal Annotation
Provides precise temporal boundary information, achieving temporal alignment between video and text, supporting fine-grained video-language association research.
ASR Subtitles
Text descriptions are generated through Automatic Speech Recognition (ASR) technology, enabling large-scale automated annotation and reducing manual annotation costs.
Pre-training Ready
The data format is optimized for direct use in the pre-training process of video language models, eliminating the need for complex data preprocessing steps.
Segment-level Data
Videos have been semantically pre-segmented into meaningful temporal segments, each corresponding to an independent text description, facilitating model training and evaluation.
Applicable Scenarios
Empowering various video AI tasks from basic research to industrial applications
Video Understanding
Train models to perform temporal reasoning on video content, understanding the sequence of events and causal relationships
Action Recognition
Develop systems that can recognize and classify human actions, applied in scenarios such as sports analysis and security monitoring
Video Description
Build models that describe video content in natural language, generating accurate video text descriptions
Video Retrieval
Use text queries to search and retrieve video clips, achieving cross-modal video content search
Quick Start with YT-Temporal-1B
Quickly access the YT-Temporal-1B dataset through the AceDataCloud API
import requests
# Set your API token
API_TOKEN = "your_api_token_here"
# Request YT-Temporal-1B dataset
response = requests.get(
"https://api.acedata.cloud/datasets/yt-temporal-1b",
headers={
"Authorization": f"Bearer {API_TOKEN}",
"Accept": "application/json"
},
params={
"limit": 10
}
)
# Parse the response
data = response.json()
for clip in data.get("clips", []):
print(f"Video ID: {clip['video_id']}")
print(f"Start: {clip['start_time']}s")
print(f"End: {clip['end_time']}s")
print(f"Caption: {clip['caption']}")
print("---")
3 Steps to Get Started Quickly
From registration to usage, you can start your video AI research in just a few minutes
Register an Account
Register your Ace Data Cloud account at platform.acedata.cloud to gain immediate access to the platform.
Get API Key
Create an API Key in the console for authentication and dataset access authorization, secure and convenient.
Call Dataset API
Use your API Key to call the YT-Temporal-1B dataset interface and start exploring large-scale video-text data.
Start Exploring the YT-Temporal-1B Dataset
Ultra-large-scale video-text data, temporally precise alignment. Whether you are a video AI researcher or a multimodal model developer, YT-Temporal-1B is the ideal data foundation.