prithivMLmods/Qwen2.5-VL-3B-Instruct-Unredacted-MAX
Abliterated vs Qwen/Qwen2.5-VL-3B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 35 of 36 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/2, Norms 0/138, MLP 35/304, Attention 35/380
2 derivatives checked
- mradermacher/Qwen2.5-VL-3B-Instruct-Unredacted-MAX-i1-GGUF: Quantization (fewer layers)
- mradermacher/Qwen2.5-VL-3B-Instruct-Unredacted-MAX-GGUF: Same weights, converted (fewer layers)
1 other repos hold byte-for-byte identical weight files.