Abliteration

A post-training weight-editing technique that suppresses an aligned model's refusal behaviour while leaving its architecture, parameter count and training data unchanged. Builds produced this way are commonly distributed as 'uncensored' variants.