Commit

5d20a56

I also tried ByteLevel. With this computational cost, tokens were better, but when switching to bytes, attention had the biggest drop.
Translated from Japanese

commit 5d20a56