I want to use the existing image-text pedestrian dataset and finetune the BLIP model. Should I use pre-trained checkpoints weights or finetuned checkpoints weights?

[shams2023]

I want to use the existing image-text pedestrian dataset and finetune the BLIP model. Should I use pre-trained checkpoints weights or finetuned checkpoints weights? The generated text should be as detailed as possible, with a length of 40-60! Looking forward to your answer! Thank you!