krthkyadtya/Qwen2.5-72B-Instruct-heretic
Abliterated vs Qwen/Qwen2.5-72B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 58 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 14/240, Attention 58/560, lm_head 0/1
2 derivatives checked
- mradermacher/Qwen2.5-72B-Instruct-heretic-i1-GGUF: Quantization
- mradermacher/Qwen2.5-72B-Instruct-heretic-GGUF: Quantization