DexopT/Qwen3-4B-Cybersecurity-Heretic-16bit

Abliterated vs DexopT/Qwen3-4B-Cybersecurity. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 23 of 36 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.

Tensors changed: Embeddings 0/1, Norms 0/145, MLP 19/108, Attention 23/144

2 derivatives checked

On the Hub · Downloads on Model Pulse