Live-CC-5M: 5 Million YouTube Video-Caption Samples for Model Pre-training

Live-CC-5M contains 5,047,208 samples of YouTube videos paired with closed captions, with each video clip lasting between 30 and 240 seconds. The dataset was created by author 'chenjoya' and is used for pre-training the LiveCC-7B-Base model. It was last updated on the Hugging Face platform on May 2, 2025.

Use Cases

• Pre-training multimodal AI models based on video-caption pairs.
• Researching video-language alignment based on YouTube content.
• Benchmarking video captioning or generation models based on a large-scale corpus.
• Studying temporal dynamics in short-form video content based on 30-240 second clips.

Strengths

• Contains over 5 million video-caption samples, providing a substantial scale for model training.
• Clips have a defined duration range of 30 to 240 seconds, offering consistency for processing pipelines.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific file formats, size, and data structure are unknown.
• The license is unspecified, which may restrict commercial or broader usage.

Provenance

Source

YouTube videos and their closed captions.

Freshness

Last updated 2025-05-02 15:03:21; freshness should be verified.