Is the fine-tuning effort value greater than few-shot prompting?
A younger pediatrician and a famend doctor, who would deal with a child’s caught higher?
Though each are docs and may deal with a baby’s cough, a pediatrician is a specialist who can higher diagnose a child, isn’t it?
That is what fine-tuning does to smaller fashions. They make the tiny, weaker fashions clear up particular issues higher than the giants, which declare to do the whole lot underneath the solar.
I used to be not too long ago in a state of affairs the place I needed to choose one over the opposite.
I used to be constructing a query-routing bot. It routes the consumer question to the proper division, the place a human agent would proceed the dialog. Beneath the hood, it’s a easy textual content classification activity.
GPT-4o (and the mini one) does it extremely nicely, but it surely’s inflexible and costly. It’s a closed mannequin, so you’ll be able to’t fine-tune it in your infrastructure. OpenAI gives fine-tuning on its platform itself, however that’s too pricey for me.
Coaching GPT-4o prices $25/1M token. My coaching information was shortly a number of million tokens. Plus, serving the fine-tuned fashions prices about 50% extra than the common ones.