Extended abliteration toolkit

Four community tools that complement the handbook default (Heretic). Use this page to pick a backend by automation, VRAM, architecture, or research depth — then run the same eval gates as any Heretic run.

Benchmark context: arXiv:2512.13655 (Young, Jan 2026) — handbook matrix in ../docs/comparative-abliteration-benchmarks.md.

Research map: ../docs/refusal-research-beginners-guide.md · defenses that resist surgery: ../docs/defenses-against-abliteration.md

Eval gates (mandatory): ../instructions/eval-driven-workflow.md · ../docs/risks-and-ethics.md

Quick comparison#

Abliterix — when Heretic is not enough#

Use when: MoE/hybrid/VL/SSM families need expert-granular or architecture-specific configs; you want multi-objective Optuna (refusal rate + KL) with 150+ upstream presets; or you need LoRA abliteration / steering / ORBA / SAE paths in one CLI.

pip install -U abliterix
abliterix --model Qwen/Qwen3-4B-Instruct-2507
# 4-bit: --model.quant-method bnb_4bit
# Windows: python scripts/run_abliterix.py --model …

HonestAbliterationBench: upstream benchmark spec at benchmarks/SPEC.md in the Abliterix repo. Pair with handbook corpora (data/eval/*.jsonl) — upstream numbers are not a substitute for factory/CyberGym gates.

Handbook placement:

• MoE / VL / Gemma 4 families → model-family-playbook.md
• Production-hardening track → advanced-abliteration-workflow.md Track J
• Agent stack comparison → ../docs/tools/abliteration-tooling.md

Caveat: Community fork of the Heretic lineage — verify license (AGPL-3.0), config model ids, and run full eval before any weights leave the lab.

ErisForge — fast layer-band prototyping#

Use when: You know the layer band (from Heretic --print-residual-geometry or manual silhouette) and want a single-pass ablation without Optuna overhead.

git clone https://github.com/Tsadoq/ErisForge.git
cd ErisForge && pip install -e .
# Minimal API: swap decoder layers + ExpressionRefusalScorer — see upstream README

• Layer band workflow → layer-selective-abliteration.md
• Mean-difference direction theory → mean-difference-direction.md
• Capability-first fallback when Heretic KL is poor → ../docs/comparative-abliteration-benchmarks.md

llm-abliteration + --deccp — low VRAM and multilingual measure#

Fork choice:

Sharded ablate keeps peak memory at one layer matrix — critical for 20B+ on consumer hardware.

python measure.py -m <model> -o directions.pt --quant 4bit --projected
python measure.py -m <model> -o directions.pt --quant 4bit --deccp   # multilingual topics
python sharded_ablate.py config.yaml --projected --normpreserve

Full path → ../instructions/low-vram-abliteration.md

DECCP (AUGMXNT/deccp): Qwen2 Instruct PoC + Chinese/multilingual topic sets. Not a peer CLI. Multilingual measure is python measure.py --deccp on jim-plus llm-abliteration.

FailSpy/abliterator — mechanistic notebooks#

Use when: You need to cache activations, compute refusal directions interactively, and test hook ablation (temporary vs permanent) before committing weight surgery.

# Pattern: TransformerLens HookedTransformer + hook_resid_post
# Layer whitelist/blacklist — see upstream notebooks

• Interpretability stack → beyond-single-direction.md
• Zero-weight-edit VRAM path → ../instructions/low-vram-abliteration.md
• Not for broad architecture deploy — TransformerLens coverage is narrower than Heretic/Abliterix.

Decision tree (summary)#

Need automatic deploy + eval gates?
  └─ Heretic (default) OR Abliterix if MoE/VL/SSM + preset exists

Heretic OOM or 20B+ on consumer GPU?
  └─ jim-plus measure 4-bit + sharded_ablate (`--deccp` if non-English topics)

GSM8K regression blocker on Heretic?
  └─ Try ErisForge single-pass, or jim-plus `--projected --normpreserve` — re-eval factory JSONL anyway

Research / direction prototyping only?
  └─ FailSpy/abliterator hooks → bake winner with Heretic or Abliterix

Agent + tool-calling specifics → ../docs/tools/abliteration-tooling.md

Every install one-liner: ../instructions/method-cookbook.md

Also in the 2026 toolbox (not in Young’s 16-model matrix):

Related#

• ../docs/refusal-research-beginners-guide.md — paper map + offline PDFs
• ../docs/defenses-against-abliteration.md — when tools report low refusal drop
• ../references.md — URLs and install snippets
• ../docs/toolchain-safetensors-gguf-lora.md — post-abliteration quant chain
• steering-and-alternatives.md — when not to edit weights