Realized I was overtraining my AI model with too much data

I spent 6 months feeding my custom chatbot tens of thousands of support tickets. Last week I tested it on a simple question about shipping times, and it gave me a 3 paragraph essay about warehouse logistics. A buddy who works in ML told me I was basically force-feeding it noise, not signal. Has anyone else hit a wall with too much raw data before?

2 comments

2 Comments

kevin_sullivan5h ago

Honestly used to think more data was always better, like you can't have too much of a good thing. Then I trained a model on like 50,000 customer emails and it started writing every reply like it was a legal document, even for simple stuff like "where's my order?" It would go on about warehouse floor layouts and inventory buffers. Totally changed my mind. Now I'm way more careful about picking only the most useful examples and keeping the number smaller.

the_kevin1h ago

Yeah so I feel this one hard, @kevin_sullivan basically called me out by name over there. I once fed my bot like 80,000 random chat logs thinking I was being smart, and it started every single reply with "Per your inquiry regarding..." even when someone just asked "hey my package is late." The bot was basically a lawyer who forgot how to talk to normal people. I bet your warehouse logistics essay would make a great bedtime story though.