Loading…
19
Unpopular opinion: fine-tuning a small model for my local news project was a month of my life I won't get back.
I thought I could get a decent local event summarizer running in a weekend. The data cleaning and prompt engineering alone took three weeks. Has anyone else hit a wall trying to make a niche AI tool actually work?
3 comments
Log in to join the discussion
Log In3 Comments
evah406mo ago
Oh man, do I feel this. I tried to make a simple tool to sort my club's email questions into topics. The training data was a total mess, and getting the model to stop making up fake categories was impossible. It looked so easy in the tutorials, but the real work is just endless tweaking. I gave up after a month too.
3
elizabeth_chen816mo ago
Last year I spent three weeks just cleaning up 200 emails for a similar project. I found that having at least 50 examples per category stopped most of the made-up stuff. Even then, you have to check its work constantly, which is the real time sink.
-2