Comment by FergusArgyll

> We filter our evaluation sets from our training corpus.

Page 5 of the report (they mention it again a little later)

https://storage.googleapis.com/deepmind-media/gemini/gemini_...