SanjiWatsuki

/

Silicon-Maid-7B

like 93

The model is extremely good at conversation and following instructions <3

but, it still has some assistant in it and some nsfw avoidance

it often try to avoid explicit language (such as using derriere a lot) and sometimes starts lecturing and tries to change topic away from explicit
it also talks a lot about consent and making sure everyone is alright all the time

as most assistants it also apologizes a lot and has a hard time talking back to you as a character would normally

Been running in in a discord setting for a month now, it beats many 13b maid models outright,
i wish we could somehow improve it, by lowering the nsfw avoidance and censorship training it has, coming from some of the data/base model in the mix

Most 13b NSFW maid models have another problem, they go straight into some bdsm stuff and are way too thirsty
need something in between, not bdsm heavy, more willing to engage in nsfw, more explicit language and lower censorship/preaching, then it would be perfect!

awesome model overall, cheers!

Did you find one thats just as good with less censorship and less nsfw avoidance?

Did you find one thats just as good with less censorship and less nsfw avoidance?

No, it is still one of the best models,
I have since replaced it with WizardLM2, the 'semi-uncensored' version before they took down al the new Wizard LM2 models
wizard lm2 is also censored, but given the right prompts it will not shy away from things and is very engaging
I found wizard LM2 to be much better for the type of chats I wanted, rather than Meta's Llama 3 which seems to avoid absolutely everything and write very poor stories
Mistral 0.2 was also not an upgrade form Silicon Maid, and also not the dolphin 2 version
perhaps if Wizard LM2 or Mistral 0.2 is trained with the Silicon Maid training set it could create something superior
my best answer for you to check out is the Wizard LM2, i use the 8 bit quant from https://huggingface.co/bartowski/WizardLM-2-7B-exl2