compare with BLIP2

#30
by tmamox - opened

thank you for great repo.
BLIP2 in coco score recall@5: 97% https://paperswithcode.com/sota/image-to-text-retrieval-on-coco?metric=Recall%405
jina-clip-v2 in coco score recall@5: 81.5%
Why the big gap?

Jina AI org

hi @tmamox i believe the model you shared above is not the BLIP2, but a variation of BLIP2 fine-tuned on MSCOCO. There are many reasons to leads a performance difference, e.g. model size, supported languages, training data composition, training strategy etc. I suggest you to build your own benchmark and evaluate on it. MSCOCO is a bit overused.

@bwang0911 great. I will try.

tmamox changed discussion status to closed

Sign up or log in to comment