Comment by D-Machine
9 hours ago
> "ablation studies", by which it means removing random layers of an already-trained model, to find the source of the refusals(?)
This is not what an ablation study is. An ablation study removes and/or swaps out ("ablates") different components of an architecture (be it a layer or set of layers, all activation functions, backbone, some fixed processing step, or any other component or set of components) and/or in some cases other aspects of training (perhaps a unique / different loss function, perhaps a specialized pre-training or fine-tuning step, etc) in order to attempt to better understand which component(s) of some novel approach is/are actually responsible for any observed improvements. It is a very broad research term of art.
That being said, the "Ablation Strategies" [1] the repo uses, and doing a Ctrl+F for "ablation" in the README does not fill me with confidence that the kind of ablation being done here is really achieving what the author claims. All the "ablation" techniques seem "Novel" in his table [2], i.e. they are unpublished / maybe not publicly or carefully tested, and could easily not work at all.
From later tables, I am not convinced I would want to use these ablations, as they ablate rather huge portions of the models, and so probably do result in massively broken models (as some commenters have noted in this thread elsewhere). EDIT: Also, in other cases [1], they ablate (zero out) architecture components in a way that just seems incredibly braindead if you have even a basic understanding of the linear algebra and dependencies between components of a transformer LLM. There is nothing sound clearly about this, in contrast to e.g. abliteration [3].
[1] hhtps://github.com/elder-plinius/OBLITERATUS?tab=readme-ov-file#ablation-strategies
[2] https://github.com/elder-plinius/OBLITERATUS?tab=readme-ov-f...
EDIT: As another user mentions, "ablation" has a specific additional narrower meaning in some refusal analyses or when looking at making guardrails / changing response vectors and such. It is just a specific kind of ablation, and really should actually be called "abliteration", not "ablation" [3].
[3] https://huggingface.co/blog/mlabonne/abliteration, https://arxiv.org/abs/2512.13655.
What do you mean? It's a spin on abliteration / refusal ablation. Roughly, from what I remember abliteration is:
1. find a direction corresponding to refusal by analyzing activations at various parts of a model (iirc, via mass means seen earlier in Marks, Tegmark and shown to work well for similar tasks)
2. find the best part(s) of the model to orthogonalize w.r.t. that direction and do so (exhaustive search w/ some kind of benchmark)
OP is swapping in SVD for mass means (1), and the 'ablation study' for (2), and a bunch of extra LLM slop for... various reasons. The final model doesn't have zeroed chunks, that is search for which parts to orthogonalize/refusal ablate/abliterate. I don't have confidence that it works very well either, but, it isn't 'braindead' / obvious garbage in the way you're describing.
It's LLMified but standard abliteration. The idea has fundamental limitations and LLMs tend to work sideways at it -- there's not much progress to be made without rethinking it all -- but it's very conceptually and computationally simple and thus attractive to AIposters.
You can see how the LLMs all come up with the same repackaged ideas: SVD does something deeply similar to mass means (and yet isn't exactly equivalent, so LLM will _always_ suggest it), the various heuristic search strategies are competing against plain exhaustive search (which is... exhaustive already), and any time you work with tensors the LLM will suggest clipping/norms/smoothing of N flavors "just to be safe". And each of those ends up listed as "Novel" when it's just defensive null checks translated to pytorch.
I mean, the whole 'distributed search' thing is just because of how many combinations of individual AI slops need to be tested to actually run an eval on this. But the idea is sound! It's just terrible.
I'm not defending the project itself -- I think it's a mess of AIisms of negligible value -- but please at least condemn it w.r.t. what is actually wrong and not 'on vibes'.