heretic-org/LFM2.5-VL-3B-heretic
Abliterated vs LiquidAI/LFM2.5-VL-3B. Looks like abliteration (refusal removal): only the output projections (out_proj, w2) changed, in 19 of 30 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/5, other 13/70, MLP 19/198, Norms 0/186, Attention 5/248
1 derivatives checked
- AX1Y2JP/LFM2.5-VL-3B-heretic-GGUF: Same weights, converted (fewer layers)
1 other repos hold byte-for-byte identical weight files.