README.md

metadata

license: cc-by-nc-4.0
datasets:
  - NingLab/MuMOInstruct
language:
  - en
base_model:
  - mistralai/Mistral-7B-Instruct-v0.3
pipeline_tag: text-generation
tags:
  - chemistry
  - molecule optimization
library_name: transformers

Model Sources

• Repository: https://github.com/ninglab/GeLLMO
• Paper: https://arxiv.org/abs/2502.13398

Usage

For instructions to run the model, please refer to our repository.

Bias, Risks, and Limitations

While our models are designed for research and drug discovery applications, they come with ethical and safety considerations:

1. Potential for Misuse: Although the model is not explicitly designed to generate toxic, controlled, or harmful compounds, adversarial prompts or unintended biases in the pretrained model may lead to the generation of undesirable molecules.
2. Unintended Harmful Outputs: The model does not inherently filter out molecules with high toxicity, abuse potential, or environmental hazards. Users must implement additional safeguards to prevent misuse.
3. Absence of Built-in Safety Mechanisms: The model does not incorporate explicit regulatory or safety filters (e.g., toxicity or compliance checks). It is the responsibility of users to validate generated molecules for safety and ethical considerations.

We urge users to adopt best practices, including toxicity prediction pipelines, ethical oversight, and responsible AI usage policies, to prevent harmful applications of this model.

Citation

If you use the trained model checkpoints, datasets or other resources, please use the following citation:

@article{dey2025gellmo,
      title={GeLLMO: Generalizing Large Language Models for Multi-property Molecule Optimization}, 
      author={Vishal Dey and Xiao Hu and Xia Ning},
      year={2025},
      journal={arXiv preprint arXiv:2502.13398},
      url={https://arxiv.org/abs/2502.13398}, 
}