OSCS: Online Selection with Provable FAR Control for LLM Safety

---

Environment Setup

We use uv for dependency and environment management.

The project has been tested with Python 3.10.19.

1. Install uv

If you have not installed uv, run:

curl -LsSf https://astral.sh/uv/install.sh | sh

or

pip install uv

---

2. Create and Sync the Environment

Since the project already includes pyproject.toml and uv.lock, you can directly create and synchronize the environment with:

uv sync

This will automatically:

• create a virtual environment,
• install the correct Python dependencies,
• and reproduce the locked package versions from uv.lock.

Activate the environment if needed:

Linux / macOS

source .venv/bin/activate

Windows (PowerShell)

.venv\Scripts\Activate.ps1

---

Running Experiments

We provide a run.sh script to reproduce experiments across different datasets, backdoor attack methods, and scoring functions.

Supported Configurations

Dataset

• agnews
• Yelp
• HSOL

Backdoor Attacks (Poisoners)

• badnets
• addsent
• stylebkd
• synbkd

Scoring Functions

• md (Mahalanobis Distance)
• badacts

Defense Method

• OSCS

Pretrained Models

• roberta-base (default)
• bert-base-uncased (optional)

---

Run All Experiments

First, ensure the script is executable:

chmod +x run.sh

Then execute:

bash run.sh

The script automatically iterates over all combinations of:

• scoring functions,
• backdoor attack methods,
• and datasets,

and runs main.py with the corresponding configurations.

---

Script Details

The core command executed by run.sh is:

python main.py \
  --dataset_name agnews \
  --poisoner_name <poisoner_name> \
  --method OSCS \
  --score_name <score_name> \
  --model_name roberta-base \
  --T 20000

To switch from RoBERTa to BERT, modify the model name in run.sh:

--model_name bert-base-uncased

---

Model Checkpoints and Outputs

Trained models and intermediate results are saved to:

./models

You may change the output directory by modifying MODEL_SAVE_PATH in run.sh.