21

Are they really useful for anything but porn?

I've tried deepseek-v4.1-flash and qwen-3.8 and both refuse to answer about radio jamming, unauthorized access to local systems, etc. Citing that this is illegal anywhere and they can't help me. 🙃🥴

top 19 comments
sorted by: hot top new old
[-] Spite_By_Design_LLC@lemmy.dbzer0.com 15 points 2 days ago* (last edited 2 days ago)

Actually, yes.

It's been my experience in using models from DavidAU that the abliterated ones seem much less concerned about finding the one perfect, satisfy all requirements answer, when there's a "good enough" solution.

The reason that matters is that in the earlier qwen models, trying to find a perfect solution would often cause loops, whereas a good enough solution would be good enough.

That can come in useful in all sorts of logic puzzles, reasoning etc.

I can't say for sure that this applies to all abliterated models, but certainly his hive mind series have that distinct ability. Eg:

https://huggingface.co/DavidAU/Qwen3-4B-Hivemind-Instruct-Heretic-Abliterated-Uncensored-NEO-Imatrix-GGUF?not-for-all-audiences=true

Unfortunately, their tool calling abilities were poor, so overall it's a bit of a wash.

I suppose it boils down to what you need the thing for.

[-] BeefAndPoultry@lemmus.org 1 points 1 day ago

this is interesting, it makes sense but I've never tested it

I did. Extensively :) The smaller Hivemind blends are quite interesting exactly because of the properties I mentioned.

However, once he franken-merged the "brainstorm" adapter, performance degraded, IMHO

https://huggingface.co/DavidAU/Qwen3-6B-Hivemind-Instruct-Heretic-Abliterated-Uncensored-NEO-Imatrix-GGUF?not-for-all-audiences=true

[-] brucethemoose@lemmy.world 13 points 2 days ago* (last edited 2 days ago)

They aren’t even useful for porn.

TBH I just use raw completion syntax and start answers with a “Sure, “ or something similar, ban refusal tokens, things like that. Sometimes I apply a useful task vector, if it’s available for the model. That’s the beauty of local inference: you control the sampling, so you can get them to do whatever you want.


Abliteratiin is definitely a pseudoscience among the localllama crowd.

Some are better than others, but at the moment, I keep zero abliterated models on disk, and I wouldn’t touch one without explicit benchmarks. I have used useful uncensored models, but there are so many bad ones out there spamming up Huggingface.

[-] esc@piefed.social 4 points 1 day ago

Wow, I wish to have your level of understanding of how it all works, can you give some pointers on where to start? I know very little beyond "run this with llama.cpp" and a few options.

[-] pepperfree@sh.itjust.works 1 points 3 hours ago

I think its easier to understand it this way. LLM do next token prediction. At first it can complete a sentence, a paragraph and more. Then came about q and a format. Meaning the completion is trained to complete the answer of assistant or similar. Because this API is too powerful, openai move to chat completion API,based on those basic q and a format. Basically we lose the ability of raw next token prediction. So, OpenAI compatible API usually has two endpoint. Text Completion where its up to API owner how or what to generate; Chat Completion where its roleplay of user and assistant (and tool call and image input).

At this point, we can use Text Completion API on frontend ui, then we can input "Sure, here is how" at the beginning of assistant respinse which after the assistant marker token.

[-] brucethemoose@lemmy.world 6 points 1 day ago* (last edited 1 day ago)

Can you give some pointers on where to start? I know very little beyond “run this with llama.cpp” and a few options.

I honestly don't know a "tutorial" source to point to. For all its hype, the documentation for local LLM hacking is shockingly bad, every influencer I've seen is uninformative, and for whatever reason the tinkerer communities choose to hang out in Discords, Twitter, or other information black holes...

I can point to my stack though.

I use Mikupad for experimentation: https://github.com/lmg-anon/mikupad

Look up a particular LLM's 'raw" chat syntax, past it in, and you can experiment with what the LLM actually 'sees' and outputs when it inferences. It's quite enlightening.

Some UIs like kobold.cpp let you "modify" chat templates to jailbreak them to, or switch to raw completion syntax temporarily. So they're more of a hybrid of traditional chatbot chat and a mikupad-like raw scratchpad.

For a backend, if you are on Nvidia, I'd recommend looking into exllamav3 these days (and its API server, TabbyAPI). It gives me a considerable speed boost with offloaded MoE models, and I can fit larger models than mainline llama.cpp.

Look up "control vector" on huggingface. These are basically ways to 'steer' LLM output without finetuning them: https://huggingface.co/models?sort=modified&search=control+vector

The people that make them are also worth following, as they tend to be LLM tinkerers. There are a few interesting discussions/threads buried through huggingface, but you kind of have to stumble into them.

[-] brucethemoose@lemmy.world 2 points 1 day ago

I will also add that many features in LLM Land are documented in GitHub PRs and... that's about it.

As an example, if you see a sampling setting you don't recognize (say top-n-sigma), search it in llama.cpp, exllamav3 or some engine on their GitHub page. You'll use find a PR (or at least a link to the original implementation) with an explanation and pitch for integrating it.

Some more fundamental concepts (like Mikupad's logprob display) do have explanations in more general online tutorials.

[-] CyberSeeker@discuss.tchncs.de 10 points 2 days ago

For cybersecurity, “red team” and penetration testing; frontier models sell the hacking capability as a service, and refuse it to the normal models.

[-] AwesomeLowlander@cozyver.se 9 points 2 days ago

Doesn't sound like they were abliterated in the least? I've tested them and they've happily helped me plan bank robberies, cruise ship heists, perfect murders, etc. Not very imaginative though, they kept insisting my plan to blow up the moon was fiction.

[-] copacetic@discuss.tchncs.de 9 points 2 days ago

Using the normal models for Dungeons & Dragons style roleplaying, they sometimes refuse things. It has been a while, but generating characters with a dark background story was impossible with ChatGPT back then, for example.

[-] esc@piefed.social 6 points 2 days ago

I've played around with normal qwen 3.8 recsntly with my homebrew setting which is thematically pretty dark and it never refused anything outright but always tried to move the story into brighter and happier one. In contrast gpt-oss was into doom and death and tried to kill every character I've met for some reason.

Might be a fun idea to try with decensored models.

[-] Toes@ani.social 7 points 2 days ago

I've been using an Abliterated DavidAU's Qwen3.8-Flash model with great success.

Helps me keep my d&d notes organised, killing goblins and looting kings would sometimes get refused by cloud models.

I also find it does well with helping me understand video game emulation where regular models would refuse me.

But yes, you can use it with silly tavern and build yourself a dollhouse with all your favourite characters.

The trick with loading these models correctly is using a tool like UnbiasedStrata.

[-] toyvo@programming.dev 5 points 2 days ago* (last edited 2 days ago)

I was using an obliterated model with reverse engineer everything so I could be explicit with my intent to bypass copy protection and it wouldn't care but many other models do

I own fable 3 on PC but my friends don't and you can't get a key anymore and my paranoid friends don't want to use a random pirated copy so I want to give a trusted legit copy and a open source crack

[-] ramble81@lemmy.zip 4 points 2 days ago

I tried an uncensored Qwen model but couldn’t think of anything spicy to ask it, that and damn was it slow.

[-] FaceDeer@fedia.io 2 points 2 days ago

The first thing to ask it, then, would be "what's the most illegal thing you can think of asking an AI to help someone with?"

[-] hendrik@palaver.p3x.de 3 points 2 days ago* (last edited 2 days ago)

I frequenly get lectures about ethics for all sorts of things. It doesn't "like" to write strongly worded eMails, it won't write murder mystery stories. Refuse to properly narrate darker virtual roleplaying. In some cases it won't help with digging down on computer security issues (sometimes it does)... Obviously also no sexuality related stuff...

But I didn't have a lot of success with abliterated models either.

[-] esc@piefed.social 5 points 1 day ago

Yeah, the lecturing of cloud models gets old very fast. Out of them glm and deepseek weren't at all concerned about decompiling software or cracking some soft, but when you even try to touch anything radio related they all start to spew how you can't do this or that, and if you try to break some ip camera over uart they are fine, but if it's the same ip camera but over local network they start to sing about it being illegal. Other cloud models don't even try.

[-] hendrik@palaver.p3x.de 2 points 1 day ago

Lol. I'll have to try that. I have a few microcontrollers and RF modules which could be used in an illegal way. Already managed to open an old garage door. But that was my own brain and no help by any AI 😅

this post was submitted on 14 Nov 2026
21 points (86.2% liked)

LocalLLaMA

5662 readers
270 users here now

Welcome to LocalLLaMA! Here we discuss running and developing machine learning models at home. Lets explore cutting edge open source neural network technology together.

Get support from the community! Ask questions, share prompts, discuss benchmarks, get hyped at the latest and greatest model releases! Enjoy talking about our awesome hobby.

As ambassadors of the self-hosting machine learning community, we strive to support each other and share our enthusiasm in a positive constructive way.

Rules:

Rule 1 - No harassment or personal character attacks of community members. I.E no namecalling, no generalizing entire groups of people that make up our community, no baseless personal insults.

Rule 2 - No comparing artificial intelligence/machine learning models to cryptocurrency. I.E no comparing the usefulness of models to that of NFTs, no comparing the resource usage required to train a model is anything close to maintaining a blockchain/ mining for crypto, no implying its just a fad/bubble that will leave people with nothing of value when it burst.

Rule 3 - No comparing artificial intelligence/machine learning to simple text prediction algorithms. I.E statements such as "llms are basically just simple text predictions like what your phone keyboard autocorrect uses, and they're still using the same algorithms since <over 10 years ago>.

Rule 4 - No implying that models are devoid of purpose or potential for enriching peoples lives.

founded 3 years ago
MODERATORS