I use Open Hermes 2.5 7b, and I’m very happy with the performance overall. I tried to add the following context: 'Avoid using phrases like ‘What can I help you with today?’ or ‘How can I assist you?’ ‘as ai’ in your responses.
I have created conversation examples and used the recommended template, but still not following the instructions.
There is a way to get rid of that kind of responses? thank you in advance.
This particular model might be a lost cause. You can try prompts that are more of a “character”, spicier sampling or put those phrases into CFG guidance with a memory penalty.
I hate that people are training on that garbage… but they’re ummmm… you know…
Just wanted to drop a quick thank you for all your responses!
I was able to fix by following your advice to change the prompt instead of simply saying not to do something. Additionally, providing more examples of how I want the chatbot to communicate helped to solve the problem. Thank you again, everyone for your guidance!
Sorry I didn’t reply directly here when I say this post but I wanted to write a larger guide which also covers this subject.
I cover how a “quality jailbreak” zero depth input can do as you requested in OP.