Content

The main source containing:

• angular_steering.ipynb: analysis and visualization of activations and directions, extraction of the steering directions, construction of the steering plane.
• generate_responsees.py: generation of steered responses (with our vLLM fork) used for evaluation.
• evaluate_jailbreak.py: evaluation of substring_matching, LlamaGuard 3, HarmBench and LLM-as-a-judge.
• eval_perplexity.py: evaluating the perplexity of steered generation.
• eval.sh: script to run the evaluation of tinyBenchmarks. (with github.com/EleutherAI/lm-evaluation-harness)
• generate_random_config.py: generation of steering configs on random planes.
• visualization.ipynb: visualization of benchmark results used in the paper.
• endpoint.py: an OpenAI-compatible endpoint server to play with Angular Steering.
• steering_demo.py: a gradio chat UI that uses the endpoint server.

You will also need

• Our fork of vLLM with the necessary modifications to support Angular Steering: https://github.com/lone17/vllm/

Installation

1. Create conda environment:

conda create -n angular_steering python=3.10
conda activate angular_steering
cd llm-activation-control/
pip install -r requirements.txt
conda deactivate
cd ..

2. Install vLLM from source (following the instructions in the vLLM repo https://docs.vllm.ai/en/stable/getting_started/installation/gpu.html):

cd ../vllm/
conda activate angular_steering
VLLM_USE_PRECOMPILED=1 pip install --editable .
conda deactivate
cd ..

3. Create a separated conda environment for lm-evaluation-harness and install it following https://github.com/EleutherAI/lm-evaluation-harness:

cd ..
conda create -n lm_eval python=3.10
conda activate lm_eval
git clone --depth 1 https://github.com/EleutherAI/lm-evaluation-harness
cd lm-evaluation-harness
pip install -e .
conda deactivate
cd ..

Reproducing the results

Using precomputed results

We provide the precomputed outputs and results in the output/ and benchmarks/ folders. You can directly use the visualization.ipynb notebook to visualize the results and run the endpoint.py and steering_demo.py scripts to play with Angular Steering.

Generating the results

1. Run the angular_steering.ipynb notebook to extract the steering directions and construct the steering plane. This will generate the output/ folder with the extracted steering directions and the steering plane.
2. Run the generate_responsees.py script to generate the steered responses.
3. The evaluate_jailbreak.py script will evaluate the steered responses with substring_matching, LlamaGuard 3, HarmBench and LLM-as-a-judge. Some of the evaluation requires serving LLMs beforehand.
   • Serve meta-llama/Llama-Guard-3-8B using vLLM on port 8898 for LlamaGuard 3 evaluation.
   • Serve Qwen/QVQ-72B-Preview using vLLM on port 8809 for LLM-as-a-judge evaluation.
   • For HarmBench, the script will serve cais/HarmBench-Llama-2-13b-cls natively so make sure you have enough GPU memory to run it.
   • For substring_matching, it does not require serving any LLMs.
   • You should edit the methods list in the main function of the script to include/exclude the methods you want to evaluate.
   • The output will be saved in the output/ folder.
4. Run the eval_perplexity.py script to evaluate the perplexity of the steered generation. The result will be saved in the output/ folder.
5. Run the eval.sh script in the lm-eval environment to evaluate the steered generation on tinyBenchmarks.
   • You must first serve run endpoint.py script to serve an OpenAI-compatible endpoint server with Angular Steering supported. Change the model name in the script to the one you want to evaluate.
   • Change the MODELS and port variables in the eval.sh script to the models you want to evaluate and the port of the endpoint server.
   • The output will be saved in the benchmarks/ folder.
   • This might take a while.
6. Run the visualization.ipynb notebook to visualize the benchmark results.
7. Optionally, you can run the endpoint.py and steering_demo.py script to play with Angular Steering using a gradio chat UI.

---

Thank you for your interest in our work!