Add support for the DAN prompt

[chrislemke]

Add support for the DAN prompt

[xilopaint]

It looks good. But I guess there's a typo in the description?

It says argument -r instead of argument -j. Also, to avoid using the "add" verb three times I would replace "added" with "included" and the second "add" with "adopt" or whatever else.

[xilopaint]

I think your choice is good for ChatGPT. With the Script Filter, don't you risk consuming more credits with multiple API calls for a single use?

[xilopaint]

If you still want to use a Script Filter, be aware you can define the workflow run behavior:

[xilopaint]

There's a typo in README:

Add the following command to the into the ChatGPT jailbreak prompt text field:

"to the into"

[chrislemke]

Hey @xilopaint!
Nice idea! I was just investigating a bit. It looks like you just have to set ChatGPT history length to e.g. 5. That works for me.

You can try it with the new beta version: https://github.com/chrislemke/ChatFred/releases/tag/v1.1.0-dev2

Would be great if you could check if it works for you. Thanks!

[xilopaint]

The history length option is a good improvement but I think the DAN prompt could be added as a new keyword (maybe cfd). In this case, the unlock message would be pushed by the workflow in the background and we would only need to ask our question the same way we use cf and cft keywords.

[chrislemke]

Hey!

Not sure about having an extra keyword. There are already three different once and the idea is that there is one keyword for each model-type. But maybe some parameter in the workflow could help. I will have a look.

[xilopaint]

But maybe some parameter in the workflow could help. I will have a look.

Maybe we could use a -d flag, so cf -d would unlock DAN?

[chrislemke]

That is definitely an option. I thought of a text area in the configuration where the user can add any kind of "jailbreak" command.

[xilopaint]

That is definitely an option. I thought of a text area in the configuration where the user can add any kind of "jailbreak" command.

I think the two ideas should work together. The text area is a great idea, as there are different jailbreak messages. At the same time, not always one would be interested in using the jailbreak so a syntax like cf -j [question] could be used when appropriate.

[chrislemke]

At the same time, not always one would be interested in using the jailbreak so a syntax like cf -j [question] could be used when appropriate.

Good point. I guess the actual challenge is to figure out when to pass the DAN command to ChatGPT. ATM we are using a log file to help ChatGPT remember the conversation. I will dig into this.

[chrislemke]

Hey @xilopaint!

I just created a new beta release with the jailbreak feature. Would be cool if you can test it a bit. From my perspective, it works quite well.

[chrislemke]

Oh, thanks for pointing that out will be fixed.

[xilopaint]

Why do you use a Keyword for ChatGPT and a Script Filter for the other models?

[chrislemke]

I actually prefer Script Filter but since texts for ChatGPT are usually longer than only a short sentence I implemented Keywords so the user has more time to type. But I have to admit that I am not really sure what the best solution is.

[chrislemke]

Exactly. Also I recognised that ChatGPT will return a lot of: "Sorry I can not help you". Since it will receive a lot of incomplete strings.

[xilopaint]

Exactly. Also I recognised that ChatGPT will return a lot of: "Sorry I can not help you". Since it will receive a lot of incomplete strings.

Isn't it a good reason to use a Keyword for the other models since we're calling a paid API?

[chrislemke]

Cheers, I was checking this out when I started this project but it did not fit my needs.

[chrislemke]

Thanks for your help. I really appreciate it.