# Abliteration

**Abliteration** is the standard open-weights technique for "uncensoring" an instruction-tuned model — removing or attenuating its refusal behavior. It works by identifying the [[steering-vectors|steering vector]] that corresponds to refusal (built from matched prompt pairs that elicit a refusal versus comply) and either subtracting it during inference or, in the weight-edit variant, zeroing its projection out of the relevant weight matrices.

Common assumption — including Sean Goedecke's, before the HN comments on his steering post corrected him — is that "uncensored" community releases are LoRA fine-tunes. Most aren't. They are abliterations, and the technique sits closer to neurosurgery than to training: pick the direction, push the model along it (or away from it).

## Two flavors, different tradeoffs

- **Runtime steering.** Add the inverse refusal vector to the activations at inference time. Per-request, reversible, and can be applied only when needed. Capability damage is bounded because the weights are untouched.
- **Weight orthogonalization.** Bake the change into the weights by removing the refusal direction's contribution to attention/MLP matrices. Doesn't need a custom runtime; the model just behaves uncensored. Costs: capability degradation across unrelated tasks, because the same neurons participate in many directions.

antirez, in the Hacker News thread on [[steering-vectors-interesting-again|Goedecke's post]], made the case for the runtime form: weight edits damage capability globally, while runtime steering pays its cost only where applied. This is the same general argument that runs across [[steering-vectors|steering as a whole]] — a narrow, switchable, non-destructive intervention is the niche where vector arithmetic genuinely beats a fine-tune.

## Why it matters even if you don't ship uncensored models

The technique is the most-shipped real-world application of steering in the open-weights world. It demonstrates that the conditions Goedecke identifies as steering's sweet spot — behavior baked in by training, prompts can't override it, you have local weights — are not hypothetical. Refusal happens to be the case where the open-source community had the strongest motive to develop the tooling, but the same pattern generalizes to any trained-in behavior somebody wants to selectively turn off.

It also matters for [[anti-llm-discourse|safety arguments]]: closed-API safety arguments often assume that refusal training is robust. Abliteration is the standing counterexample for open-weights models, and the techniques transfer to anything else of similar shape (sycophancy, persona, formatting constraints).

## Related

- [[steering-vectors]] — the underlying technique
- [[claude-emotion-concepts]] — Anthropic's interpretability stack for finding equivalent vectors at scale
- [[deepseek]] — open-weights models where this has been demonstrated