Panda-70M
Video Description Dataset
Panda-70M is a large-scale video description dataset released by Snap Research, containing 70 million high-quality video-text pairs. The videos come from HD-VILA-100M and have been filtered through semantic segmentation and multi-model scoring, while the text is generated through multi-teacher cross-validation to produce accurate descriptions, widely used for video understanding, generation, and retrieval in various multimodal tasks.
Dataset Highlights
A large-scale, high-quality video description dataset that provides a solid foundation for multimodal research
Ultra-large-scale Video Descriptions
Contains 70 million video-text pairs, making it one of the largest high-quality video description datasets, providing ample data support for large-scale model training.
Quality-driven Screening Process
Videos are semantically segmented, and multi-dimensional scoring is performed using CLIP and LLaMA to ensure each data point has high semantic relevance and information density.
Multi-teacher Cross-validation
Text descriptions are generated through cross-validation by multiple Visual-Language Models (VLM), effectively reducing bias and noise from a single model, enhancing annotation accuracy.
Temporal Dynamic Understanding
Data descriptions capture dynamic actions and scene transitions in videos, focusing not only on static frames but also on coherent semantic expressions over time.
Open Source Apache 2.0 License
Adopts the Apache-2.0 open-source protocol, freely usable for both academic research and commercial applications without additional authorization.
Endorsed by Snap Research
Strictly validated and released by the Snap Research team, with methodologies peer-reviewed and benchmark-tested, possessing academic authority.
Applicable Scenarios
From video generation to multimodal pre-training, covering mainstream research directions
Video Generation
Train text-to-video generation models (like Sora), using large-scale high-quality video-text pairs to improve generation quality
Video Description
Automatically generate video descriptions and subtitles, applied to video accessibility and content understanding
Video Retrieval
Text-based video search and recommendation system, achieving cross-modal precise matching
Multimodal Pre-training
Joint video-text representation learning, building a general multimodal foundation model
Data Preview
The following are a few example records from the dataset, including video ID, description text, duration, resolution, and CLIP score
{"video_id": "v_0a1b2c3d", "caption": "A woman walks through a sunlit garden, pausing to examine blooming roses.", "duration": 12.4, "resolution": "720p", "clip_score": 0.89}
{"video_id": "v_4e5f6g7h", "caption": "A drone flies over a coastal city at sunset, revealing the harbor below.", "duration": 8.7, "resolution": "1080p", "clip_score": 0.93}
{"video_id": "v_8i9j0k1l", "caption": "A chef demonstrates pasta-making technique, rolling dough by hand.", "duration": 15.2, "resolution": "720p", "clip_score": 0.86}
3 Steps to Get Started Quickly
Quickly integrate Panda-70M into your research workflow from browsing to training
Browse the Dataset
Browse the Panda-70M dataset on the Ace Data Cloud platform, view video samples and metadata information.
Download Data
Download video fragment files and descriptive annotation data, with support for selecting different resolutions and data subsets on demand.
Load and Train
Load the data into your video understanding or video generation training workflow, and start model training and experimentation.
Start Exploring Panda-70M Data
70 million high-quality video-text pairs, Apache-2.0 open source license, available immediately. Whether you are a video generation researcher or a multimodal AI developer, Panda-70M is an indispensable data resource.