search

LAION-AI / CLIP_benchmark

CLIP-like model evaluation
MIT License
590 stars 75 forks source link

What is the composition of the dataset for model training #121

Open geekchen007 opened 7 months ago

geekchen007 commented 7 months ago

Why laion2b_e16(ViT-B-32::laion2b_e16) does perform well in Chinese/English search? e.g. "猫/cat" "狗/dog"，but performs poorly in Japanese or French What is the composition of the dataset for model training？ A CLIP ViT-B/32 model trained with the LAION-2B English subset of LAION-5B？laion2B-multi-chinese-subset？or other？

rom1504 commented 7 months ago

LAION-2B English

There may be a little bit of Chinese in that?

But are you sure it really perform well in Chinese?

You can run the multilingual benchmark with some multilingual models to compare

geekchen007 commented 7 months ago

I did not use a particularly strict dataset for comparison, but I did search for common Chinese words such as "cat", "dog", "dance", "red clothes", and the top-k result was correct.