LLaVA CC3M Pretrain 595K: A Subset for Visual Instruction Tuning

595,000 image-text pairs form a subset of the CC-3M dataset, filtered for balanced concept coverage. It was created by liuhaotian for the pretraining stage of visual instruction tuning, aiming to build large multimodal models. The dataset was last updated on July 6, 2023.

Use Cases

• Pretraining multimodal models for feature alignment based on image-text pairs.
• Training models for visual instruction tuning based on the described concept-balanced subset.
• Benchmarking vision-language model performance on a filtered version of CC-3M data.
• Generating synthetic captions for images using the referenced BLIP captioning method.

Strengths

• Contains 595,000 data points, providing a substantial base for model training.
• Filtered from CC-3M for more balanced concept coverage, which may reduce bias.
• Includes BLIP-generated synthetic captions, offering an additional reference source.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but other scale metrics like file size and format are unknown.
• Last updated 2023-07-06 08:51:35; freshness should be verified for current research.

Provenance

Source

Subset of the CC-3M dataset.

Collection Method

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

Freshness

2023-07-06