richardyoung/Deepseek-R1-Distill-Qwen-32b-uncensored
Abliterated vs deepseek-ai/DeepSeek-R1-Distill-Qwen-32B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 46 of 64 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Attention 40/448, MLP 42/192, Norms 0/129, lm_head 0/1
1 derivatives checked
1 other repos hold byte-for-byte identical weight files.