Serious question, has anyone else tried training a small AI model on a single, weird dataset?

I was trying to get a simple text generator to write better descriptions for my online shop. The normal advice is to feed it tons of varied data, but that wasn't working for my specific niche. On a whim, I gave it only the exact product descriptions from my own best-selling items, about 150 of them, and nothing else. After two days of training, the output was suddenly spot-on for my brand voice. It stopped making generic statements and started copying the specific, detailed style I use. I'm curious if this focused approach has worked for others with very specialized tasks, or if I just got lucky.

3 comments

3 Comments

john6501mo ago

My experience says more data is always better.

victorhill1mo ago

That "more is always better" idea shows up everywhere and it's just not true. My closet is full of cheap shirts I never wear because I bought a bunch on sale. More stuff just means more clutter to sort through. It's the same with streaming services having thousands of shows, but most of them are bad. You just waste more time looking for the one good thing.

the_eric1mo ago

Depends on the data quality though. Bad data just makes the mess bigger.