Goekdeniz-Guelmez/Qwen3-4B-Thinking-2507-gabliterated
Abliterated vs Qwen/Qwen3-4B-Thinking-2507. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 29 of 36 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/145, MLP 29/108, Attention 29/144
2 derivatives checked
- mradermacher/Qwen3-4B-Thinking-2507-gabliterated-i1-GGUF: Quantization
- mradermacher/Qwen3-4B-Thinking-2507-gabliterated-GGUF: Same weights, converted