expand to multilingual captions

[rom1504]

Multilingual clip can be used as a filter https://huggingface.co/sentence-transformers/clip-ViT-B-32-multilingual-v1

try it at https://colab.research.google.com/drive/1kOKsf-X0nS2_ael7scPa9raf0pBmxUgW#scrollTo=ScN8GQwN3xXq

First experimentations show that clip is still ok at latin languages (french, portuguese, spanish, french) very bad at non latin, especially so for non european mclip keeps the same ordering of performances, but it still ok for the most difficult languages so maybe we want to decrease 0.3 to like 0.26 for mclip

[rom1504]

interesting things to do:

• [x] evaluate on more samples (for example take a bunch of samples from laion, translate the captions and run mclip and clip to get the matches)
• [x] integrate mclip in the worker script

[rom1504]

probably https://github.com/ARKseal/crawlingathome-worker/blob/master/clip_filter.py

[rvencu]

please use this, it is the latest version in use https://github.com/rvencu/crawlingathome-gpu-hcloud/blob/main/clip_filter.py

[rom1504]

this is ongoing, already 700M samples extracted

[rom1504]

done and released