Salesforce

/

blip-image-captioning-base

like 425

Hi! In the model card, it is specified that this model is pre-trained on COCO, but I don't see this configuration in the official repo:

The pre-trained checkpoints are computed on a 14M images dataset (according to the paper made of images from COCO, VG, and Conceptual Captions), or on a 129M images dataset (previous dataset + LAION).

Is the HF checkpoint a new configuration computed only on COCO, or is it one of the fine-tuned checkpoints?

Thank you for the amazing work.