NSFW models on HuggingFace chat?

#650
by MarvelSanya - opened

Hi all! A long time ago I fell in love with HuggingChat (online chat) for its ability to use text models without downloading them to your PC. At that moment I especially liked the Mixtral model, which was uncensored. Over time, models were updated and replaced with newer ones.
At the moment we have good models, but the problem is that almost all of them are censored and do not allow NSFW content. I am more than sure that many users would like more opportunities to communicate with their favorite chat.
In my case, I often used HuggingFace for role-playing games (dark fantasy with adventures in the DnD style).
Unfortunately, the only model that allowed me to play with NSFW content was Command R. But with updates, it became worse and ignored most prompts, taking away control from the player (Command R is simply not suitable for role-playing games and completely takes control away from the player over their character). Other models are simply censored and with them it is almost impossible to get the chat to describe how your character died in the game, brutally dealt with the enemy and other adult topics.
Anyway, as someone who doesn't have the ability to use text models locally, I'm curious: is it possible to add uncensored models to HuggingFace chat in the future? And do other users need this?
Thank you in advance for your feedback! P.S. Sorry for bad English

I think the problem is that the setting gradient of the repeat penalty is 0.1, so it is impossible to set it to the optimal value of 1.05.

In addition, I think the 2023 version of cR+ is much better than the current version in role-playing. Maybe it will be more interesting to bring command-r-plus-v01 come back.

In addition, I think the 2023 version of cR+ is much better than the current version in role-playing. Maybe it will be more interesting to bring command-r-plus-

I completely agree here. The version that is now in the hug chat practically does not allow the player to play and simply ignores prompts, taking away control of the player’s character. The model perfectly describes what is happening and generates events, but at the same time it places more emphasis on the narrative without taking into account the fact that the user has autonomy in the role-playing game.
The old version didn't have such obvious problems with this.

Now, Deepseek R1 is the best. This is beyond my expectation. I thought that this Chinese company's model would be have a rigourous censorship. Maybe they are considering cooperating with Sex Doll manufacturers? 🤣

Now, Deepseek R1 is the best. This is beyond my expectation. I thought that this Chinese company's model would be have a rigourous censorship. Maybe they are considering cooperating with Sex Doll manufacturers? 🤣

Let's test! Today, Command R is the best model in many aspects for role-playing games and stories (of those that are uncensored). Later I will write about my impressions of the new Deepseek R1 model.

Now, Deepseek R1 is the best. This is beyond my expectation. I thought that this Chinese company's model would be have a rigourous censorship. Maybe they are considering cooperating with Sex Doll manufacturers? 🤣

the model is very good and after testing it for several days I came to the conclusion that the current model CohereForAI/c4ai-command-r-plus-08-2024 is much better than the Deepseek R1 model for several reasons.
Firstly, command-r-plus describes events and scenes much more colorfully and creatively. Secondly (and more importantly), command-r-plus follows the prompts and never censors its responses. In contrast, Deepseek R1 is very bad at this and constantly tries to minimize, reduce or avoid content that it understands is being censored. Even with obvious prompts, she can ignore them and refuse to answer.

As a result, I think that command-r-plus is now the best model in HuggingChat for NSFW content including stories, descriptions, role-playing games, etc. Unfortunately, it has its drawbacks (since it often takes away control from the player during the game) but, despite them, it is better than Deepseek R1.

Sign up or log in to comment