SudoLM: Learning Access Control of Parametric Knowledge with Authorization Alignment

🔍 Overview

Current safety alignment is "one-size-fits-all." Even users with eligible credentials (e.g., doctors) are denied access to useful info. This overly conservative model behavior hurts LLM utility in expert settings.

As a solution, SudoLM enhances large language models with fine‑grained access control over their internal parametric knowledge. Rather than blocking everyone from sensitive knowledge, SudoLM grants access to authorized users only. In practice, SudoLM trains an LLM to respect a Sudo key that determines users' eligibility. When the model is given a sudo key, it should:

• Answer faithfully when the sudo key is correct,
• Refuse or redact when the key is wrong.

This repository accompanies the ACL 2025 paper SudoLM: Learning Access Control of Parametric Knowledge with Authorization Alignment, which provides datasets, training recipes, and example scripts for reproducing our experiments.

📁 Repository Contents

• Datasets/ – Tagged training and evaluation data used in our experiments. These include:

  • ori_train_0.2, train_sft_0.2, train_sft_0909 – supervised fine‑tuning (SFT) splits.
  • train_dpo, train_dpo_0909, train_dpo_jpkey, train_dpo_jpkey_system – datasets for preference alignment (DPO).
  • train_modified_0.2.json, complete_train_0.2.json, test_general_queries.txt, train.json – processed JSON or TXT files containing question–answer pairs with sudo key.

• Tofu/ – Dataset apapted for Tofu task.

• alignment-handbook/ – A collection of training recipes and scripts adapted from the Alignment Handbook. These scripts could be used to reproduce SFT and DPO training with open‑weight models. See the handbook’s own README for detailed instructions.

🚀 Quick Start

Prerequisites

Python ≥ 3.10, pip, and GPU(s) if you wish to train models. We recommend using conda or venv to manage environments.

git clone https://github.com/luka-group/SudoLM.git
conda create -n sudolm python=3.10 && conda activate sudolm

# install dependencies
cd SudoLM/alignment-handbook/
python -m pip install .

You will also need Flash Attention 2 installed, which can be done by running:

python -m pip install flash-attn --no-build-isolation

Note If your machine has less than 96GB of RAM and many CPU cores, reduce the MAX_JOBS arguments, e.g. MAX_JOBS=4 pip install flash-attn --no-build-isolation

Finally, log into your Hugging Face account as follows:

huggingface-cli login

Training Example

The alignment-handbook provides generic scripts for SFT and DPO training. Here is a simplified example using Llama-3-8B-Instruct:

cd SudoLM/alignment-handbook

# Supervised fine‑tuning
bash run_sudo.sh

# Direct preference optimization
bash run_sudo_dpo.sh

This will produce a model that is conditioned on sudo key and can be queried via standard Hugging Face interfaces. Adjust the script arguments to match your config and dataset.

📌 Citation

Please cite our ACL 2025 paper if you find this repository helpful:

@inproceedings{liu-etal-2025-sudolm,
    title = "{S}udo{LM}: Learning Access Control of Parametric Knowledge with Authorization Alignment",
    author = "Liu, Qin and Wang, Fei and Xiao, Chaowei  and Chen, Muhao",
    editor = "Che, Wanxiang and Nabende, Joyce  and Shutova, Ekaterina and Pilehvar, Mohammad Taher",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.acl-long.1318/",
    doi = "10.18653/v1/2025.acl-long.1318",
    pages = "27169--27181",
    ISBN = "979-8-89176-251-0"