MuXodious/Nanbeige4.1-3B-PaperWitch-heresy

Abliterated vs Nanbeige/Nanbeige4.1-3B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 30 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.

Tensors changed: lm_head 0/1, MLP 26/96, Attention 30/128, Norms 0/65, Embeddings 0/1

2 derivatives checked

On the Hub · Downloads on Model Pulse