Live-WhisperX-526K: 526,000 Video-Text Pairs for Instruction Tuning

Live-WhisperX-526K contains 526,000 video-text records developed by chenjoya in 2025 to train the LiveCC-7B-Instruct model. The collection features synthetic user prompts generated by GPT-4o to support multimodal instruction tuning for video-to-text tasks.

Use Cases

• Instruction tuning of multimodal models using GPT-4o generated prompts
• Video-to-text generation for live captioning tasks
• Training models for video understanding and temporal alignment

Strengths

• 526,000 record scale
• Instruction-tuned format using GPT-4o prompts
• Linked to peer-reviewed research (Arxiv 2504.16030)

Limitations

• Potential synthetic data bias from GPT-4o generated prompts
• Restricted to academic and educational use per author description
• Lack of detailed column schema in public metadata

Provenance

Source

chenjoya, Show Lab (Arxiv 2504.16030)

Collection Method

Synthetic generation via GPT-4o and automated transcription

Freshness

Last updated August 2025

Geography

United States