Stop sequence for few-shot learning with GPT-J on HF API

I am trying to find a a way to declare a stop sequence and show the model a few examples of what I want to generate, i.e.:
prompt_tweet = “”"
key: market
tweet: take feedback from nature and markets, not from people

key: children
tweet: Maybe we die, so we can come back as children.

key: hugging face
(Example taken from few-shot learning guide for GPT Neo)

The problem is that I cannot find any indication to such function in the documentation, but I know that is possible to implement that, there are some websites available that generate text with GPT J few shot learning.

Any help?