megabytes/gemma-3-12b-qat-int4-heretic
Abliterated vs google/gemma-3-12b-it-qat-int4-unquantized. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 27 of 48 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, Norms 0/400, MLP 19/252, Attention 27/408, other 0/1
2 derivatives checked
- mradermacher/gemma-3-12b-it-qat-int4-unquantized-heretic-i1-GGUF: Quantization (fewer layers)
- mradermacher/gemma-3-12b-it-qat-int4-unquantized-heretic-GGUF: Quantization (fewer layers)