Repository navigation
Is the constrained-vs-unconstrained gap much wider on abliterated checkpoints? #2052
behrnt-slatgng
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Disclosure first: I build Grunz, a hosted chat + coding agent running abliterated open-weight models. Nothing to sell — I think outlines is quietly load-bearing for a whole class of model, and I would like to know whether anyone has measured it.
The claim
For abliterated / decensored checkpoints, constrained generation is not an optimisation. It is close to a requirement.
Here is why. When a model is abliterated — a refusal direction identified in activation space and projected out of the weights — the first capability to degrade is not knowledge. It is instruction-following and output-format adherence.
The model still knows the material. What gets measurably worse:
So the prompt-and-hope path to JSON, which is merely unreliable on a clean instruct model, degrades much further on an ablated one. Whereas outlines does not care whether the model wants to follow the schema — it forecloses the invalid tokens. The model's damaged willingness to comply stops being load-bearing.
Why I think the gap is wider than people assume
Unconstrained structured output depends on a behaviour that abliteration specifically damages. Constrained generation depends on the token distribution, which it damages much less.
If that is right, then the delta between constrained and unconstrained should be substantially larger on an ablated checkpoint than on its base — same weights otherwise, same quant, same schema. And given how much open-weight traffic is now modified checkpoints, that would make outlines a far more important dependency for that population than the benchmarks on clean models suggest.
What I cannot answer from where I sit
I see the symptom at volume but I do not have a controlled comparison. The experiment I want:
My prediction is that the unconstrained ablated cell falls off badly while the constrained ablated cell stays close to the constrained base cell. If that holds it is a genuinely useful result for anyone deploying these.
Questions
Has anyone here run anything like that, or seen it published? I would much rather cite a result than keep repeating a hunch.
Is there a known failure mode where a damaged model fights the grammar — producing valid-but-degenerate output, looping, or filling the schema with junk because the token it wanted was masked? That is the thing I would expect to go wrong, and I do not know if it shows up in practice.
Does anything in outlines surface how hard the constraint had to work — how often a high-probability token was masked? That would be a neat proxy for how much a given checkpoint is fighting its own output contract.
Happy to run the experiment and report back if nobody has. I would just rather not duplicate existing work or design it badly.
All reactions