Fine-tuning Gemma 4 locally

I thought an interesting flex would be fine-tuning Gemma 4 (Google's latest offline model, released Friday) to talk like a pirate. Voila! (demo is at the 4 minute mark):

https://www.youtube.com/live/WuCxWJhrkIM

I looked at how to do it and it requires coming up with tons of examples to train on, so I asked Claude to help since that would take very long to do. I also thought the training itself would take hours, but it was done so quickly I missed it while I was doing something else. Fine-tuning took just 90 seconds.

If you want the technicals of it Claude kept a journal and wrote it up (I just put the image and conclusion):

https://taonexus.com/publicfiles/apr2026/pirate-gemma-journal.html

You can use fine-tuning to change Gemma 4's style or behavior.