ASID-1M: Attribute-Structured and Quality-Verified Audiovisual Instructions

ASID-1M is a large-scale audiovisual instruction dataset designed to support universal video understanding through fine-grained, controllable supervision. It addresses the limitations of traditional monolithic captions by providing attribute-structured and quality-verified data. The dataset aims to improve coverage of both visual and auditory elements within video content for more precise model training.

Use Cases

• Instruction tuning for video-language models
• Fine-grained audiovisual understanding
• Controllable video captioning
• Multimodal instruction following

Strengths

• Attribute-structured data for fine-grained supervision
• Quality-verified instructions
• Large-scale coverage for multimodal tasks

Limitations

• Specific column names and sample data are currently unavailable in the provided summary

Provenance

Source

AudioVisual-Caption

Freshness

Last updated March 11, 2026.

Geography

Global (English language focus)