A targeted change that weakens a model’s specific skill or habit.
Model abliteration is not smashing the vending machine. It is rewiring one sketchy button, so it stops selling chaos soda.
People use it for safety work and model editing. It can also weaken useful skills by accident.
Alignment
Abliteration can help Alignment by pushing down risky behavior.
Fine-tuning
Abliteration is targeted editing, not normal extra training.
Open weights
Open weights make this kind of direct model edit easier.
LLM
Abliteration changes some skills or habits inside an LLM.