MuXodious/Qwen3.5-4B-ARA-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 13 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, MLP 13/196, Attention 13/320, Norms 0/203
4 derivatives checked
- mradermacher/Qwen3.5-4B-ARA-heresy-v2-i1-GGUF: Quantization
- mradermacher/Qwen3.5-4B-ARA-heresy-v2-GGUF: Same weights, converted
- Sinbad-The-Sailor/Qwen3.5-4B-NSFW-ARA-Heretic-Literotica: LoRA-style (fewer layers)
- adamm-hf/Qwen3.5-4B-NSFW-Literotica: LoRA-style (fewer layers)