wangzhang/Qwen3.5-4B-abliterated
Abliterated (fewer layers) vs Qwen/Qwen3.5-4B. wangzhang/Qwen3.5-4B-abliterated matches part of Qwen/Qwen3.5-4B (426 of its 738 tensors; the rest of Qwen/Qwen3.5-4B has no counterpart). On the shared part: Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 29 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, MLP 25/96, Attention 29/224, Norms 0/105
3 derivatives checked
- mradermacher/Qwen3.5-4B-abliterated-i1-GGUF: Quantization
- mradermacher/Qwen3.5-4B-abliterated-GGUF: Same weights, converted
- adsyamsafa/Nixia-1.1-4B-GGUF: Quantization