MuXodious/Qwen3.5-4B-SOMPOA-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 10 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, MLP 6/196, Attention 10/320, Norms 0/203
6 derivatives checked
- mradermacher/Qwen3.5-4B-SOMPOA-heresy-v2-i1-GGUF: Quantization
- mradermacher/Qwen3.5-4B-SOMPOA-heresy-v2-GGUF: Same weights, converted
- LoliRimuru/SlopPrompt: LoRA-style
- nuofang/Qwen3.5-4B-SOMPOA-heresy-v2-GGUF: Quantization
- MuXodious/Qwen3.5-4B-SOMPOA-heresy-v2-MTP: Same weights, repacked
- Nubinu/Qwen3.5-4B-MiniFantasy: LoRA-style
1 other repos hold byte-for-byte identical weight files.