PharMolix/BioMedGPT-Mol

BioMedGPT-Mol

BioMedGPT-Mol is a multimodal molecular language model jointly released by PharMolix Inc. and the Institute of AI Industry Research (AIR), Tsinghua University. It is built for both molecular understanding and generation, supporting a wide range of tasks including chemical name conversion, molecular captioning, property prediction, reaction modeling, molecule editing, and property optimization. Trained with a well-structured multi-task curriculum, BioMedGPT-Mol shows remarkable performance across diverse molecule-centric discovery benchmarks. More technical details can be found in the technical report.

Get started

• Download the model and config files.

• Evaluation on Benchmarks

  • The test set is available in testset. If you use the dataset for evaluation, please consider citing the related papers:

    @article{yu2024llasmol,
        title={Llasmol: Advancing large language models for chemistry with a large-scale, comprehensive, high-quality instruction tuning dataset},
        author={Yu, Botao and Baker, Frazier N and Chen, Ziqi and Ning, Xia and Sun, Huan},
        journal={arXiv preprint arXiv:2402.09391},
        year={2024}
    }
    
    @article{li2024tomg,
        title={TOMG-Bench: Evaluating LLMs on text-based open molecule generation},
        author={Li, Jiatong and Li, Junxian and Liu, Yunqing and Zhou, Dongzhan and Li, Qing},
        journal={arXiv preprint arXiv:2412.14642},
        year={2024}
    }
    
    @article{dey2025mathtt,
        title={$$\backslash$mathtt $\{$GeLLM\^{} 3O$\}$ $: Generalizing Large Language Models for Multi-property Molecule Optimization},
        author={Dey, Vishal and Hu, Xiao and Ning, Xia},
        journal={arXiv preprint arXiv:2502.13398},
        year={2025}
    }
    
    @article{biomedgpt-mol,
        title={BioMedGPT-Mol: Multi-task Learning for Molecular Understanding and Generation},
        author={Zuo, Chenyang and Fan, Siqi and Nie, Zaiqing},
        journal={arXiv preprint arXiv:2512.04629},
        year={2025}
    }
    

  • Update the configuration and run inference using the provided scripts, and the outputs will be saved in the logs directory.

    - logs
    ---- biomedgpt_mol
    -------- mumoinstruct
    ------------ logs
    ------------ results
    -------- openmolinst
    ------------ logs
    ------------ results
    -------- smolinstruct
    ------------ logs
    ------------ results
    

    # SMolInstruction
    bash evaluation/scripts/inference_smolinstruct.sh
    
    # OpenMolInstuct
    bash evaluation/scripts/inference_openmolinst.sh
    
    # MuMoInstruct
    bash evaluation/scripts/inference_mumoinstruct.sh
    

  • Update the configuration accordingly and execute the evaluation scripts. The computed metrics will be stored as metrics.json in the results directory, e.g., /logs/biomedgpt_mol/mumoinstruct/results/metrics.json.

      # SMolInstruction
      bash evaluation/scripts/evaluate_smolinstruct.sh
    
      # OpenMolInstuct
      bash evaluation/scripts/evaluate_openmolinst.sh
    
      # MuMoInstruct
      bash evaluation/scripts/evaluate_mumoinstruct.sh
    

• 🔥Explore our OpenBioMed platform for more discovery tasks.

Cite Us

If you find our open-sourced models helpful to your research, please consider citing:

@article{biomedgpt-mol,
  title={BioMedGPT-Mol: Multi-task Learning for Molecular Understanding and Generation},
  author={Zuo, Chenyang and Fan, Siqi and Nie, Zaiqing},
  journal={arXiv preprint arXiv:2512.04629},
  year={2025}
}