How to train FLAN-T5 to summarization task with a custom dataset of legal documents in pt-br?

So, I would like to create a small proof-of-concept using (already extracted in txt files) +- 4.000 legal text divided in:

1. 2.000 initial petitions / complaints *.txt files
2. 2.000 summaries of each initial petition (txt files too)

PS.: all text files are in brazilian portuguese (pt-br)

So how can I use these txt files to train a new transformer able to generate new summaries (using flan-t5) ?

from How to train FLAN-T5 to summarization task with a custom dataset of legal documents in pt-br?