DuoNeural/TurboGemma4E2B
Abliterated vs google/gemma-4-E2B-it. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 30 of 35 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: MLP 25/505, Norms 0/453, other 0/240, Attention 30/647, Embeddings 0/6
1 derivatives checked
- mradermacher/TurboGemma4E2B-GGUF: Same weights, converted (fewer layers)