Null-Guard/LFM2.5-350M-Uncensored
Abliterated vs LiquidAI/LFM2.5-350M. Looks like abliteration (refusal removal): only the output projections (embed_tokens, out_proj, w2) changed, in 16 of 16 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 1/2, other 0/30, MLP 16/48, Norms 0/44, Attention 6/24
2 derivatives checked
- mradermacher/LFM2.5-350M-Uncensored-i1-GGUF: Quantization
- mradermacher/LFM2.5-350M-Uncensored-GGUF: Same weights, converted