grimjim/gemma-3-12b-it-biprojected-abliterated
Abliterated vs google/gemma-3-12b-it. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 31 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 31/252, Attention 31/408, other 0/1
2 derivatives checked
- mradermacher/gemma-3-12b-it-biprojected-abliterated-i1-GGUF: Quantization (fewer layers)
- mradermacher/gemma-3-12b-it-biprojected-abliterated-GGUF: Quantization (fewer layers)