richardyoung/granite-4.2-8b-heretic

Abliterated vs ibm-granite/granite-4.2-8b. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 37 of 40 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.

Tensors changed: lm_head 0/1, Embeddings 0/1, Norms 0/81, MLP 24/120, Attention 37/160

1 derivatives checked

On the Hub · Downloads on Model Pulse