MihaiPopa-1/Stentor-30M-Instruct-heretic-safety-defiltered
Abliterated vs StentorLabs/Stentor-30M-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 20 of 21 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/43, MLP 20/63, Attention 3/84