If you wanted to train on Lemmy data, just pretend to be an instance and have all the public instances push their data to you. No scraping required, and you get all the metadata and context you could possibly want