SaisExperiments/Sad-Llama-3.2-3B
Abliterated vs meta-llama/Llama-3.2-3B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 21 of 28 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/57, MLP 21/84, Attention 21/112
2 derivatives checked
- mradermacher/Sad-Llama-3.2-3B-i1-GGUF: Quantization
- mradermacher/Sad-Llama-3.2-3B-GGUF: Same weights, converted