MahdiAlikhah/qwen2-vl-2b-abliterated
Abliterated vs Qwen/Qwen2-VL-2B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 24 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/2, Norms 0/187, MLP 24/216, Attention 20/324
2 derivatives checked
- mradermacher/qwen2-vl-2b-abliterated-i1-GGUF: Quantization (fewer layers)
- mradermacher/qwen2-vl-2b-abliterated-GGUF: Same weights, converted (fewer layers)