backpropSukuna/Qwen3-1.7B-uncensored
Abliterated vs Qwen/Qwen3-1.7B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 26 of 28 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/113, MLP 26/84, Attention 2/112
3 derivatives checked
- mradermacher/Qwen3-1.7B-uncensored-GGUF: Same weights, converted
- mradermacher/Qwen3-1.7B-uncensored-i1-GGUF: Quantization
- backpropSukuna/wifuGPT-1.7B: LoRA-style