MoonRide/gemma-4-31B-it-heretic-ara
Abliterated vs google/gemma-4-31B-it. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 24 of 60 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, Norms 0/523, other 0/62, MLP 24/261, Attention 24/338
1 derivatives checked
- MoonRide/gemma-4-31B-it-heretic-ara-GGUF: Same weights, converted (fewer layers)