Tried fine-tuning a small LLM on my own chat logs this weekend

So I've been messing around with local AI models for a few months now, mostly just running stuff like Llama 3 on my desktop. This weekend I decided to take a few thousand messages from my group chat logs and try fine-tuning a tiny 1.5B model on them. I thought it would just memorize inside jokes and be useless. But after about 6 hours of training on my RTX 3060, the thing actually started finishing my sentences in a way that felt freakishly like me. Like it suggested a response to a buddy about our camping trip that was exactly the kind of sarcastic joke I'd make. It wasn't perfect, it still hallucinates some stuff. But man, seeing a model trained on your own voice just hit different. Has anyone else tried training a model on personal data like texts or emails and got something weird or useful out of it?

2 comments

2 Comments

kevin_carr20d ago

That 3060 grind is real, dude. I felt that part about it nailing your sarcastic camping joke because I had the same eerie moment when mine predicted my exact "oh cool" deadpan reaction to a friend's bad take.

dixon.amy20d ago

The trick is to lean into it and just let the AI have those moments because fighting it only makes it worse. When mine predicted my exact dry response to a coworker's story about their cat, I stopped worrying and started using it as a weird kind of mirror. Just accept that these things are scarily good at reading us and move on with your camping joke intact.