For experimentation? Absolutely. It can often be done overnight for smaller models and reasonably sized GPUs (24GB+). It'd become a lot less practical with hug

For experimentation? Absolutely. It can often be done overnight for smaller models and reasonably sized GPUs (24GB+). It'd become a lot less practical with huge datasets, but I'd guess that a lot of fine tuning tasks aren't really that large.