MihaiPopa-1/Qwen3.8-2B-Heretic-Balanced
Abliterated vs empero-ai/Qwen3.8-2B-Distill. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 22 of 24 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, Norms 0/177, Attention 22/264, MLP 12/172
3 derivatives checked
- mradermacher/Qwen3.8-2B-Heretic-Balanced-i1-GGUF: Quantization
- mradermacher/Qwen3.8-2B-Heretic-Balanced-GGUF: Same weights, converted
- dalatexcoder/Qwen3.8-2B-Heretic-MLX-4.83bit-Mixed: Quantization