6
Serious question, has anyone else compared fine-tuning a model versus prompt engineering for specific tasks?
I spent a week trying to get a base GPT model to write consistent product descriptions just by crafting prompts, and the results were all over the place. Then I spent a day fine-tuning a smaller model on 50 examples, and it nailed the tone and format perfectly on the first try. Has anyone else found that fine-tuning is the real game-changer for getting reliable outputs?
3 comments
Log in to join the discussion
Log In3 Comments
elizabethhayes1mo agoTop Commenter
Read a case study that proved the same thing.
4
jessica_hall491mo ago
Case studies can be interesting, but they often focus on extreme examples. The real world is usually a lot more messy and less clear-cut. It's hard to know if the findings apply broadly or just to that one specific situation they looked at.
1
lane.cameron1mo ago
My team wasted three months on prompt engineering for a customer service bot. The prompts got so long and complex they were basically code. We finally fine-tuned a model on last year's actual support tickets, about 2000 of them, and it started giving usable answers in two days. It felt like we were trying to steer a boat with a broom before that.
2