darrellbest/Qwen3.5-9B-Heretic
Abliterated vs Qwen/Qwen3.5-9B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 23 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: lm_head 0/1, Embeddings 0/5, MLP 23/211, Attention 23/336, other 0/1, Norms 0/221
1 derivatives checked
- darrellbest/Qwen3.5-9B-Heretic-GGUF: Same weights, converted