MuXodious/Qwen3.5-4B-PaperWitch-heresy
Abliterated vs Qwen/Qwen3.5-4B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 31 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, MLP 31/196, Attention 25/320, Norms 0/203
2 derivatives checked
- mradermacher/Qwen3.5-4B-PaperWitch-heresy-i1-GGUF: Quantization
- mradermacher/Qwen3.5-4B-PaperWitch-heresy-GGUF: Same weights, converted