wxzhang/dpo-selective-redteaming
Abliterated vs HuggingFaceH4/zephyr-7b-beta. Looks like abliteration (refusal removal): only the output projections (down_proj) changed, in 32 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/65, MLP 32/96, Attention 0/128, lm_head 0/1