ElementMerc/gemma-2-2b-it-abliterated

Abliterated vs google/gemma-2-2b-it. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 25 of 26 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.

Tensors changed: Embeddings 0/1, Norms 0/105, MLP 14/78, Attention 17/104

1 derivatives checked

On the Hub · Downloads on Model Pulse