DreamFast/qwen3-4b-heretic
Abliterated vs Qwen/Qwen3-4B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 32 of 36 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/145, MLP 27/108, Attention 32/144
4 derivatives checked
- liodon-ai/qwen3-4b-heretic-imatrix-GGUF: Not comparable yet
- N-Bot-Int/ElaNore3-4B_ADJUSTED_DPO-merged: LoRA-style
- N-Bot-Int/ElaNore3-4B-merged: LoRA-style
- N-Bot-Int/ElaNore3-4B_ADJUSTED_merged: LoRA-style
2 other repos hold byte-for-byte identical weight files.