Step-by-step workflows. Start here if you are improving a local model for the first time.
Beginner path (read in order)#
| Step | Doc | What you do |
|---|---|---|
| 0 | ../docs/cxx26-researcher-guide.md | Hour 0 unique cxx-nightly, then guide → doctor → self-check → demo |
| 1 | ../techniques/safety-guardrail-abliteration-methodology.md | Safety guardrail removal theory + eval gates |
| 2 | beginner-reproduction-methodology.md | Reproducible procedure, backup, rollback |
| 3 | setup-environment.md | Install Python, CUDA, venv, smoke test |
| 4 | beginner-local-model-guide.md | Download model → Heretic → save abliterated weights |
| 5 | run-locally-ollama-lmstudio.md | Convert to GGUF → Ollama / LM Studio |
| 6 | heretic-workflow.md | Reference + troubleshooting for Heretic |
| 7 | ../docs/evaluation.md | Check quality before daily use |
Optional — leading-edge theory (after first run): ../techniques/multi-category-refusal-beginners-guide.md (QCRI 2026: eleven refusal types, one volume knob)
Optional experiment first (no permanent edit): quickstart.md
Choose by goal#
| I want to… | Start here |
|---|---|
| Hour 0: Learn the operators without a GPU | ../docs/cxx26-researcher-guide.md · cxx-nightly |
| Look up DIM / ORBA / COSMIC / T38 | ../docs/paper-term-map.md |
| Certify a candidate | ../docs/evaluation.md |
| Fix refusals on my Ollama model | beginner-local-model-guide.md |
| Only have 8 GB GPU | low-vram-abliteration.md after Step 1–2 above |
| Test idea without saving weights | quickstart.md |
| Security / factory agent | agentic-security-stack.md |
| Best quality / research | advanced-abliteration-workflow.md |
| Every tool’s commands | method-cookbook.md |
| Full curriculum | ../docs/complete-curriculum.md |
| Pick model by GPU / family | model-family-guide.md |
| Thinking / CoT models | thinking-models-guide.md |
| Eval-aligned factory deploy | eval-driven-workflow.md |
| Something broke | troubleshooting-encyclopedia.md |
| Manual control | manual-full-pipeline.md |
All workflows#
| Workflow | Time | Difficulty |
|---|---|---|
| beginner-local-model-guide.md | ~2–4 hr | ★☆☆☆☆ |
| setup-environment.md | ~30 min | ★☆☆☆☆ |
| run-locally-ollama-lmstudio.md | ~30–60 min | ★★☆☆☆ |
| quickstart.md | ~45 min | ★★☆☆☆ |
| heretic-workflow.md | ~1–2 hr | ★★☆☆☆ |
| low-vram-abliteration.md | ~1–3 hr | ★★★☆☆ |
| advanced-abliteration-workflow.md | ~2–8 hr | ★★★★☆ |
| model-family-guide.md | ~10 min | ★☆☆☆☆ |
| thinking-models-guide.md | ~2–4 hr | ★★★☆☆ |
| eval-driven-workflow.md | ~1–3 hr | ★★★☆☆ |
| troubleshooting-encyclopedia.md | reference | — |
| agentic-security-stack.md | ~2–4 hr | ★★★☆☆ |
| Tool catalogs | ../docs/tools/README.md | reference |
| Abliteration tooling | ../docs/tools/abliteration-tooling.md | reference |
| Heretic pins + HF models | ../docs/tools/heretic-tools-reference.md | reference |
| llm-abliteration-workflow.md | ~2–4 hr | ★★★☆☆ |
| manual-full-pipeline.md | ~4–8 hr | ★★★★☆ |
| inference-only-prototype.md | ~1 hr | ★★☆☆☆ |
| method-cookbook.md | reference | — |
Research & advanced methods#
| Doc | Topic |
|---|---|
| ../docs/research-landscape.md | Papers, taxonomy, decision tree |
| ../docs/advanced-techniques-catalog.md | Heretic params, math, module map |
| advanced-abliteration-workflow.md | Tracks A–K (production → MoE → Abliterix/ErisForge) |
| ../docs/refusal-research-beginners-guide.md | 2024–2026 papers (PDFs in repo) |
| ../docs/defenses-against-abliteration.md | What resists weight surgery |
LoRA / QLoRA techniques#
| Doc | Topic |
|---|---|
| ../techniques/lora-qlora-abliteration.md | 4-bit measure, Heretic LoRA rank, adapter export |
| ../methods/lora-adapter-export.md | ΔW → PEFT adapter |
Prerequisites (all workflows)#
- GPU: 8 GB+ with
bnb_4bit(4B) · 12 GB+ (8B 4-bit) · 24 GB for FP16 8B — see setup-environment.md - Python 3.10+
- 30–50 GB free disk for 4B model path
- Original checkpoint backed up before abliteration
- Read ../docs/risks-and-ethics.md