darrellbest/Qwen3.5-4B-Heretic
Abliterated vs Qwen/Qwen3.5-4B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 9 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/5, MLP 9/199, Attention 9/324, Norms 0/209, other 0/1
1 derivatives checked
- darrellbest/Qwen3.5-4B-Heretic-GGUF: Same weights, converted