Difference between text-generation and text2text? Canonical way to provide multiple demonstrations?

I have a sequence of demonstrations:

first demo => first output
second demo => second output
nth demo => nth output

Given a n+1th demo, I want the model to generate a candidate for the n+1th output.

Should I format this into a giant list separated by '\n', and append '=>' at the end to get this result? This is what I would do in the OpenAI Playground.

What’s the appropriate pipeline to use for this task? Is it text-generation, text2text, or something else? All data (both demos and outputs) is plaintext (ASCII). I’m currently aiming for gpt2-medium, which I will later probably have to fine-tune.

GPT3 (text-davinci-002 on OpenAI Playground) can do this task without fine-tuning, but I have my doubts about GPT2.