This repository contains the code and resources for "Reasoning-Driven Retrosynthesis Prediction with Large Language Models via Reinforcement Learning," which introduces RetroDFM-R.

• [2025年11月22日] We released the parameters of RetroDFM-R-8B on 🤗 Hugging Face.
• [2025年07月23日] The paper of RetroDFM-R is available on arXiv: Reasoning-Driven Retrosynthesis Prediction with Large Language Models via Reinforcement Learning.

We recommend using the provided Docker image for training. Follow the installation instructions for OpenRLHF to set up your environment.

# Follow OpenRLHF installation guide
# https://github.com/OpenRLHF/OpenRLHF/tree/main?tab=readme-ov-file#installation
# We recommend using the Docker image for training.

Once inside the Docker container, install rdkit:

pip install rdkit

To set up the environment for inference, follow these steps:

conda create -n retrodfmR python=3.10
conda activate retrodfmR
pip install -r requirements.txt

You can specify the CUDA version if needed (e.g., for CUDA 12.8):

pip install vllm --extra-index-url https://download.pytorch.org/whl/cu128 # Replace cu128 with your CUDA version

All data used in this work are sourced from publicly accessible datasets:

• SMILES/IUPAC Name Conversion Data: Paired SMILES and IUPAC names are obtained from PubChem, used to construct the name conversion data.
• Retrosynthesis Data:
  • USPTO-50K: Accessed via the GLN repository: https://github.com/Hanjun-Dai/GLN (specifically, the schneider50k dataset).
  • USPTO-FULL: Also obtained from the GLN repository: https://github.com/Hanjun-Dai/GLN (specifically, uspto_multi dataset).

We provide the processed test data on Hugging Face: https://huggingface.co/datasets/OpenDFM/retrodfm-R-inference

Ensure your Docker container is successfully launched before initiating training.

Navigate to the train directory:

cd train

Then, execute the following training script:

• For Continual Pretraining:

  bash examples/scripts/train_continual_pretrain.sh

• For Cold-Start Distillation:

  bash examples/scripts/train_cold_start_distill.sh

• For Reinforcement Learning:
  • On USPTO-50K:

    bash examples/scripts/train_dapo_retrodfm_R_50k.sh

  • On USPTO-FULL:

    bash examples/scripts/train_dapo_retrodfm_R_full.sh

After downloading the processed test data (as mentioned in Data Preparation), you can run the inference script. The following command will perform inference using beam search and test augmentation:

conda activate retrodfmR
cd inference && bash eval.sh

Please cite our paper if you find our work useful:

@misc{zhang2025retrodfmr,
 title={Reasoning-Driven Retrosynthesis Prediction with Large Language Models via Reinforcement Learning},
 author={Zhang, Situo and Li, Hanqi and Chen, Lu and Zhao, Zihan and Lin, Xuanze and Zhu, Zichen and Chen, Bo and Chen, Xin and Yu, Kai},
 year={2025},
 eprint={2507.17448},
 archivePrefix={arXiv},
 primaryClass={cs.CE},
 url={https://arxiv.org/abs/2507.17448}, 
}