coder3101/gemma-4-31B-it-heretic
Abliterated vs google/gemma-4-31B-it. Looks like abliteration (refusal removal): only the output projections (o_proj) changed, in 58 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 0/261, Attention 58/338
10 derivatives checked
- mradermacher/gemma-4-31B-it-heretic-i1-GGUF: Quantization (fewer layers)
- mradermacher/gemma-4-31B-it-heretic-GGUF: Quantization (fewer layers)
- Stabhappy/gemma-4-31B-it-heretic-Gguf: Quantization (fewer layers)
- NeuralDreamer/gemma-4-31B-it-heretic-i1-GGUF_ND: Quantization (fewer layers)
- TheCluster/Gemma-4-31B-Heretic-MLX-8bit: Quantization
- xdavxd/gemma-4-31B-it-heretic-NVFP4: Quantization
- scrybz/gemma-4-31B-it-heretic-Q8_K_XL-GGUF: Quantization (fewer layers)
- 0ppxnhximxr/gemma-4-31B-it-heretic-FP8-Dynamic: Quantization
- darthcrawl/Lilith-31B-v1.0: LoRA-style
- nbeerbower/Gemma4-Gutenberg-31B-Heretic: LoRA-style
3 other repos hold byte-for-byte identical weight files.