Advice on replicating selection for other languages

#2
by Varmer - opened

Great to see your educational domain selection from Fineweb 2 for Japanese!

I was wondering if you had any advice on how you trained the classifier for data selection as a lot of the documentation seems to be in Japanese :) I am working on multilingual BabyLM with collaborators and we are aiming to carry out similar domain selection for a range of other languages including Dutch, Hungarian, Japanese, Indonesian, Korean and more.

Any help would be much appreciated!

Marcell Fekete

@Varmer Thank you for your interest in my FineWeb 2 Japanese educational domain selection project, Marcell!

Your multilingual BabyLM project sounds excellent. I'm happy to share some thoughts on classifier training for domain selection:

For the classifier, mMiniLMv2-L6-H384 should be sufficient - it's compact and fast. Creating a multilingual classifier would likely be the most straightforward approach.
For scoring, I recommend using the additive scoring method from fineweb-edu, which has proven effective.
High-performance multilingual models with permissive licenses like deepseek-r1 are now available. These would be excellent for generating educational scores across multiple languages.
You might also want to look at https://huggingface.co/datasets/epfml/FineWeb2-HQ, which uses embedding-based score filtering. This could provide an alternative approach to your selection process.

Best of luck with your multilingual BabyLM project!

Your need to confirm your account before you can post a new comment.

Sign up or log in to comment