bunnycore/Llama-3.2-3B-Mix
Abliterated vs chuanli11/Llama-3.2-3B-Instruct-uncensored. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj) changed, in 28 of 28 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 1/1, Norms 0/57, MLP 28/84, Attention 28/112, lm_head 0/1
2 derivatives checked
- mradermacher/Llama-3.2-3B-Mix-GGUF: Same weights, converted
- powermove72/Llama-NextGenmix-4b: Partial fine-tune