#Which model is best for roleplay and doesn’t put out bad responses?

14 messages · Page 1 of 1 (latest)

sonic hull
#

Hey, uh…I need some help on trying to pick a good non-moderated Ai model to use for roleplaying. I’ve tried everything. Starting from Hermes 3 with 405B, to Mistral Nemo, to Llama 3.1 Nemotron 70B Instruct, to Meta Llama 3 70B Instruct, to Meta Llama 3.2 90B Vision Instruct, to Claude 3.5 Haiku (2024-10-22) which I know is moderated but I used it for the context, to Claude 3.5 Haiku (2024-10-22) with self-moderation that’s non-moderated, to Meta Llama 3.3 70B Instruct, to Mistral Ministral 8B, to Meta Llama 3.1 8B Instruct, to Google Gemini Flash 1.5, to Google Gemini Flash 1.5 Experimental, and lastly, to Qwen QVQ 72B Preview. I know this is a whole lot of models and words that I’m typing and speaking about with it being a mouthful, but I just need help on trying to find a very good model.

And basically, not just any model, I need a good Ai model that’s both for roleplaying and can handle a chatbot that’s over 60k tokens. The site I use with these different OpenRouter models is Chub Venus. Yet, no matter how many times I reroll a response or switch models, I get tons of errors without an answer from the chatbot coming through or if I do get a response, I either end up getting a response that’s super long despite my max new token being under a thousand that also tends to control my RP persona despite me having a pre-history prompt telling the bot to avoid doing so instead of don’t/do not/shouldn’t/should not/won’t/will not. Or…, I end up getting a response that’s full of gibberish nonsense despite my generation temperature being below 1 as I’ve had it at 0.70, which literally isn’t too low but neither is it too high. Or I even end up getting a response that’s disappointing with it being comprehensive yet it doesn’t follow the previous bot’s message correctly or even my previous message correctly (as a result of me using the google models).

IMG_6063.png
#

Either way, those are major issues I tried to fix by myself without complaining yet it didn’t work. I’ve attached an image to show one of the errors I’ve faced on Chub from trying to use one of the models I’ve tested. Not only that but, as I know the chatbot I’m trying to talk with has over 60k tokens, I obviously use OpenRouter models that’s at least 128k token context and over such as me using a 131k token context model or even a 200k token context model or even higher than that, yet they all fail regardless, despite them having much more space to handle an over 60k token bot. And basically, I love Hermes 3 with 405B as it used to work so well with coherent and not overly long responses since I used it way more than the other models yet I had to switch that back and forth with the Mistral Nemo model since Hermes 3 with 405B sadly became an incoherent mess when like I said, my temperature on Chub was at 0.70, not 1 or higher.

I don’t know what I’m doing wrong but I just need help to fix this issue and I’ve only got a few credits left that’s above negative numbers and above 0 to try and use another Ai model. Don’t worry, I’ll replenish my credits as I don’t need help with that, just that I’m frustrated on what model to use for roleplaying that’s higher than 60k tokens for the token context between chatbot and user persona.

#

If someone can come up with a solution to my problem, then let me know. I’ll be patient and wait for something to come up.

bitter pier
#

You must have something wrong in your setup, because I'm getting incredible RP from even Sonnet 3.5 with t=1, min_p=0.05, and a good system prompt
Sole issue with Sonnet, is it's incredibly expensive, and you should cache the tokens of your context to reduce a bit the price

sonic hull
#

I could report this to the Venus Chub server for Chub issues but this is still on topic when I say about the OpenRouter models.

sonic hull
#

You don’t have to share the system prompt if you don’t want to as I’ll try to keep making/using my own prompts or using other people’s prompts from here or Chub server. I just only wanted to see if you have a good prompt with the usage of {{char}} and {{user}}.

#

Also, I’ve just turned on logging in my OpenRouter settings but I turned off model training as I’m not too sure about that.

#

Hopefully that’ll work. The logging thing.

sonic hull
#

I’ve tried both 3.5 sonnet models with the normal newest one and the normal June 20th 2024 one and they both have filters that my own jailbreak can’t even get through as they are hard to get past by. Plus, Chub sadly doesn’t have the self-moderated models from OpenRouter so I can’t use sonnet at all, even with enough funds for it.

#

So, I’m stuck using other OpenRouter models above 60k token context.

bitter pier
#

I do not use Venus Chub