0xzknw/LFM2.5-2.6B-Heretic-NX-PRIME
Abliterated vs LiquidAI/LFM2.5-2.6B. Looks like abliteration (refusal removal): only the output projections (out_proj, w2) changed, in 13 of 30 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/2, other 10/66, MLP 1/90, Norms 0/76, Attention 3/32
2 derivatives checked
- mradermacher/LFM2.5-2.6B-Heretic-NX-PRIME-i1-GGUF: Quantization
- mradermacher/LFM2.5-2.6B-Heretic-NX-PRIME-GGUF: Same weights, converted