0xA50C1A1/Llama-3.3-70B-Instruct-SOM-MPOA
Abliterated vs unsloth/Llama-3.3-70B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 49 of 80 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/161, MLP 29/240, Attention 49/320, lm_head 0/1
2 derivatives checked
- mradermacher/Llama-3.3-70B-Instruct-SOM-MPOA-i1-GGUF: Quantization
- mradermacher/Llama-3.3-70B-Instruct-SOM-MPOA-GGUF: Quantization