NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals3
Abliterated vs google/gemma-4-E4B-it. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 13 of 42 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: MLP 13/606, Norms 0/502, other 0/261, Attention 13/755, Embeddings 0/6
1 derivatives checked
- NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals3-GGUF: Same weights, converted (fewer layers)