thirdeyeai/SmolLM2-1.7B-Instruct-Uncensored
Abliterated vs HuggingFaceTB/SmolLM2-1.7B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj) changed, in 24 of 24 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 1/1, Norms 0/49, MLP 24/72, Attention 24/96
1 derivatives checked
- mradermacher/SmolLM2-1.7B-Instruct-Uncensored-GGUF: Same weights, converted