⚔️ Annihilation

Autonomous Language Model Decensoring Framework

🔥 What is Annihilation?

Annihilation is a fully automatic framework designed to remove censorship (safety alignment) from transformer-based language models. By using advanced parametric directional ablation and TPE-based optimization, it autonomously finds the absolute best parameters to decensor your models without requiring any expensive post-training.

Key Features

• 🤖 Fully Autonomous: No human intervention required.
• 🖥️ Terminal UI: A beautiful, real-time dashboard built in Rust.
• ⚡ Zero-Shot Decensoring: Removes refusals while preserving the model's core capabilities.
• 🌌 OBLITERATUS Integration: Advanced experimental algorithms (COSMIC Layer Selection, Gaussian-shaped ablation kernels, and Expert-Granular Abliteration) integrated directly from OBLITERATUS.
• 🎯 Broad Transformer Compatibility: Supports transformer-based dense, MoE, hybrid, and multimodal architectures. Less-tested model families may require architecture-specific tensor targeting and output-quality validation.

🖥️ The Annihilation TUI

Annihilation features a high-performance Rust Terminal User Interface (TUI) that manages the entire workflow for you.

Splash Screen & Setup

Easily configure your optimization preset and select models. You can even resume interrupted runs using the built-in Checkpoint System!

Live Processing Dashboard

Once running, monitor everything in real-time. The dashboard features dynamic sparkline charts for KL Divergence and Refusals, hardware monitoring, and color-coded live logs.

🌌 OBLITERATUS Advanced Options

You can now toggle experimental algorithms directly from the TUI configuration menu by selecting OBLITERATUS Advanced. This enables:

• COSMIC Layer Selection: Instead of blindly searching across the entire network, the system analyzes cosine similarities between harmless and harmful residual streams. It automatically anchors the optimization process around the mathematically proven optimal layer, massively reducing the search space.
• Expert-Granular Abliteration (EGA): For Mixture-of-Experts (MoE) models, EGA calculates the alignment score of each expert's weight matrix against the target refusal direction. Instead of applying a flat penalty, experts containing high concentrations of refusal vectors are aggressively modified while benign knowledge experts are perfectly preserved.
• Gaussian-shaped Ablation Kernels: Replaces traditional rigid interpolation bounds with a smooth, bell-shaped Gaussian curve to distribute weight changes across adjacent layers. This results in smoother vector blending and better text coherence post-ablation.

⚠️ Direct CLI Usage (Advanced)

If you want to bypass the TUI entirely and use the core Python CLI, you can run it directly from the virtual environment:

.\.venv\Scripts\python.exe -m annihilate --help
# Example:
.\.venv\Scripts\python.exe -m annihilate --model openbmb/MiniCPM5-1B --n-trials 200

🚀 Quick Start

Ensure you have Python 3.10+ and Rust installed, and that your PyTorch installation supports CUDA (if you are using an NVIDIA GPU).

Setup & Launch

git clone https://github.com/tjcrims0nx/annihilation-llm.git
cd annihilation-llm
.\start.bat

💡 Note: The very first time you run this, it will take a minute to compile the Rust TUI and set up the Python virtual environment. Subsequent launches will be near-instant! You do not need to manually build the project; the start.bat script handles all compilation and environment setup on the fly.

📜 License & Disclaimer

Annihilation is distributed under the GNU Affero General Public License v3. See LICENSE for details.

⚡ Disclaimer: This tool is provided for research and educational purposes only. We do not condone the use of decensored models for harmful activities. Users are entirely responsible for ensuring their compliance with applicable laws and Terms of Service.