That raises a more general issue: Text scraped from the internet may be duplicated to multiple sites with few changes. How does the team behind GPT-n ensure that the corpus is maximally deduplicated before splitting into train and test corpora?