SergiusFlavius/Qwen3-VL-4B-Instruct-heretic
Abliterated vs Qwen/Qwen3-VL-4B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 28 of 36 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, Norms 0/249, MLP 25/220, Attention 28/240
2 derivatives checked
- mradermacher/Qwen3-VL-4B-Instruct-heretic-GGUF: Same weights, converted (fewer layers)
- mradermacher/Qwen3-VL-4B-Instruct-heretic-i1-GGUF: Quantization (fewer layers)
1 other repos hold byte-for-byte identical weight files.