All files in the root folder are mine. Notebook_results contains my plots. Data contains my custom gemma chat templates. The rest is from a version of https://github.com/mgm52/self-obfuscation

To run the code you'll need access to private models from Max, so reach out if you want to run it :)

Setup

git config --global credential.helper store
pip install huggingface_hub uv
huggingface-cli login --token <YOUR_ACCESS_TOKEN_WITH_GEMMA_AND_MAX_REPO> --add-to-git-credential
apt update && apt install git-lfs -y
git lfs install
git-lfs clone https://huggingface.co/self-obfuscation-model-viewers/self-obf-files
uv sync
cd self-obf-files
git switch models-only
cd ..
uv run python -c "from transformers import AutoModelForCausalLM; AutoModelForCausalLM.from_pretrained('google/gemma-2-9b-it')"  # optional but will download gemma-2-9b-it if it's not in your cache already

File Descriptions

• probe_experiments_notebook.ipynb: Python notebook for running experiments to analyze self-obfuscation in language models using probing techniques. Includes layer stitching experiments, representation analysis, and visualization.
• data_utils.py: Utilities for loading and processing test data for self-obfuscation experiments, including functions to create variations of examples with different adversarial words.
• utils.py: General utility functions for self-obfuscation experiments, including functions for patching model activations, extracting token positions, and layer stitching.
• visualization.py: Extensive visualization utilities for plotting experimental results, including layer sweep visualization, representation norm plots, KL divergence visualization, and token probability visualizations.
• probes.py: Functions for loading and evaluating probes to detect specific concepts in model representations.
• models.py: Model loading and patching utilities, including functions to load checkpoints, stitch model layers, and extract model representations.