richardyoung/Agents-A1-4B-heretic

Abliterated vs InternScience/Agents-A1-4B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 20 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.

Tensors changed: Attention 20/320, MLP 19/196, Embeddings 0/4, Norms 0/203

1 derivatives checked

On the Hub · Downloads on Model Pulse