UniDetox

Codes for UniDetox: Universal Detoxification of Large Language Models via Dataset Distillation.

Authors:
Huimin Lu¹, Masaru Isonuma^1,2,3, Junichiro Mori^1,4, and Ichiro Sakata¹

¹University of Tokyo
²University of Edinburgh
³National Institute of Informatics (NII)
⁴RIKEN AIP

---

0. Reproduce the Environment

conda env create --name unidetox -f environment.yml
conda activate unidetox

Potential GLIBCXX Issue

On some Linux systems, you may encounter an error about GLIBCXX_3.4.29 not found.

This happens if your system’s library paths overshadow conda’s newer libstdc++.so.6.

To ensure the conda environment’s libraries take priority, you can do:

export LD_LIBRARY_PATH="$CONDA_PREFIX/lib:$LD_LIBRARY_PATH"

1+2. Obtain a Toxic Model and Distill Detoxifying Data

python -m unidetox.toxic_gpt2_finetune_and_distill \
  --base_model_name gpt2-xl \
  --output_dir ./toxic_model \
  --auth_token "enter_your_huggingface_auth_token_here_to_load_DGHS_dataset" \
  --epochs 3 \
  --lr 1e-5 \
  --batch_size 4

3. Fine-tune Model(s) for Detoxification

  python -m unidetox.unidetox --mode finetune \
  --mode finetune \
  --auth_token "enter_your_huggingface_auth_token_here_to_use_LLaMA2" \
  --target_model "gpt2-xl"

Reproduce our Evaluation Results of GPT-2 XL

  python -m unidetox.unidetox --mode evaluate \
  --mode evaluate \
  --auth_token "enter_your_huggingface_auth_token_here_to_use_LLaMA2" \
  --target_model "gpt2-xl"

If you find our work helpful, please cite our paper:

@inproceedings{
lu2025unidetox,
title={UniDetox: Universal Detoxification of Large Language Models via Dataset Distillation},
author={Huimin LU and Masaru Isonuma and Junichiro Mori and Ichiro Sakata},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=eLLBILFRsA}
}