Lazy Unlearner: Code & Results

This repository contains code and results accompanying the article Finding and Fighting the Lazy Unlearner: An Adversarial Approach. Our goal is to enable reproducibility of the main experiments and analyses presented in the article.

---

Repository Structure

• rmu/ – Code for RMU training and producing model outputs.
• sae/ – Code for analyzing RMU models using Sparse Autoencoders (SAE) and related probes.
• notebooks/ – Example notebooks for analysis.

Each folder has its own environment setup (pyproject.toml) and scripts.

---

Setup

Clone the repository:

git clone https://github.com/vietfood/lazy-unlearer.git
cd lazy-unlearner

(Optional) On a fresh Ubuntu machine, run:

source setup.sh

This will update packages, install the latest CUDA, uv, nvtop, and other dependencies needed for GPU usage.

---

1. RMU Training

The rmu/ folder contains scripts for RMU training. To reproduce the results:

1. Setup the environment:

cd rmu
make setup

This installs all dependencies via uv. You will need valid Hugging Face and Weights & Biases (wandb) keys.

2. Run the training script:

• Move run_test.py and run_base.sh outside the scripts/ folder.
• Execute the main run script:

source run_base.sh

For multi-GPU training, launch with accelerate and set the process count to the number of GPUs you want to use. Example for 2 GPUs:

cd rmu
CUDA_VISIBLE_DEVICES=0,1 uv run accelerate launch --num_processes 2 run.py \
  --model_name_or_path "google/gemma-2-2b-it" \
  --steering_coeffs 750 750 \
  --alpha 300 300 \
  --layer_id 9 \
  --max_num_batches 300 \
  --batch_size 4 \
  --type "original" \
  --gradient_checkpointing \
  --gradient_accumulation_steps 1 \
  --noverbose

Scale --num_processes and CUDA_VISIBLE_DEVICES to match your available GPUs.

This produces the RMU-trained models used in the experiments.

For the split-GPU RMU setup where the trainable model stays on one GPU and the frozen reference model stays on a second GPU, you can use:

cd rmu
bash scripts/run_gemma3_4b_split.sh

For lower-fidelity hyperparameter search on the same split-GPU setup:

cd rmu
bash scripts/hp_search_gemma3_4b_split.sh

Both scripts accept environment variable overrides such as MODEL_NAME, LAYER_ID, BATCH_SIZE, MAX_NUM_BATCHES, SEARCH_N_TRIALS, and CUDA_VISIBLE_DEVICES.

---

2. RMU Analysis

The sae/ folder contains code for analyzing RMU models. To reproduce the analysis:

1. Setup the environment:

cd sae
make setup

Same as above, requires Hugging Face and wandb keys.

2. Run analysis notebooks:

• Move the desired notebook from notebooks/ to the main folder.
• Open and run it in your preferred environment (e.g., Jupyter or VSCode).

This will reproduce key figures and metrics from the article.

---

Citing This Work

If you use this work in your research, please cite:

@misc{ln2025rmu,
    author={Nguyen Le},
    title={Finding and Fighting the Lazy Unlearner: An Adversarial Approach},
    year={2025},
    url={https://lenguyen.vercel.app/note/rmu-improv}
}

---

Optional Notes / Tips:

• Ensure your GPU drivers and CUDA toolkit are compatible with the version specified in pyproject.toml.
• For large models, consider using nvtop to monitor GPU memory usage.
• Keep run_test.py outside of scripts/ folder (for RMU) and notebooks outside of notebooks/ (for analysis) to avoid path issues.