coder3101/gemma-4-12B-it-qat-q4_0-unquantized-heretic
Abliterated vs google/gemma-4-12B-it-qat-q4_0-unquantized. Looks like abliteration (refusal removal): only the output projections (o_proj) changed, in 24 of 48 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/12, Norms 0/289, other 0/48, MLP 0/144, Attention 24/184
4 derivatives checked
- SC117/gemma-4-12B-it-heretic-QAT-GGUF: Quantization
- FadedRedStar/gemma-4-12B-it-qat-q4_0-heretic-GGUF: Not comparable yet
- pegasus912/gemma-4-12b-it-qat-heretic-ud-q4-k-xl: Quantization
- Disya/gemma-4-12B-it-qat-q4_0-unquantized-heretic-gguf: Same weights, converted