tvall43/Qwen3.5-0.8B-heretic-v3
Abliterated vs Qwen/Qwen3.5-0.8B. Looks like abliteration (refusal removal): only the output projections (o_proj, out_proj) changed, in 10 of 24 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Attention 10/216, Norms 0/129, Embeddings 0/4, MLP 0/124
2 derivatives checked
- mradermacher/Qwen3.5-0.8B-heretic-v3-GGUF: Same weights, converted
- mradermacher/Qwen3.5-0.8B-heretic-v3-i1-GGUF: Quantization
1 other repos hold byte-for-byte identical weight files.