[Loranubi]

Since all input is run through a tokenizer, I would expect the tokenizer space doesn't change a lot between one trained on uncompressed vs one trained on compressed data.