rawcell/Moonlight-16B-A3B-Instruct-bruno
Abliterated vs moonshotai/Moonlight-16B-A3B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 17 of 27 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/82, MLP 429/2781, Attention 17/108, lm_head 0/1
2 derivatives checked
- mradermacher/Moonlight-16B-A3B-Instruct-bruno-i1-GGUF: Quantization
- mradermacher/Moonlight-16B-A3B-Instruct-bruno-GGUF: Quantization