A/b Test Base Vs Ablated Behavior On Identical Weights Via Alpha

Runtime refusal-direction ablation for Ternary Bonsai 2 27B: applies y -= alpha*(y·r)r at 129 residual writer sites in MLX, leaving the ~1.72-bit quantized weights bit-identical and alpha/layer tunable at inference.

AI Tools | Python · abliteration · uncensored-llm