alt Hacker NewsSpeaking to the "uncensored model" angle: there's little reason to distribute abliterated weights anyway. Instead of orthogonalising the weights that write back to the residual stream, you can just orthogonalise the activations themselves. It's equivalent.
Orthogonalising activations at runtime is computationally cheap. Just distribute the refusal vectors (few thousand floats per layer), then run against the stock weights. Antirez's DS4 already supports this: https://github.com/antirez/ds4/blob/8db1d1d155cb0400a86a86b9...
Abliterated weights are just a bad habit we've gotten into. It's also deeply suboptimal from a precision point of view to take a model that's already been QATed and distributed in pre-quantised form (DeepSeek V4, Kimi K2.5 or K3...), modify its weights, and re-quantise it. Similarly, abliterated models regain some of their refusal behaviour when they're re-quantised after abliteration -- avoidable by keeping the two separate.
Replies
Interesting. Is there a paper that explains this in a bit more detail, like [0] for abliteration (underlying the Heretic software, afaik)?
Yeah I use a custom fork of llama.cpp that has an abliteration feature that basically does this. It's sloppily vibe coded and I don't have time to coordinate on a way to do this cleanly upstream, but it's absolutely possible and saves a lot of time and bandwidth from being wasted
Can you explain this a bit to a non-expert?
I haven't wrapped my mind around this
So... distribute a LoRA (or equivalent) that modifies the base weights with the abliteration vectors. That makes sense as it would be possible to try different abliterations and keep the storage space down.
> Similarly, abliterated models regain some of their refusal behaviour when they're re-quantised after abliteration
Thanks for this information, Q4 seemed fine but they reappeared again in Q5 with an vengeance, I couldn't understand why. Very Strict and I've only found one jail break that barely works around 60% of the time.
Who is offering the refusal vectors or instructions on how to generate them?
I believe abliterated models are mostly still created at this point because they're "universal": they can not only be run locally, and on cloud GPUs, but also on "managed inference" providers (i.e. services where you hand them a model URI, and they blindly fetch it, load it, and give you inference access to it through standard text/chat-completion APIs. Think HuggingFace Spaces, or Google CoLab, or CloudFlare Workers AI.)
Such managed inference providers have (for now) plausible deniability of behaving ethically (at least enough that they don't get boycotted / scare away investors) due to them being "blind" to what gets run on their systems. They're acting as the inference equivalent of data transit carriers.
But I don't think it would be possible for managed inference providers to publicly expose "runtime activation steering" in the way antirez's DS4 does, without that reading much more explicitly as them inviting unethical workloads.
(Yes, there are other things you can do with runtime steering. But almost all of those things are workload-specific, relying on you privately tuning to the needs of your own dataset. And if you can do that, you can run inference without the help of a managed inference provider. The only time a customer will come along with a pre-made runtime-steering vector file in hand, is if that vector is an alignment-orthogonalization vector.)