MuXodious/Qwen3.5-4B-PaperWitch-heresy-v2
Abliterated vs Qwen/Qwen3.5-4B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 32 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, MLP 32/196, Attention 30/320, Norms 0/203
2 derivatives checked
- mradermacher/Qwen3.5-4B-PaperWitch-heresy-v2-i1-GGUF: Quantization
- mradermacher/Qwen3.5-4B-PaperWitch-heresy-v2-GGUF: Same weights, converted