T
21

The feedback that made me stop overcomplicating my AI training data

I was feeding my model massive datasets full of every edge case I could think of for a project in Austin. A buddy who works at a startup in the Bay Area told me my precision was actually hurting recall because the model was drowning in noise. Cut my training data by 40% focusing on the core 3 use cases, and my F1 score jumped from 0.68 to 0.81. Has anyone else had to unlearn the more data is better mindset?
3 comments

Log in to join the discussion

Log In
3 Comments
brian_ramirez
Tbh I had the exact same wake up call. Spent months cramming every possible scenario into my dataset thinking I was being thorough, then my metrics actually got worse after I trimmed it down. Cutting the noise was what finally made the model actually learn what mattered.
3
caseym48
caseym481mo ago
Less is more with data, isn't it? Cramming everything in just buries the signal under a pile of useless junk. That trimming step is the real secret sauce.
2
james_singh7
Does trimming down to just one core case work better than keeping three?
2