The three delimiters are <unk>(token_id=0) , reserved_0(token_id=3), reserved_1(token_id=4)

[<unk>]sytem_message[reserved_0]user_message[reserved_1]response

@BitVoyage The tokenizer cannot separate reserved_0/reserved_1 from English text.
Why not add <> to <reserved_0> and <reserved_1> for easier identification?

Good suggestion! We have fixed it in the next version.