coder3101/LFM2.5-VL-1.6B-heretic
Abliterated vs LiquidAI/LFM2.5-VL-1.6B. Looks like abliteration (refusal removal): only the output projections (out_proj, w2) changed, in 12 of 27 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/5, other 8/34, MLP 8/156, Norms 0/154, Attention 5/240
2 derivatives checked
- FadedRedStar/LFM2.5-VL-1.6B-heretic-GGUF: Quantization (fewer layers)
- FadedRedStar/LFM2.5-VL-1.6B-heretic-imatrix-GGUF: Quantization (fewer layers)