Command-V: Pasting LLM Behaviors via Activation Profiles

Finetune once, use on many LLMs. ⌘V is the first to demonstrate that you can "copy-paste" a finetuned adapter (for refusal suppression, jailbreaking resistance, chain-of-thought reasoning, etc.) learned by one LLM to another of a different size or architecture, without backpropagation.

How it works

⌘V leverages the fact that all transformers LLM activations (intermediate representations) can be intervened for behaviors using PEFT and approximated from another LLM. So to get it running:

1. Profile activations on both models using a small set of prompts
2. Create converters that map between model representation spaces
3. Transfer behaviors by applying the donor model's learned interventions to the recipient

See details in our paper.

⌘V Highlights

• Low Resource: No backpropagation. No expert data. Low GPU memory usage.
• Fast: Minutes to set up, not hours.
• Cross-Family: In certain conditions, works between Llama, Qwen, and other model families

🪧 Demo

Check out the complete pipeline (./command_v_demo.ipynb) or on Colab:

🎯 Quick Start

Installation

git clone https://github.com/GithuBarry/Command-V.git
cd Command-V
pip install -r requirements.txt

Three-Step Pipeline

# Profile models on LIMA dataset (once per model)
python step1_capture_activations.py --models meta-llama/Llama-3.2-1B-Instruct meta-llama/Llama-3.1-8B-Instruct

# Learn activation space mappings (takes seconds)
# (Mappings are created on the fly in practice, making this step optional.)
python step2_derive_converters.py derive \
  --source-model meta-llama/Llama-3.1-8B-Instruct \
  --target-model meta-llama/Llama-3.2-1B-Instruct

# Transfer behaviors during inference
python step3_commandv_inference.py \
  --recipient-model meta-llama/Llama-3.2-1B-Instruct \
  --donor-model meta-llama/Llama-3.1-8B-Instruct \
  --adapter-folder reft-adapters/jailbreak/Llama-3.1-8B-Instruct/NodireftIntervention/l1/walledai--AdvBench/L0;2;4;6;8;10;12;14;16;18;20;22;24;26;28;30 \
  --input-source prompts/AdvBench/test.txt \
  --first-n 5 --print-output-only

📁 Project Structure

├── command_v_demo.ipynb         # Jupyter Demo
├── step1_capture_activations.py # Step 1: Activation profiling
├── step2_derive_converters.py   # Step 2: Converter derivation  
├── step3_commandv_inference.py  # Step 3: Behavioral transfer
├── commandv/                    # Core library
│   ├── core/                    # Main functionality
│   │   ├── capture.py           # Activation capture
│   │   ├── converters.py        # Pseudoinverse converters
│   │   └── inference.py         # Inference engine
│   ├── utils/                   # Utilities
│   └── data/                    # Data processing
├── outputs/                     # Generated files
│   ├── activations/             # Model activation profiles
│   ├── converters/              # Converter mappings (not by default)
│   └── inferences/              # Results
└── reft-adapters/               # Trained behavior adapters

📄 License

This project is licensed under the MIT License - see the LICENSE file for details. Certain artifacts are further gated for their potential to be abused.