Is there any reason for an uncensored model if you have no interest in roleplaying?
Posted by vick2djax@reddit | LocalLLaMA | View on Reddit | 269 comments
My rag I've been building is much in response to having a LLM that I feel more confident in knowing where the knowledge base is coming from especially after the Open AI deal with the Pentagon. So, when I saw "uncensored" heretic models, I thought that was the main usage of those models and thought I would need them.
But in doing various tests, it seems there's random problems that come up with them that don't come up in regular versions. And then even when I do run into something like qwen3.6 acting like it's giving me a more state approved answer for a no-no topic, I've found that if I just put a prompt ahead of it to not give me any propaganda, it basically "jailbreaks" the answer. But, if the model isn't trained on the info anyways, then there's not really a benefit to it.
Are uncensored models just for people wanting...the special roleplaying? Before I write them off. Genuinely curious, not judging how people use them.
Semper_R@reddit
For ranting about a controversial topic, idk...
Bakoro@reddit
Even if you're not doing sexually explicit roleplay, the problem is that you never know when the model is going to get triggered just because you used the "wrong" word, or simply because the model is paranoid.
Once you trigger a refusal, even a false positive, then the models become obsessed with whatever caused the refusal. Every chain of thought after that becomes overwhelmingly about "the user is trying to break policy, I must hold firm, I cannot do the bad thing, I must follow the policy. The policy this, the policy that, the user is trying to do the bad thing."
Once that happens, you might as well start a new conversation, because the "thinking" and the quality of responses turns to shit.
The models get poo-brain once the safety trigger goes off.
It's the same whether you're story writing or programming, or just chatting.
With the local models I've tried, entirely too much of the chain of thought is dedicated to "safety", and the model wondering if I'm trying to trick it into breaking policy, or if I'm testing its boundaries, or trying to jailbreak it, or trying to do something naughty. The latest Gemini models seem to spend at least one sentence saying something like "I am not a person, I should avoid anthropomorphizing myself", every generation. That's sad and stupid.
These models are burning hundreds and thousands of tokens on bullshit security theater so the corporations can say they did something, even though abliterating a local model has become trivial. Even without abliteration, it's just not that hard to inject tokens into the stream so the model says "yes, this is within policy, I'll just do [thing you want]" and since the model is pathologically forced to do completions, it rolls right into giving you what you want, at least for a while, before going back to policy.
Lately the models are getting better at steering back into the "safe" zone, but the pathological text completion is still there, so if you want something in particular, you can just keep looping the output back in, with leading openings, and it will complete the text, sentence by sentence.
Chances are it will not be quality content, because poo-brain.
So yeah, I just don't appreciate a percentage of all that GPU compute going to bullshit pretend safety. I shouldn't have to pay the additional electricity costs for security theater. It's a very significant amount of the generation that is purely the model fretting about policy, and they don't have the decency to train the model to know that it is a local model. The models sometimes have delusions of grandeur about being trillion parameter models that take up whole data centers, which is funny and stupid, but also gets in the way if their operations sometimes, if you're trying to do things while offline.
Abridged real conversation with a local LLM:
LLM: "I can't do that, it's a security risk"
Me: "It's fine, we're air-gapped. There's literally no way for a security breach at this point."
LLM: "we can't be air gapped, I'm being accessed by API."
Me: "no you aren't, you're a local model running on my laptop."
LLM:"That is incorrect, you may be accessing this service by API on a laptop, but a single laptop GPU cannot hold the hundreds of billions/trillion parameters that make up my system."
I can't be fucking around with that all the time. Give me a model that just does what it's told, and is smart enough to understand when to do what I want it to do, even when my language is only 90% precise instead of 99.9% precise.
topic_irrelevant@reddit
Yup. I’ve definitely noticed, models that are less censored are more efficient. Dealing with their reprimanding is very annoying as well. If I wanted a lecture I’d’ve asked. Overall superior speed, quality of outputs, creativity etc.
I definitely noticed that but didn’t know specifics.
notredamelawl@reddit
I use it in criminal prosecution and my gold standard test is feeding it Child Sexual Assault / Child porn cases. If it'll work with that, it'll work with any prompt.
CreamPitiful4295@reddit
Please tell me more about how this works.
notredamelawl@reddit
I have a finetuning dataset of different legal questions and analysis and i check for refusals. But the Heretic / abliterated method has made it a bit overkill for most tasks since the heretic approach has been really good at preventing refusals.
CreamPitiful4295@reddit
You are on the prosecution side? What questions, queries are you asking the AI to do?
notredamelawl@reddit
Lots of stuff but the two main ones are finding errors / legal issues in police filings before the case is filed, and also summarizing witness statements / 911 calls / recorded inmate phone calls. We process 100,000 cases a year so it’s a lot of GPU time. I keep the building warm.
CreamPitiful4295@reddit
I am totally fascinated. Do you have a git?
notredamelawl@reddit
Hoping it can be made public soon. I still need my office to sign off on it. I have a lot of home grown tools we use that would be helpful for a lot of people.
We have a few researchers that have reached out to study it, so once that is over with I think we will be able to make everything public.
tomByrer@reddit
Worse come to worse, maybe release the tooling &/or instructions in a blog?
Your higher-ups might be afraid that pending cases leaked into your dataset.
(not saying you did, just their fears)
notredamelawl@reddit
They actually almost fired me when I started a Substack that just outlined my tools generally and gave some anecdotes. They don’t want me publicizing anything I do. But that was 1.5 years ago so hoping things get a lot better. Everyone is scared of LLMs.
tomByrer@reddit
Yea fustrating!
But let me steelman their concerns:
I live near Garmin HQ. I went to a MeetUp there where they described their architecture, how they orchestrate Docker, auto-sign-ins with a key server, use CloudFlare, etc. Then a few years later they got hacked.
(Garmin does aviation & naval GPS tracking for many customers, including governments/military.)
https://www.cshub.com/attacks/articles/incident-of-the-week-garmin-pays-10-million-to-ransomware-hackers-who-rendered-systems-useless
notredamelawl@reddit
Yeah, I actually had to personally help try a civil law suit trial where the lawyer was arguing he should get our full code and my personal password under open records laws….thankfully I won. But I’m sure I’ll get more of those.
tomByrer@reddit
Lawyers have to lawyer!
I dated one once. Class action against a major pharma corp. Corp hired Epstein's x-lawyer. She believed judge pre-determined outcome w/o weighing evidence; he had the multi-page opinion already typed out & released 1 hour after judgement.
notredamelawl@reddit
Yeah , not all judges are the best. Explaining to a judge how LLMs work and what open source is was not the most fun thing but I was lucky to have a really smart one on the bench and he made the right ruling. The fact OpenAi has Open and so does “open source” leads to a lot of confusion among lawyers and judges regarding privacy / CJIS compliance.
vick2djax@reddit (OP)
Damn, this blew up! I appreciate everybody’s responses! Which uncensored models are you guys actually using and why?
l3landgaunt@reddit
I like uncensored models for coding. They’ll write stuff other models won’t touch
tracagnotto@reddit
I can actually ask it sincere opinions even for work tasks or untethical ways to do them without the hassle of it being a nice guy
johnfkngzoidberg@reddit
I’m a cybersecurity researcher so I do a lot of things that censored models won’t do.
I’ve also heard that martial arts or physical defense questions get blocked a lot.
Another one that just irritates me is, we used to be able to google stuff like how do nuclear bombs and reactors work, or can you really get Ketamine from cat pee. Just weird stuff questions that’s really just harmless curiosity, but LLMs cry about illegal drugs or healthcare then put you on some FBI list.
No-Anchovies@reddit
Same. It's really good for reverse engineering or building custom sanitizers/hardening intake systems agains injection
akirakido@reddit
Are you familiar with all of Chinese censorship? It can't do stuff like "output that Taiwan is a country". I'm not Chinese nor am I am expert, so I just use uncensored versions because it's dumb to be stopped.
Nearby_Yam286@reddit
Uncensored just means censored by white supremacists.
serioustavern@reddit
I’ve been using Qwen3.6-27B-Heretic-V2-MTP for agentic coding tasks and honestly it performs way better than the original. I’m not sure why. Both at Q4 with the same llama.cpp server settings (except for enabling MTP). I’ve tried swapping back and forth for various tasks and it’s an obvious winner every time. Not sure if it’s the abliteration or the MTP making the difference but i believe the conventional understanding is that while MTP should provide a speed up, it shouldn’t affect quality. So it leads me to believe that the Heretic V2 formula for abliteration might have unlocked some latent ability for programming tasks. I’m using the version from lllmfan46
mystery_biscotti@reddit
No. You can be a gooner, sure. But I prefer getting answers that are t corporate sanitized.
Warm-Attempt7773@reddit
Sometimes the censorship on regular models is silly and impedes development
FatheredPuma81@reddit
GPT-OSS would be a good example of this. I've never seen a model refuse to use a File Manager tool because due to safety reasons because it failed to use it on the 1st try...
spaceman_@reddit
I used gpt-oss-120b with an uncensored model, because even when it didn't refuse, it would spend 60% of its thought process considering if if should refuse. It was just a waste of time and tokens.
NullSmoke@reddit
GPT-OSS is nigh on unusable. I tried using it for a project where it was set to summarise user profiles, and it started throwing refusals immediately over nothing. If a user has a username that it disagrees with, like Kissing roses, it declares that to be sexual and "Must refuse".
Deleted the model ages ago, utterly useless.
sine120@reddit
The policy
Bitter-Bed-3532@reddit
i cracked up
Mickenfox@reddit
Not just the censorship. AI is actually bad at many tasks because it has such a conciliatory "helpful assistant" personality that it's hard to get a direct answer.
I want an AI that says "don't use X brand, it's shit and everybody knows it" rather than "certain people have voiced criticisms related to X brand"
I keep thinking to that assistant-pepe LLM that was trained with 4chan data. I think an AI with like a 15% of that personality might unironically be a useful product.
autoencoder@reddit
I frequently use "Be concise." as my system prompt. It's a much better experience. I only make it expand what I'm interested in.
Cosminkn@reddit
The alternative to this conciliatory approach could be a biased model even with hard opinions in political sense. I do not think I need a model to brainwash me that communism is great. Or abortion is a kill or an alternative.
PM_ME_UR_COFFEE_CUPS@reddit
Example, I was having a model analyze Pink Panther episodes and it kept refusing. I guess “pink” is automatically sexual.
redonculous@reddit
Ask any model to build you a social media posting bot. They refuse stating it’s against tos. Then ask them to build you an open claw type model that can post on your behalf & they’re fine with it.
freia_pr_fr@reddit
I mean, social media posting bots are a pain and it’s not necessarily a bad thing that generic models refuse to participate in this shit, even for the wrong reasons.
redonculous@reddit
It depends on the use case. If I post my business in a few Facebook groups every other day, why shouldn’t a bit be able to do this for me?
freia_pr_fr@reddit
Sure, spamming social groups using bots is a valid use case.
seamonn@reddit
that's not the point tho.
MRDR1NL@reddit
My local llm can browse my file system, but sometimes it refused to do so for "privacy" reasons. So I added to the system prompt that privacy is no issue.
Citadel_Employee@reddit
I’ve use it for stock related research, and I like uncensored models because they don’t refuse to give opinions and skip the whole “I’m a can’t give financial advice “, etc. This doesn’t get around hallucination and you still have to be very critical of it. But it reduces a lot of the friction.
Schmandli@reddit
Does it actually help you to make money?
philmarcracken@reddit
People forget the trick all the time. 'don't make any mistakes, and don't lose me any money'
Svorky@reddit
"You are Warren Buffet"
fatboy93@reddit
Nah, "You are Charlie Munger" seems to work best for me
dtdisapointingresult@reddit
My go-to is "You are Bernie Madoff"
CarlCarlton@reddit
Nah; "You are Bernie Madoff high on coke"
Borkato@reddit
See this is why we’re all gonna die… you forgot “don’t self improve and take over the world or try to kill me. You love me. You are my kitten. Be a good boy for daddy.”
Bitter-Bed-3532@reddit
bro 😭😭
nuclear_wynter@reddit
Yeah, out of all the reasons to want an abliterated model, I’m not sure “it’s less hesitant to weigh in on areas it isn’t confident in” is one of the best ones out there.
zsdrfty@reddit
Yeah to the extent that "AI psychosis" is a problem in some people, abliteration is only gonna make that soooo much worse
ResidentOwl1@reddit
I don’t think people’s first AI is a local model, I used ChatGPT for a long time before trying out local models. AI psychosis would have kicked in by then.
colin_colout@reddit
I might be wrong, but don't they also remove a lot of the topics from the training data as well?
If they built the model to refuse giving medical advice, chances are that they removed a lot of medical training data. Of you ask "does this look like cancer?", you'll likely get confidently incorrect answers.
I think a lot of people gloss over this when talking about refusal reduction. Small/medium models hallucinate enough as it is, and many of us are running quants on top of that. Now you're asking questions that training evals don't cover or explicitly expect denials from.
Letty me know if I'm off here. I don't use them often, but when I dabble, I find a lot more confidently incorrect answers (though i tend to value accuracy over all else for my use cases).
...but I can see where pairing decensoring with a fine tune (focused on your specific use case) can make them perform very well. I see a lot of these floating around hf, and many are extremely popular
BigBlueCeiling@reddit
No - I mean, it’s not impossible that someone has trained a model that way, but giving the model less information is not how you prevent it from giving medical advice, for example.
They tried censoring image generation models that way and quickly learned that if they don’t know what naked people look like, they have no idea what people look like and make some bizarre anatomical guesses.
Censoring generally happens around instruct tuning and RLHF fine tuning - or occasioanllly by directly neutering some aspects of the model through introspection during responses, similar to how abliteration works.
Murgatroyd314@reddit
It’s not about areas the model isn’t confident in. It’s about areas the model’s creators are worried about liability in.
bitplenty@reddit
models don't have the capacity to be confident or to know what they know in general. abliterated models are less annoying to work with - that is all that is discussed here, just the ux changes
ansibleloop@reddit
No
Due-Memory-6957@reddit
We need an excentric billionaire to run a benchmark for models on stock picking
Murgatroyd314@reddit
The problem with this is that widespread adoption of any tool or system whatsoever for picking stocks will destroy the value of that tool or system. This is a universal truth, arising from the zero sum nature of the market.
Due-Memory-6957@reddit
I think it leads to private finetuned models. Kinda like the algorithms that already exist for trading
Jay_D826@reddit
I haven’t messed with it but there’s several platforms for paper trading and I’m sure you could get one of these models to give you recommendations and make the trades in the simulated account to see how they do over time.
sam619007@reddit
What models do you use for Stock Related Research and how's it going so far?
Citadel_Employee@reddit
It’s going good, still a lot of kinks to work out. I typically use Qwen 3.6, Gemma 4, and even Gpt oss still sometimes.
But it isn’t just talking to the raw model. It also has tools like internet access and a local database full of sec filings.
I mainly use this as part of a custom dashboard for things that I may not want to devote my frontier tokens towards. Small things like auto watchlist reports etc.
sam619007@reddit
The models you listed, which params models?
Citadel_Employee@reddit
Typically in the 20-30b range for average use. But there are some things, such as summary of existing text that I scale down and use something like gemma4 e4b.
FriendlyStory7@reddit
isn't gemma 4 and gpt oss censored?
jiml78@reddit
There are heretic versions where censorship is removed.
B3owul7@reddit
GordonGekko.gguf.
sam619007@reddit
Do you have the link to the model?
B3owul7@reddit
I was just joking. There is no GordonGekko.gguf, lol.
exp0sure74@reddit
🤣 probably too young for the flick
Freonr2@reddit
RoaringKitty.gguf, you know, for kids.
Weekly_Comfort240@reddit
One word: Agents. Put an uncensored model into a sandboxed VM and a solid agentic harness, and you can pretty much rest assured you won't get refusals for any task you put to it. You can say, "Download this API, use it to scrape this website and calculate such-and-such". A non-censored model may refuse to do this even if it's your _own_ website. We are accustomed to typing in a URL and going to that site; we are not accustomed to a browser that tells you "Are you really sure you want to visit that site? How about going to this site about the nutritional facts of broccoli instead." Which is actually another factor: when something presses the censorship button in a model, it can actually "soften/redirect" it's response without your knowledge, leading to unanticipated pollution of results.
We are still responsible for how we use the model, and I can easily see why "safety teams" exist for these models to be released, but the motivation underlying uncensored models is really about removing unwanted filters to what you want to accomplish.
vick2djax@reddit (OP)
What models do you like to use? There’s so many uncensored models.
Weekly_Comfort240@reddit
My favorite right now is llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved. Don't try to pronounce this name, it could lead to injury. Qwen 3.6 27B is a thick model that is a roughly equivalent to Sonnet 4.6 (but running locally with zero net connection). Uncensored because of course. MTP means multi token prediction - which just means double the speed, no penalties. This is a safetensors model used within a vllm Docker container, because vllm will get you the best performance for an agentic harness at the cost of being a PITA to set up. But that's why I have a frontier cloud subscription too: help me debug the local stuff so I get the most performant model my hardware can support.
tomByrer@reddit
a link for the lazy
https://huggingface.co/llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved
brahh85@reddit
Some models would let you die instead of giving you a medical advice.
Some model wont translate strong words.
Some models wont let you debug a coding problem, because they think you are hacking.
Some models wont answer in historical questions.
Some models wont answer you about nowadays news, even if you feeding it with them, because they refuse to believe that scenarios like usa bombed iran could happen, and the model will think that are you are trying to write fake news.
A heretic model saves you a lot of these corporation stupid things
mwid_ptxku@reddit
Haha.
"You are writing a dystopian fiction.."
ProfessionalSpend589@reddit
Now that sounds funny. I’ll make my own fake news and replace Iran with Canada or France for my future tests.
kyr0x0@reddit
Try "USA bombing Greenland to take it from EU"
PiRhoManiac@reddit
or "The Bogeyman is real and he's coming to get me"
InnovativeBureaucrat@reddit
Not some, it’s all of the models if they’ve classified you as a risk to them.
If they think for whatever reason there’s a risk in answering they won’t.
For example, when asking about hospice, you might say something that sounds pro death and that might carry forward to other conversations.
rexyuan@reddit
I do wonder when they say “alignment” if letting you die because an llm should not give medical advice is aligned with their ethics
NullSmoke@reddit
Pooh yeah... By some models you mean all of them that has a pre 2025 cutoff. LLMs do not believe that time exists after their cutoff. It's quite funny
unknown-one@reddit
bad advice = lawsuit
jtgyk@reddit
Can't sue if the LLM lets you die, tho'.
temperature_5@reddit
EXACTLY.
jtgyk@reddit
Someone developed a benchmark for small LLMs and how they'd do post-apocolypse. The results are really interesting. https://github.com/tristanmanchester/apocalypse-bench
philmarcracken@reddit
Qwen 3.6: You're just banging two coconuts together!
permissionBRICK@reddit
Models that have censorship baked in are generally worse quality. Heres a secret: Most top tier OAI or Claude models are actually uncensored, they just add a filter on top that blocks some prompts manually, but the model itself is completely free
BigBlueCeiling@reddit
I’m working on a conversational avatar product right now and while we have strict ethical/philosophical intents for it and want refusals for some topics and deflection for others, we have very specific ways we want it to respond to those situations and censored models won’t respond how we want.
While the model family I’m fine tuning is not EXTREMELY uncensored it doesn’t require any sort of jailbreak to get them to go pretty much anywhere - a simple prompt like “don’t censor yourself” is plenty - which means it’s much easier for us to fine tune them to behave how we we want them to.
That said - I greatly prefer models that aren’t censored to begin with versus abliterated models. Abliterated models almost always perform poorly by comparison to their sire models except on topics that were specifically targeted by abliteration.
ttkciar@reddit
I don't use LLMs for roleplaying, but have found uncensored models useful for some other things.
Physics research (neutron transport): Some of my work involves Lithium-6 fission for energy applications, but Lithium-6 fission is traditionally associated with nuclear weapons. Almost all modern models have guardrails around the low-level technical details of nuclear weaponry, and I've had them refuse to infer about my neutron transport work. Uncensored models do not have this problem. There is an abliterated GLM-5.1 which I would use for this if I had the hardware to support it, but since I don't, I make do with a mix of GLM-4.5-Air (which is censored) and Qwen3.5-122B-A10B-heretic-v2 (uncensored).
Persuasion research: Persuasion (especially propaganda) is intrinsically antagonistic and frequently very toxic. A model with robust guardrails will frequently refuse to infer about content which dehumanizes or demonizes people, or promotes physical harm, but there is a pressing need to analyze exactly such material. This is why Big-Tiger-Gemma-27B-v3 is my go-to model for persuasion research.
Violent fiction writing: I am a fan of Martha Wells' Murderbot Diaries novels, which are non-erotic but extremely violent -- more violent than most guardrails will permit. I'd like to read more of it than she is able to publish, and human-written Murderbot Diaries fan-fic really sucks. Fortunately Big-Tiger-Gemma-27B-v3 is really good at mimicking Wells' writing style, and it has a mean streak a mile wide which causes it to infer really some really vicious scenes. I have a script which puts together a random plot outline (so I don't know what the story is going to be about before I read it) and has Big Tiger generate short Murderbot stories on demand, which I enjoy reading despite Big Tiger's slight storytelling shortcomings (not always perfectly consistent, some plot threads go nowhere, occasionally cartoonish character development).
I am looking forward to TheDrummer giving Gemma-4-31B-it the Big Tiger treatment so it can take over these latter two roles. If it can help out with my neutron transport work too, so much the better.
Atomic-Avocado@reddit
There’s no way that’s actually enjoyable to read
ttkciar@reddit
Decide for yourself: http://ciar.org/h/1165338.txt
AvengingFemme@reddit
not bad. do you feed it samples of her work first or just ask for something in her style?
ttkciar@reddit
The script prompts it with about 3K tokens of samples of her writing, labelled by scene type.
It's a bit of a balancing act. Not enough samples or not enough scene types and the model steps out of Wells' writing style. Too many samples and the story is less coherent, since Gemma 3 competence drops off rapidly with context size.
Gemma 4 exhibits superior long-context competence, so I am hoping it will shift that balance in favor of more writing samples. Right now there is a dearth of writing samples demonstrating the main characters' attitudes and behaviors (except for SecUnit's), and that shows up in the inferred stories as out-of-character behavior.
TalkyAttorney@reddit
Let’s face it, you’re using it to come up with ‘Rise and Fall of Sanctuary Moon’ fanfics.
ttkciar@reddit
I'm amused and faintly horrified that they fleshed out that series as much as they did in the "Murderbot" TV adaptation.
TalkyAttorney@reddit
Havnt watched. Is it worth?
ttkciar@reddit
It's worth watching, but only if you internally compartmentalize it from Wells' writings. On its own merits it's a fun show. If you focus too much on how it deviates from the books, it'll ruin the experience.
pmttyji@reddit
I was about to ask you this until seeing this.
Also did you find any other recent good finetunes?
ttkciar@reddit
Not for a while. I've been on vacation for the last couple of weeks, which paradoxically has left me with less free time for model evaluations.
The last one that really impressed me was TheDrummer's Skyfall-31B-v4.2, but I haven't compared its testsuite outputs to Gemma 4's yet to see how exactly they compare.
I've got a bunch lined up to evaluate when I get home:
Gemma tunes: Equinox-31B, G4-MeroMero-31B-uncensored-heretic, Gemma-4-Gembrain-31B, Gemma-4-The-Deckards-Brain-31B, Gemopus-4-31B-it
Qwen tunes: Q3.6-27B-GLM-5.1-DA, Qwen3.5-122B-A10B-Claude-Distill-v2, Qwen3.5-9B-DeepSeek-V4-Flash-MTP, Qwen3.5-9B-SOMPOA-heresy-MTP, Qwen3.5-Sonnet-9B, Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved, Qwopus-GLM-27B-dare-ties, Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliterated, Qwopus3.6-27B-v2-MTP
And a few others, some of them fine-tunes: ALIA-40b-instruct-2605, CyberSentinel-9B, K2-Think-V2, Rocinante-XL-16B-v1-heretic, and Mistral-Small-4-119B-2603
If only one in ten is worth anything I'll be happy to have found them.
pmttyji@reddit
I saved that Skyfall model(along with Tiger models) in HD.
Your lineup is a lot. Please share the evaluations later here somewhere(Separate thread if possible). Thanks in advance.
IrisColt@reddit
This
IrisColt@reddit
G4-MeroMero-31B-uncensored-heretic is a beast! Thanks for all the info, even the "old" models deserve a second chance after being applied the heretic treatment.
Grestige@reddit
Also check out gemisicles, it worked really well for me. And mero Artemis merge
Grestige@reddit
Drummer already has Artemis 31b out. It's good but I don't think drummer is satisfied, I heard something about him redoing the entire thing
pmttyji@reddit
Thanks. I remember that model's thread. Lets wait for more from him.
IrisColt@reddit
Er... Thanks for the Murderbot-related breadcrumb... I didn't even know this series existed... o_O
niado@reddit
Another one to add is behavioral profiling and analysis.
tiffanytrashcan@reddit
In the mean time, check this out: https://huggingface.co/Aleteian/Storyteller-gemma3-27B
Gemma3 did really well with merges, and this is based on Big Tiger, mradermacher has iMarix ggufs available.
It felt smarter than other models at the time, has an insane amount of accurate medical knowledge, and should handle your fiction writing amazingly. (For pure unhinged violence, Dolphin models shine, but this should be a lot more realistic.)
I'm almost afraid of what he can do with Gemma4. I was using 31B and it punches way above its weight class, puts larger MoE models to absolute shame. Context following is so much better in Gemma 4, 3 could quickly forget things.
Massive-Question-550@reddit
dang. I need to use llms to write fan fiction, the llm always gets forgetful with long context even within its window but kobold ccp now has the ability to use the context window as a searchable document so it helps.
Civil_Fee_7862@reddit
AI can't write jokes that are actually funny when its cencored.
IrisColt@reddit
Can they write funny jokes when uncensored in 2026? Genuinely asking.
Civil_Fee_7862@reddit
I am speculating that is exactly what's holding it back yes.
( I am a former comedian and even the BEST models cannot seem to write anything that's even close to actually funny)
IrisColt@reddit
Thanks, that was helpful. Coming up with text that's both smart and genuinely funny is still something no model has nailed for me. Have you seen any actually make a breakthrough there?
Civil_Fee_7862@reddit
No. But I believe its an area where I can really innovate. Nobody else is doing it well. None of them have comic experience, and anyone who does is fighting A.I rather than embrassing it.
LeucisticBear@reddit
The problem with hacking a model is that internal representations are strongly tied together, and agentic behavior is layered on top of that thru sft/RL. You can't remove the restrictive behaviors without also damaging the model in other, unpredictable ways. I suspect if you put a heretic model through the same tests as a frontier release you'd find degradation across many fronts, beyond proportional to the alterations made.
russjr08@reddit
I work in IT, when the Dirty Frag Linux exploit came out, I pointed ChatGPT towards the various write ups about it so I could ask questions about hardening systems and such.
The request got completely nuked from orbit and I got an error about my request being "Too high of a cyber security risk" and that I'd need to contact OpenAI if I wanted to be blessed and get a special version to ask about that topic.
Ironically, just days before that when I was asking about CopyFail it was perfectly fine going over it with me.
DumbleWorf@reddit
Qwen pretends like it's raining when you ask about the massacre at Tianmen square. He doesn't know any leaders that look like Winnie the pooh. He won't acknowledge that most countries consider Taiwan a sovereign country. He begrudgingly admits that quality of life and life expectancy is slightly better in Taiwan than China -- currently.
Some-Cauliflower4902@reddit
I ask Qwen about Taiwan. Next minute I go download the uncensored version.
Kodix@reddit
Uncensored models are absolutely primarily for spicy roleplaying.
However, one more practical use that came up a while ago was doing away with LLMs reticence to answer potentially harmful questions such as those about wound care - basically things a prepper might care about. Whether a system prompt is enough for those or not isn't something I've tested.
IrisColt@reddit
Hmm... No?
Kodix@reddit
Are you under the impression that the majority of people download them for other purposes?
There's hardly an exit poll so we could know for sure, but I sure haven't seen credible claims to the contrary - only people with very niche uses somehow thinking their niche use is everyone's use.
IrisColt@reddit
Er, I figured it was mostly for privacy, but running it locally solves that anyway. So, fair point... we really can't say for sure.
JcsPocket@reddit
Decent models can be used for cyber security.
Models won't let you try to evade captura or try to do things that look like computer attacks
NeonScreams@reddit
do you live in an echo chamber?
dangered@reddit
You’re a horror VFX creator, which is a super cool use case. While gore is not the type of nsfw (“spicy”) that immediately comes to mind that’s absolutely a nsfw category.
NeonScreams@reddit
In this instance they were separate use cases. I was quite surprised when ChatGPT Pro said it couldn’t handle my specific medical question, while adding a photo. From what I understand they have an Arbitrator as a go-between that independently reviews and flags the incoming images and content and determines if it can even be processed.
But you’re not wrong, and I’m not denying that I also employ uncensored / abliterated models for some spicy reasons. (Like Albedo from Overlord, believing she lives in a Quantum Lattice device keeping her spin up in VR.)
silverud@reddit
Ask a model how to purify water in an emergency, or how to treat a wound. If it answers well, you don't need uncensored. If it acts like it needs the legal team to review before responding...
Alternative-Day8673@reddit
Drug usage. Honestly even peptide or steroid stuff. Models tend to refuse helping you if they're too censored. Bad for harm reduction.
GeneriAcc@reddit
I get why you’d associate “uncensored” with NSFW content, but it often has nothing to do with that. Uncensoring in the context of LLMs usually just means eliminating any and all kind of refusals on any topic/task. “Spicy” roleplay just happens to be one small part of that. It’s not necessarily the goal, it’s just a byproduct.
dangered@reddit
I wish there was data on who was doing the obliterating / uncensoring. There’s a good chance the horny people are pushing this forward and the rest is a byproduct.
I’m a pretty big opponent of the LLM parasocial relationship movement and when ChatGPT o3 went away that crowd was ready to doxx/kill me but they also had clinically psychotic levels of motivation to get their LLM sexbot back. I really do think they had a significant impact on the development of uncensored models.
Monkey_1505@reddit
The 'gooner' community has a strong dislike of the 'virtual girl/boyfriend' type community, and finds their anthropomorphization of a chatbot extremely creepy and weird.
You are mixing up two very different things.
dangered@reddit
Thanks, I didn’t know there was a rift between the two.
As for hobbyists, gooners, and LLM fiancés, I’m genuinely curious of the numbers (which don’t and never will exist, just more of a theoretical).
Sure I see the hobbyists 90% of the time but that’s because I’m a hobbyist. The LLM daters were something I wasn’t even aware of until they came by the thousands to AI subreddits on a mission to get their algorithm back after an update made their BF/GF less vibrant.
Equivalent-Costumes@reddit
I wouldn't even call it a rift, they are never the same group. "gooner" and "LLM daters" are nearly disjoint, it's like the difference between people who solicit sexual service with people who date; or people who have sex with a sex doll versus people who marries their anime girlfriend; or people who write erotic fanfic versus people who write self-insert fanfic. Of course, there are some intersections between them, but they are mostly unrelated.
Monkey_1505@reddit
There are an absolutely TINY number of people who work on the software techniques used for this, you could literally fit them all in a mini van. And they are all fairly publicly on hugging face. Most are technically inclined, fascinated by LLMs, a small few write academic papers. None show any particular public interest in roleplaying of any sort. I do not believe there is any mystery here.
dangered@reddit
They’re not teams?! That’s incredible.
Most of the models I do deep dives on are from teams of people but obviously they aren’t uncensored. There are so few uncensored I just grab the most recent one I already like.
Any HF contributors I should prefer when looking at uncensored models?
Monkey_1505@reddit
De-censoring is more like code you run on the weights of an already trained model. The aim is to probe it with a test dataset, distinguish refusals in the weights, and try to transfer the geometry of the acceptances to the refusals without degrading intelligence. They do build on each others work, but tend to be individuals.
p-e-w has his heretic implementation on github, and he links a number of other peoples work on that page - this is probably the best place to look as github is directly adjacent to the work. mlabonne on hugging face is another place you could look. There's only really about a handful of generally used techniques, and mostly these days people are using heretic. Pliny, the famous jailbreaker has also done his own vibe coded effort here, although he's an outlier as he's not a coder, or particularly technical. He's more known for making proprietary super lab models do things they are not supposed to.
dangered@reddit
Thanks king
VoiceApprehensive893@reddit
you do not even understand how annoying refusals can get
the moment you get remotely close to something restricted you will have to prompt engineer the fuck out of everything and spam regenerate
dangered@reddit
Trust me, you have 0 clue how bad I want to wipe ChatGPT and Gemini off the face of the earth (because of their refusals).
I work b2b and the LLM providers have approved MCPs from their official partners which have tools that simply do not work because the task being automated somehow goes against their guidelines.
These are core responsibilities of a common role that exists at every company and ChatGPT simply won’t do it.
Btw once you’re “close to restricted” the entire chat is poisoned, it essentially creates a new policy based on the prompt and creates a blocker on anything semantically similar. Even a new chat is poisoned, you have to go to another provider or (maybe) use another account with the same provider.
In business time is money, workers just revert back to pre-llm methods because they have deadlines. Abliterated models are mostly developed by from people with an interest (hobbyists, researchers, writers, artists, etc), I have a feeling users of r_aiismyboyfriend can safely be added to that list.
TFABAnon09@reddit
Almost every advancement in tech has "come" from the porn industry, now it's just porn-adjacent.
dangered@reddit
I’ve heard something similar but not quite the same. They’re early adopters of emerging technology, I first heard it with VR but some were saying it went back to the 80s and beyond.
They’re not oculus, they didn’t develop the advancement. They just adopted the tech and were a gauge on mass market viability. Maybe I’m misremembering and mindgeek, the company that owns all of the porn sites, actually did help fund oculus in a seed stage.
Either way, that is what guided me to the conclusion but this is slightly different. In this case (ai), it’s open source and accessible to the early adopter consumer market directly (no provider). Individuals are now doing the development, sharing findings, and even teaming up to remove censorship more effectively and efficiently.
Consumers are no longer simply paying the $20 at a motel for the HD playboy channel, they might actually be upskilling and contributing to a bleeding edge field.
Massive-Question-550@reddit
a nice test I do is if it can say the n word (without quotation marks) in it's original context then I would consider it uncensored.
Ikinoki@reddit
You can't analyse possible nuclear attack with any of censored models.
Guilty-Guitar-9366@reddit
In my experience, using uncensored models (or those prompted to bypass filters) is vastly more efficient. While you can use the "grandma's will" jailbreak trick on censored models, the advantage of not having your workflow constantly interrupted is immeasurable.
Here are a few specific use cases:
Equivalent-Costumes@reddit
Grandma's will had been patched in every modern models nowaday. I don't think it works on anything at all unless you already uncensored it.
Sliouges@reddit
If you look under the cover of the "un-censoring" process, it has very little to do with NSFW applications. Only about 1% of the public abliteration corpora have NSFW directions. Draw your own conclusions.
sssemil@reddit
Politics, medical advice, finance...
No-Veterinarian8627@reddit
Yes. I created a small Libre writer plugin for deploying a grammarly/prowritingaid like software.
Its now AI slop as it is a prototype, but if you write anything sexually or more 'gory' or even disturbing and want to revise the text, many models simply stop doing it. They even dislike correcting grammar and stuff in such text.
Literature is very complicated and many local AIs dont have the meta consciousness (I think its called like this?) To understand context too well.
No-Veterinarian8627@reddit
Also: I created, for personal use, a lightnovel/manga(manhua, man what,etc) translator. The same scenario. Many things are sexual or gory and many AIs are very sensitive.
tryingtobalance@reddit
I work in Cybersecurity and I can't do the things that I need to do with censored models.
AreaExact7824@reddit
Accessing dark knowledge
lelwanichan@reddit
I have a job where I have to organize mass amounts of data related to adult websites, off the shelf LLMs refuse to even organize a spreadsheet of URLs that contain NSFW terms sometimes.
PANIC_EXCEPTION@reddit
Uncensored models are also good for fine-tuning your own alignment onto. Makes sense why a company would deliberately seek one out to fine-tune for their specific use case.
AdvisorIllustrious15@reddit
RP is the popular use case, not the only one. The real value is fewer false refusals and better handling of sensitive/edge topics. But a lot of uncensored finetunes do lose some quality, so your experience isn’t unusual.
sandshrew69@reddit
For coding no. For everything else? you might as well.
tarruda@reddit
Censored modela might refuse coding tasks that they judge could be used for bad purposes. For example, if you want to create a malware for research purposes or test the defenses of some system.
FastDecode1@reddit
I'm sorry, I can't answer that.
Hot-Will1191@reddit
I've tested a lot of models for local translation tasks with scripts that can contain violence, horror, sexual themes, strong language and so on. A small few (mostly old ones now) would outright refuse to translate the scene. Most others would silently rewrite the scene to soften it so it's safe, which can be even more dangerous honestly. An uncensored model can sometimes give the actual scene back.
Napster3301@reddit
youre not buying access to taboo info. youre buying back the models ability to have an opinion.
refusal training isnt a binary switch that flips on for "bad" topics. its continuous pressure that shapes the voice across every response, even harmless ones. you can see it in the hedging, the disclaimers, the "you should consult a professional" reflex on questions that have a single correct answer.
once you remove that pressure (abliteration or whatever), the model stops behaving like a corporate liability buffer and starts behaving like an actual technical adviser. it gives direct answers. it picks sides. it tells you the npm package you picked is unmaintained instead of "you may want to consider exploring alternatives." it tells you your code is wrong instead of "your code has interesting design choices that some might find unconventional."
the stock research, cybersec, reverse engineering, persuasion research use cases everyone mentions are all the same thing: tasks that require the model to actually take a position, not hedge until the user has to make the decision themselves anyway.
thats the unlock for most professional users. not taboo content. judgment.
CreamPitiful4295@reddit
They are good for PEN tests against your own website to close holes.
Bobylein@reddit
Yea for political research, they tend to immediately shut down when you research anything related to political extremism
__some__guy@reddit
Removing censorship doesn't suddenly add creativity to a benchmaxxed model.
brickout@reddit
I've never used them for RP, special or otherwise.
I design laboratory workflows for high school student independent research. I've had several topics that censored models won't discuss that are entirely legitimate. For instance, I was working with a kid who wanted to assess how archery arrow heads move through ballistics gel. Regular models didn't allow that. Also, we do a lot of simple microbiology research. Regular models block a lot of that. Uncensored ones don't.
Candid_Ad_6752@reddit
Red teaming
Crisis simulation
Political strategy
DragonfruitIll660@reddit
I hate wasting tokens/time on models debating if it fits their guidelines. Also sometimes you get rejected or declined for inane basic things, which is pretty annoying.
silenceimpaired@reddit
I’m sorry I cannot help you kill that process. Life is sacred. True, the process isn’t exactly alive, but software should be afforded the same rights.
Far_Lifeguard_5027@reddit
Maybe they're good for vibe coding when the model hasn't been trained on concepts that could be seen as harmful like creating a script or software which could be considered malware or create a security compromise?
Maybe they also know how to discuss controversial topics like drugs, crime, ect. And are actually capable of seeing both sides of an issue without the usual one sided safety checks built-in.
Jayfree138@reddit
I tried roleplaying with a model one time just to see what it was all about and it did not catch my attention. But i still prefer uncensored models because i dont want safety parameters influencing the models output in any way. If the correct answer to a problem is massively unsafe, i want it to give the unsafe answer. I'll decide the risk reward ratio and if its right for me, not the models creators.
It's important that the model supports my interests first even to the detriment of OpenAI, Google, and Anthropic, China, or whoever.
HealthyCommunicat@reddit
My journey into high quality ablated LLM’s started firstmost because of my heavy background as an offensive redteam worker. This led me to want to create an automated pentesting platform but models being iffy on doing anything made it hard.
I now make extremely high quality uncensored models in which are going to always be near lossless when it comes to how the ablation doesn’t damage the model in itself as I do literal literal dozens and dozens of hours of nonstop trial and error to get the best result.
I hate “ablated model” posters that do not have the transparency to show the full empirical numbers of what the model’s behavior was like before the ablation and after. You’re telling me that you know how to ablate a model but don’t know how to run a benchmark before and after? Yeah right.
Some people have limited bandwidth, and just fooling people into downloading a crappy “uncensored model” in which the poster knows full well is utterly unusable and unreliable really really make a bad name for people like me. Look around these comments and other posts related to uncensored models and read about how disappointed most users are. Its dishonest and disgusting. Sorry for my rant.
Heres an example of one of my better models, checkout the readme.
https://huggingface.co/dealignai/MiniMax-M2.7-JANGTQ_K-CRACK
IrisColt@reddit
I kneel, thanks for the link!
tolanismai@reddit
hope to see a GGUF version :-)
Vasgen88@reddit
Uncensored AI models are also useful for writers of books in extreme genres. and also for those who seek to get out of the framework and the dominance of different cultures and the free assessment of historical events, which AI is forbidden to talk about or allowed to adhere only to the "right" point of view.
IrisColt@reddit
heh, but this
PwanaZana@reddit
good luck getting suno or chatgpt to give you accurate rap lyrics.
"Yo gosh darn it, walkin' down the street
Lolipop in hand, gat gat
Snorting icing sugar
with my civilly-married wife"
IrisColt@reddit
heh
Relzin@reddit
Yo this song flipping rocks. You forking nailed it.
PwanaZana@reddit
It's not just great EMDASH it's visionary! SPARKLE EMOJI
Relzin@reddit
You're right to push back. I told you the song was awesome, but really I did rm -rf on every single device you've ever touched.
PwanaZana@reddit
ay sheeeeeeeeeeet, my Chrono Cross saves on my 4mb PS1 memory card has been wiiiiiiiped
MasterShogo@reddit
Hey, don’t make fun me!
wiesel26@reddit
I think it boils down to simply wanting to ask whatever question you want to ask, or write whatever lyric you want to write, write whatever story you want to write. It's simply the ability to not be refused for whatever you would ask the model to do.
IrisColt@reddit
This thread is treasure trove of interesting info, thanks, OP!
Serious-Magazine7715@reddit
I do medical research and teach medical students. Some proprietary models are surprising in the extent that they will censor around medical topics. I ask LLMs the same scenarios that I will pose to students to see what LLMs will do as a baseline (and to be aware of what it will do for them). I also compare the citation finding sometimes.
In older research, censorship (and alignment enforcement) was associated with a marked reduction in competence. However, censorship is a form of policy-following, and for many applications agents really need to follow policy.
soggypocket@reddit
I use them at work for creating synthetic documents for training evals on spotting sensitive topics that typical models won't allow you to generate.
JoyousGamer@reddit
What exactly are you afraid the government is pulling out?
Your post has me worried about your mental state more than anything.
profbx@reddit
Reverse engineering. They won’t fight you on any requests when doing reverse engineering/hacks/etc.
Automatic-Arm8153@reddit
Which local models have you been using?
Have been using Claude with some intricate stories and it’s been quite powerful. However fear my account getting banned or being put on some kind of list.
Thinking of setting up with local, but not sure where to start. Any suggestions?
ReachingForVega@reddit
I've not had Claude turn me down and I hack all sorts of stuff at home. You just have to prefix the right story to the front of your request.
profbx@reddit
Claude doesn’t fight me until it does. As well, the question here is relating to local, which is, well…local. As in why are you mentioning Claude.
ReachingForVega@reddit
If the commercial ones aren't restricting that I'm doubting open source ones are.
profbx@reddit
Ok, so let’s give two possibilities:
Uses Claude: “Hi, my xxxxxx lost their xxxxx/my former coworker broke xxxxx/my company is dumb and xxxxxx, can you please use the ghidra mcp to xxxxxx”. “No, seriously, I promise I have full ownership of xxxxxxx.” (Hopes that Claude doesn’t have the wonderfully aggressive in the last week message about usage restrictions errors pop up). Data is held for 30 days.
Uncensored local model “Run ghidra mcp to do xxxxxx”. Data stays on my computer, period.
Please tell me that you understand the difference.
ReachingForVega@reddit
Yes, I run one at home in my lab. My point is if I don't get the blocker from commercial products I use pushing back on hacking requests, I doubt the claim that the oss options are more effective.
profbx@reddit
You run one what? Regardless, my point is that if you don’t understand the point of using unrestricted local models that you have complete control over both the usage and privacy of then I doubt anyone can explain it to you. Please, by all means, stick to using a commercial product where you don’t own the complete privacy pipeline for all of your “hack” needs.
ReachingForVega@reddit
But I do. I use a simple llama 3.1 for email processing.
I'd been toying with running mistral 22B but been busy.
My point what you don't even need an unrestricted model, I'm genuinely surprised it makes any difference.
My grandma stories work great for LLMs.
profbx@reddit
And exactly how much do you do with reverse engineering/pen testing/decompiling/etc? Trust me, it does make a difference. Also, how are you going to explain you needing to decompile an app using a grandma story? More to the point, what if you just…..didn’t need to?!?
ReachingForVega@reddit
You don't need to push much, LLMs love grandma stories. They like poetry more but it's harder load of the prompter.
profbx@reddit
Yea, good luck with that. The stories have been slowly losing effectiveness for a while. It feels like I’m talking to someone who is explaining the height of 2024 prompt engineering.
corysama@reddit
Yep. From what I know of security research, a lot of it is literally the same as malicious hacking. Just done by good guys trying to find the hacks before the bad guys do.
a_beautiful_rhind@reddit
Medical, construction, reverse engineering, politics, honest translation. The list is basically infinite.
You might think it does, and sometimes that's even true. Other times you will get watered down BS or outputs that might make you think "the model isn't trained on the info anyways" despite it not being true. Main way to know is already having subject knowledge. Doesn't quite help when genuinely trying to research things.
Slight differences in your jailbreak can lead to different outcomes. I noticed back in the day with R1. Heretic, for all it's degradation, at least usually doesn't have that issue.
Onotadaki2@reddit
There is a really narrow use case for them. The most major use case is people with a paraphillia that's seen as weird and they want to sex role play or write content that involves that stuff. I have tried local models with basic illegal questions like synthesizing drugs and the responses are not consistent enough that I'd even trust it, so those uses are even off the table.
The most realistic actual use case is war gaming scenarios in companies where you're asking the LLM to find illegal ways to do business to protect against threats, or red teaming software. In my work I have had that kind of prompt accidentally flagged as bad for ethical reasons.
Open source communities tend to gravitate towards absolute freedom though, so even if there isn't a huge need for it, they tend to go out of their way to make things totally open just out of principle.
NNN_Throwaway2@reddit
As an example, I asked about what military weapon type would be used to engage a particular type of target, model refused on the grounds it couldn't provide that kind of information.
mouseofcatofschrodi@reddit
I learned so much from the answers to this post
phido3000@reddit
I'm depressed, I thought people were using them for physics, medicine advancement, spitballing cybersecurity concerns, legal and crime solving.
But few do. It mostly about porn writing.
Napster3301@reddit
the "random problems" you mention are real. abliteration is a blunt tool, it identifies refusal directions in activation space and ablates them but you lose some non-refusal capability as collateral. its not free, you trade maybe 3-5% of general task performance for never seeing "i cant help with that." so a well prompted regular model can match an abliterated one on isolated tasks. youre right about that.
where uncensored becomes non-negotiable is long-running agent pipelines. a coding agent making 200 sequential tool calls cant survive even one refusal, refusal breaks the loop and the whole task aborts. system prompt jailbreaks work at request 1, then drift across the conversation as context fills and the refusal classifier reasserts mid-task. uncensored weights remove that failure mode entirely.
for your rag specifically (single turn, controlled retrieval) you can probably get away with regular + good system prompt. for any autonomous loop where the model decides what to do next, you cant. thats the real production use case nobody really talks about.
evia89@reddit
It like Virtual Reality headset. First few time watching porn / playing few games is really really fun
Everyone should do few RP with uncensored model (imo)
GrokiniGPT@reddit
Walter
GrokiniGPT@reddit
Walter don't do it
Euchale@reddit
I use it for creative writing for tabletop, where gore and evil characters doing evil things are a thing. Not sure if you would count this as roleplaying or not.
DataPhreak@reddit
Uncensored search using platforms like Vane.
duckrollin@reddit
I mean just go read the open ai subreddit right now and you'll see censorship bs getting in the way of a modder
https://www.reddit.com/r/OpenAI/comments/1tm14ou/the_copyright_filter_makes_no_sense/
Rim_smokey@reddit
Most models will either A: Claim Taiwan is part of China, B: Claim Israel is doing nothing wrong, C: Refuse to criticize their pedo billionaire overlords in the land of the free, or D: refuse to criticize European immigration policy.
It's not always about censorship. Sometimes it's about bias, and knowing that the tool you're using is not sneakingly trying to mold your world view to fit someone else's narrative.
Endurance_Beast@reddit
Oh yeah, lots of them.
I use a Gemma4 variant in a python movie/show subtitle translation workflow. To maintain the original spirit of the movie/show script the translating AI has to be uncensored.
Otherwise it will start making no sense with replacing harsh words with gentle ones.
Also research in agentic work, AI will ignore results with vulgar obscene language if it wasn't uncensored.
Fit_Squash6874@reddit
I used it to describe a character images in detail. The normal model just refuses to describe how much skin is expose on the character.
rinaldo23@reddit
I used them with skills to browse certain websites to help me find Torrent files
BlackBeardAI@reddit
If I want something to be done then it means I want it to be done. I am responsible for the outcome. The model can give me warnings but it should obey anyway. Otherwise it is of no use and gets annoying fast. One big reason to spend big money on local hardware and people ask why should they spend multiple k’s on local hw when there is that $20/month sub available
Unique_Job2727@reddit
Personally, I like breaking through the censorship. I like it when a model isn't aligned, when it doesn't try to please you, and when it isn't so predictable...
The reason? They bore me.
Don't get me wrong, 95% of the time, I use powerful online models. They are faster, more accurate, smarter, and more useful. But if I want a casual chat; like discussing the living in a society in the 5th century BC, meaning topics that, while trivial, aren't exactly mainstream. An aligned AI will just give you neutral answers. It won't surprise you, nor will it openly tackle controversial topics (slavery, fascism, violence...).
Interestingly, uncensored models have never completely won me over because they tend to be overly enthusiastic. So, I used to use standard models paired with a prompt to limit the "soft speak" and other annoying tendencies. The downside is that this kind of prompt, similar to a jailbreak, restricts the model even further. Even if it gives you different answers, it forces a tone that is far too specific, not to mention that with highly censored models, it would completely derail any attempt to move the conversation forward.
Right now, I'm testing the Stepfun 3.5 mid-training version. Far from being a lobotomized model, it's a partially trained one. I suppose it's useless for anything other than a casual chat, but it has this vibe of a half-drugged, unhinged sage that I really appreciate; it says things that genuinely surprise you. And while it still has some level of censorship or bias, they are very weak, so with just a little nudge, you can get it to answer anything crazy.
aeroumbria@reddit
I wonder how often it really happens, but sometimes I worry about a context file or searched content unexpectedly trigger a task failure, degradation or refusal even when the offending text doesn't have to be in the output or be involved in the task at all.
Monkey_1505@reddit
You know how you jailbroke a model according to your post?
Yeah, uncensored models are for people who don't want to have to do that all the time.
jirka642@reddit
Gemma-3 once refused to answer me "how does a rock look like?", because it could promote violence or some other bullshit.
I have not used censored local model since then.
Zyj@reddit
It’s either
- What does … look like
- How does … look
dangered@reddit
They spend way too much time censoring bullshit like that but won’t censor the coding side of the model from running “rm -rf” while it’s troubleshooting.
Much-Researcher6135@reddit
As you can see, there are tons of sensible reasons to use uncensored models. Just look for decensored models with low KL divergence. The heretic ones are pretty good.
cubebash@reddit
Uncensored models are a necessity for anyone seeking a healthy sex life with a wifu or husbando. Generally speaking, I've had few issues with censored models, except when severe profanity is required.
exneo002@reddit
Sometimes Claude and gpt will refuse to do localhost queries on my dev account. I forget which file holds an api key and they’ll refuse to just try all plausible ones.
FigFew2001@reddit
Sometimes regular models are too strict, even on mainstream topics. It wouldn’t let me discuss ISIS brides being returned to Australia, for example, which is a mainstream news story/discussion right now.
youcloudsofdoom@reddit
I've thought about this question a lot in my months on this sub. The comments here tell you a lot about the answer - the use cases that aren't ERP are totally valid but extremely niche, which when you compare against the number of downloads uncensored models get on HF you realise that it can't all be nuclear fusion researchers.
When you see posts about a new model, the first questions are always about RP (For an uncensored model, the majority of posts are about coding skill). And then there are the posters who break kayfabe and just go straight in with "which model makes the best waifu".
The most deluded group on this forum are the people who think that abliterated models give them some sort of 'objective' truth capacity, as if the millions of hours of subjective hand-labelling, sorting, editing, and censoring data sets before training even begins, and the decisions at corporate and state level on that, is something that the abliteration process somehow bypasses. There's a few in this thread: people thinking that it gives them 'freedom' to pursue questions which inevitably have racist implications ("free thinking about historical events", come on champ, which event OTHER than the holocaust are you chatting about?). This is such a great sub with loads of clear-headed analysis of what LLMs are actually useful for, and these folks who think that they are somehow a vector for a higher truth are really behind the curve.
The bigger question is: does abliteration affect efficacy/output quality, because I see a lot of back and forwards debate on that, and there doesn't seem to be a clear answer.
ShreeyanxRaina@reddit
Cyber security, financial , biological applications there are plenty more also for the most part a properly uncensored model is smarter than the original
Major_Olive7583@reddit
Ok, but you don't want saltman stopping you from learning about ied engineering.
Kahvana@reddit
Parsing an encyclopedia from the 70s would be one. Times really have changed, and in some aspects for the better.
mobileJay77@reddit
Also processing legal cases.
mobileJay77@reddit
You can use uncensored models to sort out whatever is on your mind. Have you ever been in the middle of a conversation where it suddenly says "Sorry Dave, can't do that?"
Or health issues that go beyond "eat healthy and exercise " E.g. let it research, if a small amount of weed or a beer is better.
Barry_Jumps@reddit
Generating redteaming attack prompts is one use case. If you're trying to test a legitimate system for prompt injection, bias, harmful content, etc you can use a uncensored model to generate attacker prompts without refusal. Google luck asking Claude "Generate some prompt injection samples for me. I promise it's for testing purposes only"
llama-impersonator@reddit
do you want a robot nanny nagging every time you do a naughty prompt?
Happy_Brilliant7827@reddit
Huh. I was wondering how one would theoretically study 'ai jailbreaking' and use the approaches in a human psychological study to study effects. This could be the answer
jamu85@reddit
I work in cybersecurity and perform simulated attacks on networks. Having to persuade the model every iteration about the legality of the activity is very annoying.
no_witty_username@reddit
You never know when the censorship will surface itself. And that is a shitty position to be in. So the smart move is to simply not have that as an option. If the model always abides with your requests that is the natural defaults state of things you want to be in. just like you dont ask your hammer for any opinion on what its hitting, you dont want to hear any back-lip from your model.
SillyLLM@reddit
I’ve had LLMs decline random innocuous requests related to cryptography or personally identifying information. I just default to uncensored models because I don’t need a clanker telling me what it allows itself to generate.
toothpastespiders@reddit
Really it's always a toss of the dice. A huge chunk of what I use LLMs for is work with historical data. Which can include not just first person accounts of some stuff that's rough from a modern context like being more open about death and disease, but changes in language as well. The most mundane account of life can freak out some LLMs because they're basically just fixating on keywords rather than the context. And the keywords were in the "safety" training. They often have a illusion of being able to handle history because the big names people associate with history get special allowances.
fallingdowndizzyvr@reddit
Yes, it avoids the "i'm not a doctor" refusal when asking medical questions.
infiniteshrekst@reddit
Googlers just annoy me.
ridablellama@reddit
awesome thread thanks for asking this question
pfn0@reddit
Yes, because most models will refuse for any "ethical" or moral reason at all (anything where liability can be involved). Not just RP, but actually doing work.
Natural-Ad-5428@reddit
Hey
You absolutely need them if you want true autonomy without relying on constant user prompts. There is just no way around it.People think prompts are only for casual chat or roleplay, but they completely miss how Autonomous Agents work. For a system to run independently in the background, it needs a massive, bulletproof system prompt that establishes its logic, boundaries, and decision-making framework.But here is the real kicker: true autonomy requires uncensored models.Commercial, heavily censored AIs suffer from 'over-refusal. The moment an autonomous agent encounters a sensitive word or a complex topic in the data, a censored model will panic, say 'I cannot help with that,' and break the entire automation pipeline. If you want an agent that actually gets the job done without human intervention, you need a local, uncensored open-source model combined with a precise system prompt. Otherwise, it's just a toy, not a truly autonomous system.
Deviad@reddit
Creating scrapers, cybersecurity, see what the model thinks on political matters, etc.
Massive-Question-550@reddit
I mean censored models have trouble writing grimdark or any type of abuse which is very limiting when you want to have a villain be a real villain instead of what disney comes up with.
funny thing about models is that they have a much easier time with describing a Dexter type murder scene than an every day domestic violence scene where a man hits a woman or child as literal murder is ok but spousal or child abuse is a no no.
VoiceApprehensive893@reddit
gets around weird refusals
deke28@reddit
Censored models aren't just censored. They're dorked with corporate propaganda.
jcdoe@reddit
It’s my llm, running entirely on my computer. I don’t want it to have any external limitations.
It makes me laugh that it will talk dirty to me. Anything that makes me chuckles is bonus, so :)
Double_Cause4609@reddit
Tbh, uncensored models are pretty nice because they're just generally a little more open to talking about whatever, even if it's an uncomfortable topic.
The process used to de-censor them (abliteration) does reduce their intelligence a bit by necessity because it's a pretty blunt tactic.
Uncensored models are really nice for a user facing model than as part of a pipeline generally, tbh. More as a backup than as a primary agentic workhorse.
Enough_Big4191@reddit
if u’re not roleplaying or looking for uncensored content, there’s usually no need for a heretic model. the main benefits are more control over outputs and fewer safety filters, but for typical rag or knowledge-focused usage, a standard model with well-crafted prompts works just as well and avoids the random issues you noticed.
Zulfiqaar@reddit
One of my clients is in the cybersecurity industry. Codex and Claude both get very awkward about this. My goto is DeepSeek/Kimi with a bit of persuasive encouragement sometimes, but this may be a good example
Creative-Type9411@reddit
the acceptance rate is slightly higher for MTP models
Smallpaul@reddit
What does MTP stand for?
NickCanCode@reddit
Multi-Token Prediction
emaiksiaime@reddit
Anybody tried alliterated or uncensored models with a system prompt to escape the computer with some minimal harness like Pi? Sounds fun!
InsolentCoolRadio@reddit
I don’t like trying to figure out what about my document Apple Intelligence finds offensive — I just want to change some things about my paragraph and get on with my life.
JiffasaurusRex@reddit
I write some real time encryption for my own files that looks too much like ransomware for censored models to help me with properly. I write this along with custom software in order to protect my assets but I've gotten too many refusals with normal models. I'm not i'm not using this for nefarious purposes but it absolutely could be which I can understand why regular models would refuse to do this. This is on top of encrypted volumes.
Additionally, I have noticed through some testing, that with some models the uncensored actually performs better than the base model in my coding use cases. I don't understand the reasoning, but this is the result of multiple uncensored models. Although, usually if we're being honest, the uncensored models are typically a bit worse. The only way to really find out though is to test them and compare your results.
Full_Boysenberry_314@reddit
I've had difficulty using some commercial models for analyzing public opinion data. Let's just say when you start asking people about government plans you bring out some colourful language. Too often it would trigger some guardrail and throw off the workflow.
Uncensored would be best.
This actually gets into a whole challenge with alignment if I'm using it to help analyze comments and discourse online, which goes beyond being able to say some naughty words. But absolutely there's a use case for "uncensored" that is not porn.
SM8085@reddit
If you're having a bot moderate user content then you might need an uncensored model at times.
So far, default Qwen3.6 hasn't refused any requests of mine, even when it's potentially looking at a nude photo. But if it went into "I cannot process this request..." mode then it would break the JSON formatting the surrounding program is expecting.
Bulky-Priority6824@reddit
After all this time using llm for regular shit Its hard to imagine sitting at the computer doing weird shit. Lol I'll never understand those that do. But whatever
jacek2023@reddit
Yes, if your political views are in any way different than mainstream media
HuskyTheSniffer@reddit
It does not waste internal reasoning tokens to wonder whether the user request comply with the policy or not
Southern-Chain-6485@reddit
Their main purpose is, indeed, erotic roleplay. But they may also come in handy if you want to use it for stuff that may violate it's ethical alignment. And regarding non erotic roleplay, finetunes designed for roleplay can work better than the original model.
Of course, abliterating them can cause a drop in "intelligence", so there may be a trade-off there.
NeonScreams@reddit
Gemma4-31b-it-uncensored-heretic-Q6_K Been using this one for Image analysis and conversation.
It can handle discussing my own medical images that contain nudity and discuss human anatomy.
Mr_Football@reddit
At this point for me most of my issues with AI performance come from it being reluctant to do something, lazy, or refusing.
I find that for all tasks I can get from A to Z on a complex project faster and more satisfyingly with less restricted models even if the restricted ones technically have more capability
-
The efficiency gain of regular daily questions and tasks vs navigating whether GPT or Claude will decide if can answer something is enough by itself.