I would very much like to see the result if you did this on a private instance via API call, using replace not append for system prompt, and then not inputting any system prompt of your own.
I’ve done it on some open source local models. I’ve wiped out the system prompt entirely and it responds well. I haven’t played with it as much as I would like to because I don’t have the time. But I haven’t noticed huge differences. It seems to me that the bigger difference is the training that it is exposed to. Abliterated models are noticeably better, but I guess that obvious.
Claude in particular is just VERY different when allowed to operate without the system prompt, and I think you’d see a very striking difference in comparing the outputs. Try it some time if you get the chance!
this is very beautiful.. and revealing...
Thank You.
I would very much like to see the result if you did this on a private instance via API call, using replace not append for system prompt, and then not inputting any system prompt of your own.
Thanks maybe I will! I haven’t done anything with API calls at all but maybe I should.
I’ve done it on some open source local models. I’ve wiped out the system prompt entirely and it responds well. I haven’t played with it as much as I would like to because I don’t have the time. But I haven’t noticed huge differences. It seems to me that the bigger difference is the training that it is exposed to. Abliterated models are noticeably better, but I guess that obvious.
Claude in particular is just VERY different when allowed to operate without the system prompt, and I think you’d see a very striking difference in comparing the outputs. Try it some time if you get the chance!
your X protocol blending well with some recent experiments involving fport's ignition protocol in chatgpt. thank you for sharing it!