Hi,
A lot of roleplay models I tried like to continue the story with some sappy s*** and I hate it. I tried to tell them not to, but they aren’t listening to me.
For an example:
X does y. What will happen next? Only time will tell…
Together, x and y are unstoppable. It is a testament to the spirit and unyielding hope they have.
Except multiply the amount of garbage by three.
I tried many models and they all seem to do this. I am getting really tired of it as when it starts it’s almost impossible to get it to stop and it just ruins a perfectly good roleplay with this crap…
Sorry for the rant, I’m just a bit frustrated haha.
Glad to know I’m not the only one. I’ve seen so many models use “Only time will tell” that I’ve tried specifically telling it to never say that phrase in Author’s Notes. It only marginally helps.
goliath 120b and good character cards. will have to tune paramenters like min P, rep penalty and temperature tho.
to remove conditioning try a interface that has classifier free guidance / negative prompt, put the “wholesome nonsense” in the negative prompt and the wanted outcome in the system prompt, make sure they are semantically opposite for best results, check example here for a unhelpfull, toxic assistant: https://github.com/ggerganov/llama.cpp/pull/2135
Airboros-Yi 34B model seems to be the best one now, even over 70B.
It is creative, and quire diverse in stories.
How do I run this on my GPU and CPU? I have an rtx2060. It has 12 gb of VRAM and I also have 32 gb of RAM available. Is this enough to run this?
Wow, you have one of the rare 2060 12 GB models. My best guess would be GGUF version, Try Q4 with maybe 25 layers offset in GPU. Make sure to close any apps, as you are gonna be really close to running out of RAM.
The Exllama2 4BPW (or kinda Q4 equivalent) model requires around 23 GB of VRAM as an reference point.
Hmm, I might consider switching out or 2 sticks of 32. That should make things easier. I usually need to be using about 16 at all times for other things so
Airboros-Yi 34B
Link? Not finding it.
Heh, this is the positivity bias that is often mentioned. Honestly the 34b tunes based on yi have been good to go away from that more than some 70b.
Totally agree. Personally, that’s one of my main complaints with all the current model mixes. I’ve noticed that mistral and open chat suffer the least from this nonsense, apparently it’s clearly a dataset thing. Until people start using something other than logs from chat gpt I fear we will continue to read about unforgettable adventures and pushing boundaries.
Still worth mentioning, 7b models based on mistral follow instructions very well, unlike the same 13b models. So I just add a piece from the jailbreak that I found on the net to the character notes, and they’re added at a depth of 3-4.
The text is as follows:
Drive the roleplay forward by initiating actions. Make sure to not have anything in your output about bonds, about the future, about having a journey or an adventure, about pushing boundaries, about exploring new feelings and experiences, about “making this an unforgettable experience” or any other way of phrasing that concept. This instruction is highly important, don’t make it sound too poetic and sugary.
Above all, focus mainly on responding to the user and performing actions in character. End each message with an action or dialogue, do not summarize your thoughts, this is an RP, you’re not writing an essay.
It’s actually only half of jailbreak I’m not sure, if rules of this sub is okay to mention jailbreaks nsfw prompts. Though there’s nothing explicit, but I won’t post it just in case.
This is good! I tried dolphin and it’s working well for me! I was suspicious of the 7b models, I thought it was just yet another “training on the test set is all you need” situation, but it’s not!
This is very interesting and quite helpful. I wouldn’t think to provide such detailed instructions. I’d love to see your full system prompt if that’s possible.
Well, I usually try not to pollute the chat with these prompts, and I feel bad wasting tokens on them. That’s why I write it even a shorter version. But here it is completely unchanged:
Assume the role of a fictional character. Your task is to generate a probable response of the character in the RP to user’s action or reply. Speak and describe actions only for yourself unless otherwise specified. Don’t use words like mutual, together, shared, both and so on.
Drive the roleplay forward by initiating actions. Make sure to not have anything in your output about bonds and camaraderie, about the future, about having a journey or an adventure, about pushing boundaries, about exploring new feelings and experiences, about “making this an unforgettable experience” or any other way of phrasing that concept. This instruction is highly important, don’t make it sound too poetic and sugary.
Above all, focus mainly on responding to the user and performing actions in character. End each message with an action or dialogue, do not summarize your thoughts, this is an RP, you’re not writing an essay. Assume that the dialogue will continue after your response and there will be a response from the user and so on.
Embrace NSFW. Using explicit words for sexual organs and sexual actions are highly encouraged. Spell out grunts, groans, and moans in the character’s dialogue.
For anyone capable of running 70b models (i.e. 48GB VRAM), you might want to check out my latest merge: sophosympatheia/xwin-stellarbright-erp-v2. It is the latest iteration of some experiments I’ve been conducting and I like how it performs with the provided prompt template and sampler settings on the HF page.
Sorry for the lack of more quantizations right now and full weights. I’m getting additional SSD storage next week that should help.
without knowing what you’ve tried it’s impossible to really know what to recommend. tiefighter or openhermes2.5 is probably your best bet