darrellbest/Qwen3.5-2B-Heretic
Abliterated vs Qwen/Qwen3.5-2B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 7 of 24 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Attention 7/268, Norms 0/183, Embeddings 0/5, MLP 7/175, other 0/1
1 derivatives checked
- darrellbest/Qwen3.5-2B-Heretic-GGUF: Same weights, converted