This repository contains the code and resources for training models on long-form reinforcement learning tasks for our paper Semantically-Aware Rewards for Open-Ended R1 Training in Free-Form Generation.

• Upload dev dataset.
• Upload sample model generated answers
• Upload vllm response inference code

[📖Paper] [🤗RewardBERTModel] [Training Curves]

To get started with this project, follow the steps below to clone the repository and set up your environment.

1. Clone the Repository

git clone https://github.com/zli12321/long_form_rl.git
cd long_form_rl/OpenRLHF

2. Install Dependencies

Install the necessary Python packages using pip.

pip install -e .
pip install qa-metrics

cd train_reward_bert
python reward_bert.py

Once the setup is complete, you can begin training the model using the provided scripts.

1. Navigate to the Training Scripts

cd ../scripts/no-cot

2. Configure Your Training Run

Before launching the training, you must edit the grpo_preferenceBert.sh script to match your environment settings.

Open grpo_preferenceBert.sh and update the following variables:

• working_dir
• remote_rm_url
• save_path
• use_wandb

3. Run the Training Script

./grpo_preferenceBert.sh

Evaluation procedures are currently under development and will be released soon.

The planned evaluation method involves using our provided template to prompt GPT-4o to generate scores twice for each output. The final score will be the average of the two generated scores.

• Please ensure all script paths and configurations are adjusted to fit your specific setup.
• If you encounter any issues or have questions, please feel free to open an issue or submit a pull request on our GitHub repository. We welcome your contributions!

If you find our work helpful for your research, please consider citing our work.

@misc{li2025semanticallyawarerewardsopenendedr1,
      title={Semantically-Aware Rewards for Open-Ended R1 Training in Free-Form Generation}, 
      author={Zongxia Li and Yapei Chang and Yuhang Zhou and Xiyang Wu and Zichao Liang and Yoo Yeon Sung and Jordan Lee Boyd-Graber},
      year={2025},
      eprint={2506.15068},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.15068}, 
}