coder3101/Qwen3.5-4B-heretic
Abliterated vs Qwen/Qwen3.5-4B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 18 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, MLP 18/196, Attention 11/320, Norms 0/203
6 derivatives checked
- FadedRedStar/Qwen3.5-4B-heretic-GGUF: Same weights, converted (fewer layers)
- FadedRedStar/Qwen3.5-4B-heretic-imatrix-GGUF: Same weights, converted (fewer layers)
- DavidAU/Qwen3.5-4B-Deckard-HERETIC-UNCENSORED-Thinking: LoRA-style
- DavidAU/Qwen3.5-4B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING: LoRA-style
- TheCluster/Qwen3.5-4B-Heretic-MLX-mxfp4: Quantization
- filvyb/Qwen3.5-4B-heretic: Same weights, converted
1 other repos hold byte-for-byte identical weight files.