Loading…
32
Unpopular opinion: prompt templates beat fine tuning for most small teams
Played with both for a customer support bot at my last job and the template setup with like 15 examples handled 80% of tickets while the fine tuned model kept making up refund policies. Which side are you on here, or am I missing when fine tuning really wins?
1 comments
Log in to join the discussion
Log In1 Comment
blair_green3d ago
15 examples is your problem right there. That is nowhere near enough to teach a model your refund rules, which is exactly why it went off script and started inventing policies. Fine tuning wins when your task has hard rules the model has to follow every single time, like a refund window of 30 days or a specific tone for a legal notice, because you can bake those in with hundreds of labeled examples instead of hoping they stick in a prompt. Templates also eat up your context window with every call and get pricey at scale, while a tuned model is cheaper and faster per request once you are doing thousands of tickets a day. Your test sounds more like a bad data problem than proof that templates beat tuning.
6