mlabonne/gemma-3-12b-it-abliterated
Abliterated vs google/gemma-3-12b-it. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 43 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 43/252, Attention 43/408, other 0/1
4 derivatives checked
- jeffcookio/gemma-3-12b-it-abliterated-gptqmodel-4b-128g: Quantization
- bartowski/mlabonne_gemma-3-12b-it-abliterated-GGUF: Same weights, converted (fewer layers)
- matrixportalx/gemma-3-12b-it-abliterated-GGUF: Same weights, converted (fewer layers)
- IlyaGusev/saiga_gemma3_12b: LoRA-style
2 other repos hold byte-for-byte identical weight files.