Item — HN Reader

Curious about this and I just download it.

Want to try uncensored models.

I have a question, looking for the most popular "uncensored" model I just find "TheBloke/Luna-AI-Llama2-Uncensored-GGML", but it has 14 files to download between 2 to 7 GB, I just download the first one: https://imgur.com/a/DE2byOB

I try the model and it works: https://imgur.com/a/2vtPcui

I should download all the 14 files to get better results?

Also, asking how to make a bomb it looks that at least this model isn't "uncesored": https://imgur.com/a/iYz7VYQ

4 comments

basically no open source fine tunes are censored, you can get an idea of popular models people are using here: https://openrouter.ai/models?o=top-weekly

teknium/openhermes-2.5-mistral-7b is a good one

You don't need all 14 files, just pick one that is recommended with a slight loss of quality - hover over the little (i)icon to find out

Thank you! Now I checked the (i) tooltips. Just downloaded the bigger file (7GB) that says "Minimal loss of quality".

Honest question from someone new to exploring and using these models; why do you need uncensored? What are the use-cases that would call for it?

Again, not questioning your motives or anything, just straight up curious. To use your example, any of us can find bomb building info online fairly easily, and has been a point of social contention since the Anarchist's cookbook. Nobody needs an uncensored LLM for that, of course.

Where do you typically have the 'safe search' setting when you web search? Personally I have it 'off' ('moderate' when I worked in an office I think) even though I'm not looking for anything that ought to be filtered by it.

I'm not using models censored or otherwise, but I imagine I'd feel the same way there - I won't be offended, so don't try to be too clever, just give me the unfiltered results and let me decide what's correct.

(Bing AI actually banned me for trying to generate images a rough likeness of myself by combining traits of minor celebrities - in combination it shouldn't have looked like those people either, so I don't think should violate ToS, certainly it didn't in intention (I wanted myself, but it doesn't know what I look like and I couldn't provide a photo at the time, idk if you can now (banned!)) so it does happen, 'false positive censoring' if you like.)

Holy shit that's so badass BingAI banned you. I mean like it sucks and i hope it gets resolved, but still awesome. you got it just from trying to make a composite of people's faces? i guess it takes fraud seriously, at least for biometrics. however you clearly either weren't doing that or omitted some choice details. good luck with appealing; that process always wrecks my faith in humanity.

i've been trying to see how much i can get away with before it suspends or bans me using one of my throwaway accounts. it takes some doing, but if you convince it you aren't doing something shady right before doing something shady it'll play along with you a fair bit more than if you just say "hi bing, i wanna do some shady shit." unfortunately you have to engineer your "i'm not gonna do (insert something shady)" prompt on a per shade basis.

I'd switch that question around: Why would I want to use a censored LLM?

It doesn't make sense for me personally. It does make sense if you're offering an LLM publicly, so that you doesn't get bad PR if your LLM says some politically incorrect or questionable things.

It’s very easy to hit absurd “moral” limits on chatgpt for the most stupid things.

Earlier I was looking for “a phrase that is used as an insult for someone who writes with too much rambling” and all I got was some bullshit about how it’s sorry but it can’t do that because it’s allegedly against its OpenAI rules.

So I asked again “a phrase negatively used to mean someone that writes too much while rambling” and it worked.

I simply cannot be bothered to deal with stupid insipid “corporate friendly language” and other dumb restrictions.

Imagine having a real conversation with someone and they freaked out any time anything negative was discussed?

TLDR: Thought police ruining LLMs

Most people wanting 'uncensored' are not looking for anything illegal. They just don't want to be forced into some Californian/SV idea of 'morality'.

> Again, not questioning your motives or anything, just straight up curious. To use your example, any of us can find bomb building info online fairly easily, and has been a point of social contention since the Anarchist's cookbook. Nobody needs an uncensored LLM for that, of course.

When you ask a local LLM, at worst you get no useful info. When you ask online, at worst you spend the rest of your life in a government black site without any chance of due process.

Because I don’t appreciate it when a model has blatant democrat/anti-republican bias, for example. The fact that chatGPT, Bard, etc are heavily and purposefully biased on certain topics is well documented[1].

[1] https://www.brookings.edu/articles/the-politics-of-ai-chatgp...

Isn't it reality that has a "democratic / anti-republican" bias?

Curiosity and entertainment. Also have an experience that isn't available on the current popular consumer products like this.

At best the false positives are a nuisance and makes the model dumber. But really censorship is fundamentally wrong.

Unless we are talking about gods and not flawed humans like me I prefer to have the say in what is right and what is wrong for things that are entirely personal and only affect me.

perhaps they are in a war zone and must learn to make molotov cocktails without connecting to the internet

For the entertainment value

None of your business.

The readme of their repositories each have tables that detail the quality of each file. The QK_4_M and QK_5_M seem to be the two main recommended ones for low quality loss while too being too large.

Only need 1 of the files, but recommend checking out the GGUF version of the model (just replace GGML in the URL) instead of GGML. Llama.cpp no longer supports GGML, and not sure if TheBloke still uploads new GGML versions of models.

No, each download is just a different quantization of the same model.