Unimpressed by Phi-2

#39
by Aditiyadav - opened

I fine-tuned Phi for a simple task of recommending books given a query (training data from Kaggle)
Received suboptimal results during inference after changing prompts in 10 different ways.

It's a base model so you have to multi-shot prompt it, not prompt it in the normal way. It's also best-in-class for its size, but its size is quite small. It's just a little fellow.

I used few shot prompting.

who is the 36th president of the US?

who is the 36th president of the US?

I am was not.

LBJ not Barak Obamma

@Aditiyadav can you please share your email ID & LinkedIn ID. I would like to know how you have fined the model. I'm also doing the same, but I'm facing few difficulties. So, kindly guide me. Thanks!

Sign up or log in to comment