Right, it needs to be fixed at the model level.

I'm not sure if that's possible either but I'm thinking a good start would be to separate the "instructions" prompt from the "data" and do the entire training on this two-channel system.