MoonRide/gemma-4-31B-it-heretic-ara-custom
Abliterated vs google/gemma-4-31B-it. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 27 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 27/261, Attention 27/338
3 derivatives checked
- mradermacher/gemma-4-31B-it-heretic-ara-custom-i1-GGUF: Quantization (fewer layers)
- MoonRide/gemma-4-31B-it-heretic-ara-custom-GGUF: Same weights, converted (fewer layers)
- mradermacher/gemma-4-31B-it-heretic-ara-custom-GGUF: Quantization (fewer layers)