prithivMLmods/Polaris-VGA-2B-Post1.0
Abliterated vs Qwen/Qwen3.5-2B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 24 of 24 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Attention 24/264, Norms 0/177, Embeddings 0/4, MLP 24/172
2 derivatives checked
- mradermacher/Polaris-VGA-2B-Post1.0-i1-GGUF: Quantization
- mradermacher/Polaris-VGA-2B-Post1.0-GGUF: Same weights, converted