The way people talk about AI training data gaps feels off to me

I keep hearing folks say AI models just learn everything from the internet and that's it. But after spending 6 months building a small dataset for a local history project in Austin, I realized how much context gets missed. Like, my model kept confusing a 1980s flood with a 2021 one because the training data had them both labeled as 'disaster' without dates. It made me wonder if most people even check where their training data comes from or just assume it's all good. Has anyone else run into this kind of timestamp issue messing up their outputs?

2 comments

2 Comments

michael_jenkins391mo ago

Have you tried adding a separate timestamp field instead of just relying on the text itself? That's what saved my bacon when I was tagging old newspaper clippings for a neighbor's genealogy project, the model kept mixing up events from different decades because the articles used the same "local hero" language. Once I made date its own column in the dataset everything started lining up way better, it's a simple fix that most people overlook.

hart.sage1mo ago

Did you try just binning the timestamps into broader eras before training, like 80s/90s/2000s? @michael_jenkins39's approach with a separate column makes a ton of sense for keeping things straight. A buddy of mine in Dallas had similar issues with his model mixing up two different tornadoes, and splitting dates out fixed it for him too.