That moment I realized my AI training data was making my model dumber

I spent 3 months building a chatbot for local small businesses in Austin. Tested it at a coffee shop and the owner laughed at me. Said it kept recommending taco trucks from Dallas when people asked about Austin food. I was so proud of my training dataset. Turns out I grabbed a bunch of web scrapes without checking the sources. Half my data was from national chains and other cities. The AI was learning the wrong patterns because I fed it junk. Now I manually go through every source city by city. Has anyone else had their model completely miss the mark because of bad training data?

2 comments

2 Comments

the_elizabeth26d ago

Did you try training it on Reddit threads for Austin food instead?

alex52425d ago

Raw Reddit threads would be way too much nonsense to sift through. People on there will fight about whether Torchy's is "real" tacos or not for 400 comments. You'd probably end up training a bot that gets into arguments about queso and brisket instead of actually giving good recommendations.