tbilisi-ai-lab/kona2-small-3.8B
Abliterated vs microsoft/Phi-3.5-mini-instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 32 of 32 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Norms 0/65, MLP 32/64, Attention 32/64
2 derivatives checked
- mradermacher/kona2-small-3.8B-i1-GGUF: Quantization
- mradermacher/kona2-small-3.8B-GGUF: Same weights, converted