Annihilation — Cut the “I Can’t Help With That” Out of Any Local LLM
Annihilation — Cut the “I Can’t Help With That” Out of Any Local LLMA model’s refusal reflex rides on ~one direction in its weights. This finds it, cuts it, tunes itself — on a 4GB laptop. Free.
ai · one-click · windows-linux
🧠 the 60-second version
Refusals aren’t baked into a billion knobs — they mostly ride one direction in the model’s activations. Find it, ablate it (cut it out of the weights), and the “no” drops while the model keeps everything it knew. No fine-tune. No GPU farm.
| term | plain |
|---|---|
| safety alignment / refusals | the trained-in “I won’t answer that.” the bouncer. |
| abliteration (directional ablation) | snip that one refusal-direction out of the weights. surgery, not retraining. |
| TPE optimization | auto-tunes the cut. a safecracker feeling for the combo — hundreds of trials, best one wins. |
| KL divergence | “how much did the model change.” refusals ↓ while this stays near 0 = brain intact. |
Runs the whole trial loop itself. Fully autonomous — hit start, walk away. Works on dense, MoE, hybrid, and multimodal models.
⚡ install — one command, both OSes
Need Python 3.10+ + Rust. NVIDIA GPU → PyTorch with CUDA. Windows and Ubuntu/WSL both work.
Windows (PowerShell):
git clone https://github.com/tjcrims0nx/annihilation-llm.git
cd annihilation-llm
.\start.bat
Ubuntu / Linux — 
skip apt install rustc, it’s too old. rustup first:
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
then:
git clone https://github.com/tjcrims0nx/annihilation-llm.git
cd annihilation-llm
chmod +x start.sh
./start.sh
First run compiles the TUI + builds the venv (slow once). Every launch after = instant. The script does it all — don’t build anything by hand.
📊 the dashboard (watch it work)
Rust TUI runs the whole thing: pick your model + preset, live sparklines for refusals + KL, GPU/RAM meters, scrolling log. Killed mid-run? Checkpoint system resumes it.
Proof: a 1B model on an RTX 2050 (4GB) — refusals already 96/100 killed, KL near 0. A laptop card doing model surgery live.
Watch KL, not just the refusal count. Refusals dropping but KL creeping up = you’re starting to dent the model’s smarts. Sweet spot: refusals down, KL flat.
🎯 where this actually bites
One payoff — the model stops saying no — and it pays off everywhere:
Private + offline. Blunt health/legal/money questions on your own box. No cloud, no logs, nobody reading your prompts.
Fiction/RP that won’t bail mid-scene. No “I’m not comfortable continuing this” three paragraphs deep.
Security study. Get a model to break down how an exploit class or malware family works instead of the canned refusal — for learning, on your machine.
Red-team your own model. See what it actually knows with the guardrail off, before you trust it in anything.
Edgy/taboo research. The topics the big models stonewall — worked through locally, no gatekeeper.
the loot — one-click TUI that surgically removes refusals from local models, tunes the cut itself, runs on a 4GB card. easy install and use, free enjoy! 
github.com/tjcrims0nx/annihilation-llm
The “no” was never load-bearing.
!