As Joop Eggen noted, using tokenizers.UnicodeScripts instead of ByteLevel solved the issue.
tokenizers.UnicodeScripts
ByteLevel