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..