I finally spent 12 hours training my own chatbot to stop answering questions with paragraphs

I accidentally trained it on a bunch of old sales emails and it took me a whole weekend to figure out why it kept asking people for their credit card numbers in every response, has any developer dealt with a bad training dataset that wasted their time like this?

3 comments

3 Comments

derekward23d ago

Oh man, that's brutal. I once spent a whole week trying to debug a recipe bot that kept telling people to add "a pinch of the souls of the innocent" to their chili because the dataset got contaminated from some fantasy forum. It's amazing how one bad training file can just derail everything, especially when you're already 10 hours deep and your brain is fried. The credit card thing is hilarious in hindsight but I bet it was maddening when you first noticed it. Solidarity from someone who's been there with a bot that would only respond with full on corporate jargon from old press releases.

james_kim23d ago

The spooky part is those corrupted datasets probably make more engaging bots than the clean ones do lmao. Would people actually prefer a recipe bot that occasionally demands sacrifices over one that's perfectly polite?

davis.noah13d ago

Ha, this is one of those things where I gotta disagree with @derekward - those "corrupted" datasets actually make the bots way more entertaining and people remember them better, so maybe a recipe bot that wants souls is exactly what we need for engagement.