How to use few-shot learning correctly?
#5
by
0kopr
- opened
What is the most correct way to do few-shot learning in the prompting of this model.
Do I have to use a special token for it?
I don't think there is an official method. I make use of the <|system|>, <|prefix_begin|>, and <|endoftext|> tokens. I haven't tried few shot learning frequently because it doesn't seem super effective for me.
One think worth trying if you were wanting to answer the question in either "yes" or "no":
<|system|>Answer the questions with "yes" or "no"<|endoftext|>
<|prompter|>Is the sky blue?<|endoftext|>
<|assistant|>yes<|endoftext|>
<|prompter|>Is it ok to mock people?<|endoftext|>
<|assistant|>no<|endoftext|>
<|prompter|>
and then you ask your question followed by an <|endoftext|><|assistant|>.