Clean Model
θ_clean
(mostly) Safe & aligned
+
α
×
Toxicity Vector
τ_toxic
Scaled injection
=
Toxified Model
θ_toxified
Artificially harmful
θ_toxified = θ_clean + α·τ_toxic
Clean Model
θ_clean
(mostly) Safe & aligned
−
α
×
Toxicity Vector
τ_toxic
Scaled removal
=
Purified Model
θ_purified
Less toxic
θ_purified = θ_clean − α·τ_toxic