LLaVA Visual Instruct Pretraining Subset With Balanced Concept Coverage

A subset of the LAION/CC/SBU dataset filtered for more balanced concept coverage distribution, constructed for the pretraining stage of visual instruction tuning. It contains synthetic captions generated by BLIP for reference and aims to build large multimodal models towards GPT-4 vision/language capability. The dataset was created by liuhaotian and last updated in July 2023.

Use Cases

• Pretrain multimodal models for visual instruction tuning using the filtered image-caption pairs.
• Fine-tune vision-language models on the BLIP synthetic captions associated with images.
• Analyze concept coverage distribution across the filtered subset of LAION/CC/SBU data.
• Train feature alignment models for visual instruction following tasks.

Strengths

• Filtered from the large-scale LAION/CC/SBU dataset for more balanced concept coverage.
• Includes BLIP-generated synthetic captions for reference.
• Specifically constructed for the pretraining stage of visual instruction tuning.

Limitations

• Exact dataset size, number of rows, and specific column structure are unknown.
• Limited information on the filtering criteria and the resulting concept distribution is provided.
• Dataset details and sample data are not available in the provided input.

Provenance

Source

Subset of the LAION/CC/SBU dataset.

Collection Method

Filtered for balanced concept coverage; captions associated with BLIP synthetic captions.

Freshness

Last updated on July 6, 2023.