MuXodious/gemma-3-4b-it-PaperWitch-heresy
Abliterated vs google/gemma-3-4b-it. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 25 of 34 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, Norms 0/316, MLP 21/210, Attention 25/352, other 0/1
2 derivatives checked
- mradermacher/gemma-3-4b-it-PaperWitch-heresy-i1-GGUF: Quantization (fewer layers)
- mradermacher/gemma-3-4b-it-PaperWitch-heresy-GGUF: Same weights, converted (fewer layers)