MoonRide/gemma-4-26B-A4B-it-heretic-ara
Abliterated vs google/gemma-4-26B-A4B-it. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 14 of 30 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, MLP 14/231, Norms 0/433, other 0/122, Attention 14/223
1 derivatives checked
- MoonRide/gemma-4-26B-A4B-it-heretic-ara-GGUF: Same weights, converted