jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.25
Abliterated vs Qwen/Qwen3.8-27B. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj, out_proj) changed, in 64 of 64 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Norms 0/319, Attention 64/556, MLP 64/304, Embeddings 1/4, lm_head 0/1
1 derivatives checked
- mradermacher/Qwen3.8-27B-refusal-ablation-L46-a0.25-GGUF: Quantization