Not really, though. Sequences of tokens are predicted as {spam, not spam} via training on pair sets. It's like a token predictor where the output vocabulary is two tokens.