Stumbled onto a GPT fine-tuning demo at a Seattle meetup that flipped my workflow

I was at this random AI meetup in Seattle two months ago, not expecting much. Some guy from a local startup showed how he fine-tuned a GPT model on his company's internal docs for customer support. I tried it on my own small business chat logs with like 200 examples and the results were wild. My bot went from giving generic answers to actually knowing our return policy and product specs. Took about 3 hours to set up and the API costs stayed under $20 a month. Has anyone else messed with fine-tuning for a niche use case?

2 comments

2 Comments

thompson.robin3d ago

Oh wow that's cool you mentioned the 200 examples part. I heard somewhere that fine-tuning works way better if you focus on quality over quantity, like curating the best 50-100 conversations instead of dumping everything in. Wonder if that would make the setup even faster.

brookep623d ago

Wait, what? Only 200 examples? That seems way too low to get good results. I would've thought you'd need thousands at least. I tried fine-tuning a model on my customer service chat logs and I had about 800 conversations and it still gave me weird answers about our shipping policy. Maybe I need to go back and clean up my data like the other person said. What kind of chat logs were you using where 200 was enough?