What are the best prompt practices for fine-tuning a T5

Models
1

Hello everyone, I am currentling working on fine-tuning the FLAN-T5 model for article highlight generation task. That task is kind of like summarization but not exactly the same. So I really wonder what is the best prompt I shoud use when fine-tuning? Should I just use “generate the highlights of the following texts” or should I discribe what kind of response I am looking like “Generate 2-4 sentence highlights.” for or even add few examples?

Thanks!

2

I think the prompt value itself for a custom/unique case like yours does not matter, as much as, the training data which covers the nuances of differences in your requirements than OOTB summarization task that model is trained on - vanilla summarization seq2seq VS instruction based prompts. Your input and ouptut text covering needed scenarios will drive the success and ability post finetuning process.