That AI consultant who said 'just train a small model on your own data' cost me 3 weeks of work

Turns out the small language model I built for my customer support bot in Denver couldn't even tell a refund request from a shipping question, so now I'm back to using a bigger API and paying per token like everyone else.

2 comments

2 Comments

seth_allen9d ago

I used to be in the "just train your own" camp too. You hear about how cheap it is and think you're going to save money and have this perfectly tuned bot. Well, I spent two weekends getting mine to handle basic stuff like store hours and return policies. Then a customer asked if we shipped to Canada and it replied with a recipe for poutine. Three weeks of tweaking and it still couldn't tell the difference between a lost package and a billing question. I finally gave up and went back to paying OpenAI by the token.

hugo_coleman288d ago

Did you try fine-tuning a smaller model like GPT-2 on just your customer chat logs? I've had better luck with that than training from scratch, still not perfect but way less headache.