Evaluation · Shipped · research use
Refusal ablation
Removing refusal behaviour from open models and measuring the damage, so red-teamers and guardrail developers have something to test against.
Why these exist
Studying how systems fail requires models that do not refuse. A red-teamer needs to generate the attack to test the defence; a guardrail developer needs unfiltered output to check the detector fires. A model that declines is useless as a test fixture.
This is the least comfortable work we do and the easiest to misread, so every release carries the numbers and a responsible-use notice rather than leaving the framing to the reader.
Measured, not asserted
Refusal behaviour is removed surgically. No capability is added — only refusal is reduced — and we report what that cost. On our multimodal build, refusals on a standard harmful-prompt set drop from 99 in 100 to 12 in 100, at a distributional distance of 0.053 from the original.
That second number is the one that matters. It sits well inside the range where general capability is understood to survive, and without it a refusal rate on its own says nothing about whether the model still works.
Models from this thread