Attention o_proj abliteration

Target#

block.self_attn.o_proj   # Linear: num_heads * head_dim → hidden_size

Refusal signal sometimes couples through attention mixing — ablating o_proj can complement MLP edits.

Caution#

Attention layers are more sensitive — start with lower α and fewer layers than MLP recipe.

Procedure#

Same projection as MLP method but on o_proj.weight with the same r_ℓ estimated at post-attention residual (not post-MLP) if you want consistency.

Combined strategy#

1. Abliterate MLP bands at α=0.75
2. If false refusals remain, add o_proj on top 3 layers at α=0.25
3. Re-evaluate capability suite