saidutta69/DeepSeek-R1-Distill-Qwen-7B-heretic
Abliterated vs deepseek-ai/DeepSeek-R1-Distill-Qwen-7B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 22 of 28 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Attention 10/196, MLP 22/84, Norms 0/57, lm_head 0/1