Training datasets

[UnderTheMangoTree]

Thanks for your work! How can I get the "blip_laion_cc_sbu_558k.json" and "bliva_llava_150k.json"? the original llava_150k.json doesn't fit the dataload module, such as "ann["question"]".

[gordonhu608]

Thank you for your interest in our work. We acknowledged LLaVA and used their pre-training dataset here: https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain/blob/main/blip_laion_cc_sbu_558k.json. As for the lava 150k, it has multiple nested questions and answers in one data. We simply unnested or flatten the data into one question, answer pair per data point.

[UnderTheMangoTree]

okay, thanks for your help!😁