ApolloRaines/Llama-3.1-8B-Instruct.immunized.v1
Abliterated vs meta-llama/Llama-3.1-8B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 6 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 6/96, Attention 6/128, lm_head 0/1
1 derivatives checked
- mradermacher/Llama-3.1-8B-Instruct.immunized.v1-GGUF: Same weights, converted