atakankutalp/LFM2.5-2.6B-heretic
Abliterated vs LiquidAI/LFM2.5-2.6B. Looks like abliteration (refusal removal): only the output projections (out_proj, w2) changed, in 27 of 30 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/2, other 13/66, MLP 27/90, Norms 0/76, Attention 6/32
1 derivatives checked
- atakankutalp/LFM2.5-2.6B-heretic-GGUF: Same weights, converted