Hacker News Reader

Comment by isoprophlex | Hacker News Reader

So, this interacts with someone else's hosted model, or something open source you might host yourself. But these models are still more or less censored.

What does it take to build something like this that is completely sexually unshackled, that can do explicit things? I guess simply a lot of training material and money?

View on HN

Ragdoll is based on RAG - the knowledge is scoped to documents you define, and is uncensored, even if the model you're using has censorship.

For example, if you provided some raunchy URLs as Lara Croft's source of knowledge, you could have realistic conversations about the content there, however sexually unshackled they may be. To test it, I just searched for some known banned words in character.ai and tried them in character.ai. It refused to answer, tried to convince me to not talk about it, and even clarified there's a policy against these words, breaking the 4th wall. I asked the same thing to a character in Ragdoll and can confirm they will speak on it like any other knowledge you provide.

Original RAG paper: https://arxiv.org/pdf/2005.11401.pdf

Thanks for asking about this, I see this come up a lot on character.ai discussions elsewhere, and I'm going to do a NSFW case study about it soon!

Simple general purpose uncensored models without any specific dataset are already pretty good with that kind of stuff.

Interesting; would you have any specific recommendations for an intelligent (high parameter count? Something that can hold a conversation ...), uncensored, specific model that is able to do this?

If anyone is looking for a cloud option (e.g. don't have the hardware, just want to pay someone else to handle the compute while still writing about generally uncensored things like a violent RPG adventure or something explicit in the adult sense), then I guess NovelAI could fit the bill: https://novelai.net/

I guess most people focus on their image generation which is pretty cool for anime characters and things like placeholder art (say, in gamedev), but their text generation options are also pretty nice and can be used either through their web UI, or something like SillyTavern: https://github.com/SillyTavern/SillyTavern

If you do want to run something locally and have the hardware, I think people discuss models in a variety of places, like https://old.reddit.com/r/LocalLLaMA/comments/16425mk/best_ns... (quick search, was curious, NSFW discussion) and it seems like there's plenty of models on HuggingFace, though I haven't particularly looked into them.

I will absolutely say that even when I was trying to run a general purpose chatbot LLM, it was kind of impressive how quickly things are moving ahead in regards to LLMs: all of these GGML, GGUF and other formats, being able to use even consumer GPUs (though I couldn't get ROCm working with my RX580) and experimental attempts like using Vulkan.

I think most custom fine-tunes and merges on HuggingFace will do this unless they specifically mention it being censored. Even the lower param models have been surprisingly good, with relatively fast progress being made in the 7b and 11b models.

My "daily driver" is Fimbulvetr v2 11b, surprisingly slapped together by an EMT. Kunoichi 7b seems to be a pretty popular model too. These can be run locally with as little as 8 GB free RAM (preferably VRAM) with an easy install solution like LMStudio or Faraday.

You can generally find a lot of recommendations in places like SillyTavernAI or LocalLLaMa on reddit:

https://old.reddit.com/r/SillyTavernAI/comments/1brig2n/whic...

https://huggingface.co/fhai50032/BeagleLake-7B-Toxic-GGUF

This one works with GGUF-compatible llama.cpp wrappers like the fully-local, fully-private Layla app or the other examples tossed around. With some prompt tweaking it's capable of a broad variety of... tasks.

This model is very compliant and doesn't tend to go to "I'm sorry, as a large language model..." type replies.

Download ollama and use one of the dolphin models. You can literally do it in two commands:

     curl -fsSL https://ollama.com/install.sh | sh
     ollama run dolphin-phi

That puts you into a chat with an uncensored LLM. "Phi" is the smallest model though.

Just wanted to add that ollama doesn't support text-to-image yet, ragdoll implements stable diffusion's txt2img separately. I tried that Phi one, it was very bad, not even close to ready for primetime. My favorites from ollama so far are mistral for most uses and alfred for highly instructional use cases. I can't even try some of them because of the requirements, but some of the apparently high-end ones I have tried just aren't good - unusably so. Mistral is pretty awesome, glad they are fully open.

It takes peeling yourself out of your gaming chair, getting a gym membership and trying to make something of your real life.

Reply on news.ycombinator.com