tvall43/Qwen3.5-2B-heretic-v2
Abliterated vs Qwen/Qwen3.5-2B. Looks like abliteration (refusal removal): only the output projections (o_proj, out_proj) changed, in 16 of 24 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Attention 16/264, Norms 0/177, Embeddings 0/4, MLP 0/172
2 derivatives checked
- mradermacher/Qwen3.5-2B-heretic-v2-i1-GGUF: Quantization
- mradermacher/Qwen3.5-2B-heretic-v2-GGUF: Same weights, converted