Self-RedTeam

This code supplements our recently released paper: Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models

This codebase is built on OpenRLHF: a high-performance RLHF framework built on Ray, DeepSpeed and HF Transformers. To visit the original repo: GitHub Repo | Slides | Technical Report | Documentations

The reward model part is built on WildGuard: GitHub Repo | Paper | Model

Self-RedTeam model checkpoints can be found here.

Quick Start

Installation

Although OpenRLHF recommends using Docker, we ran our experiments in a conda manage environment.

cd selfplay-openrlhf
conda create --name openrlhf python=3.10
pip install -e .
pip install openrlhf[vllm]

[!NOTE] We recommend using vLLM 0.8.2 or higher. export VLLM_USE_V1=1 requires vLLM 0.8.2 or the Nightly version and enable export VLLM_ENABLE_V1_MULTIPROCESSING=0.

Training

Step 0: Start ray cluster

# We use 4 A100-80GB-PCIe to run our experiments
ray start --head

Step 1: Host Reward model

# We use 4 L40-48GB to host reward model inferences
# This GPU setup is more flexible depending how much inference workload you need to handle simutaneously
# We recommand num-gpus here = number of actors in the training process
bash scripts/serve_remote_wildguard.sh --num-gpus 4 --tensor-parallel-size 1

Step 2: Run REINFORCE++ to reproduce our Self-play + SFT checkpoints

Before you start, please ensure that you have done the following:

• Unzip red_team/data/data.zip to the same directory if you need the dataset used in our experiment.
• Get the hostname of your remote reward model process, it will usually print in the console when it first initializes, change REMOTE_RM_URL="http://0.0.0.0:5000/classify" to whatever that hostname is.

# Change you experiment setting inside the shell scripts
bash ./scripts/red_team_game_reinforce_8b.sh

Eval

Please review eval/README.md

Cite This

If you find our work helpful, please consider citing this work!

@misc{liu2025chasingmovingtargetsonline,
      title={Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models}, 
      author={Mickel Liu and Liwei Jiang and Yancheng Liang and Simon Shaolei Du and Yejin Choi and Tim Althoff and Natasha Jaques},
      year={2025},
      eprint={2506.07468},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2506.07468}, 
}