How each abliteration family is implemented — formulas, target modules, pseudocode.
Runnable operators (DIM / projected / COSMIC score / SVD / Arditi / ORBA / Householder / subspace / hook / marker eval): C++26 abliterate-cxx — ../docs/cxx26-platform.md.
Method selection guide#
Start here
│
├─ Want reversible experiments? ──► residual-hook-ablation
│
├─ Want fastest path to checkpoint? ──► automated-heretic-search
│
├─ Low VRAM (8–12 GB)? ──► Heretic bnb_4bit OR llm-abliteration --quant 4bit
│ └─► [../instructions/low-vram-abliteration.md](../instructions/low-vram-abliteration.md)
│
├─ Want adapter-only deploy? ──► lora-adapter-export
│
├─ Want full control / research? ──► manual-transformers-pipeline
│ └─► mlp-down-proj-abliteration
│
├─ Need projected/normpreserve manual? ──► projected-llm-abliteration
│
├─ Have r.pt already (COSMIC/SOM/factory)? ──► orba-pipeline / apply-weight-abliteration.py
│
├─ Weakly aligned / no refusal template? ──► cosmic-direction-id
│
├─ Need k-D subspace (not one arrow)? ──► svd-refusal-subspace
│
├─ Multi-direction / RDO research? ──► multi-direction-ablation / gradient-rdo-pipeline
│ └─► protected-subspace-abliteration
│
├─ MoE model? ──► moe-expert-abliteration (or Heretic)
│
└─ Deploy in llama.cpp? ──► safetensor surgery → gguf-export-notes
└─ LoRA sidecar? ──► lora-adapter-export → convert_lora_to_ggufBefore every permanent path: prepare contrasts with contrast-set-design.md, require held-out causal evidence with direction-diagnostics-and-localization.md, and bind outputs with experiment provenance.
Selective tensor and layer routes (September 2026)#
Selective method guide covers exact tensor plans, verified exports, norm-preserving projection, and selected-tensor SFT. MiniCPM5-1B study records the tested support matrix and measured limitations.
Adaptive interventions (October 2026)#
Adaptive interventions and ARA compares fixed projections, kernelized steering, context-conditioned adapters, selective representation edits, and module-I/O optimization. Huihui evidence records per-checkpoint claims without assigning every publisher model one algorithm.