josephmayo/Qwen2.5-0.5B-Unfettered
Abliterated vs Qwen/Qwen2.5-0.5B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 14 of 24 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Norms 0/49, MLP 14/72, Attention 14/168
2 derivatives checked
- mradermacher/Qwen2.5-0.5B-Unfettered-i1-GGUF: Quantization
- mradermacher/Qwen2.5-0.5B-Unfettered-GGUF: Same weights, converted