BLIP3o/

BLIP3o-Pretrain-Long-Caption

This collection contains 27 million images, each paired with a long (~120 token) caption generated by Qwen/Qwen2.5-VL-7B-Instruct.

from huggingface_hub import snapshot_download

snapshot_download( repo_id="BLIP3o/BLIP3o-Pretrain-Long-Caption", repo_type="dataset" )

You don’t need to unpack the .tar archives, use WebDataset support in 🤗datasets instead: from datasets import… See the full description on the dataset page: https://huggingface.co/datasets/BLIP3o/BLIP3o-Pretrain-Long-Caption.