If that's an issue for you, you do the year 2018 thing and just train classification heads for a base model LLM. No instruct tuning means prompt injection is c

If that's an issue for you, you do the year 2018 thing and just train classification heads for a base model LLM. No instruct tuning means prompt injection is curbed. Classification heads means you get results off a single forward pass.