Live data from Hacker News

Fine-Tuning TranslateGemma-4B for Better Welsh Translations on an H200 GPU

metalglot.com

1–2 of 2 posts

Re: Fine-Tuning TranslateGemma-4B for Better Welsh Translations on an H200 GPU

#2
Open source repo: https://github.com/grctest/finetuned-gemmatranslate-cy

5% of the fine-tuning took 40 minutes and cost a couple dollars to prove the process works.

Looking forwards to Flash Attention v4 to leave beta, to test fine-tuning performance on a B200 on the cloud, probably a few months away it seems?

What languages would you train TranslateGemma to be able to translate? I was originally thinking about klingon but the available datasets seemed a bit lacking..