Chemical Language Model Linker datasets and models

Dataset

The unfiltered version of the PubChem dataset used for evaluation in the Chemical Language Model Linker (ChemLML) manuscript. The original dataset comes from the PubChem database. If you use this dataset, please see the PubChem download policies and citation guidelines.

There are entires for 257,619 chemicals, each with the fields:

• description
• Name
• CID
• ANID
• SMILES
• SELFIES

The PubChem dataset is available under the Creative Commons Zero v1.0 Universal license.

Relevant citations:

• S. Kim, J. Chen, T. Cheng, A. Gindulyte, J. He, S. He, Q. Li, B. A. Shoemaker, P. A. Thiessen, B. Yu, L. Zaslavsky, J. Zhang, E. E. Bolton, PubChem 2023 update. Nucleic Acids Research 51, D1373–D1380 (2023).
• Y. Deng, S. S. Ericksen, A. Gitter, Chemical Language Model Linker: blending text and molecules with modular adapters. Journal of Chemical Information and Modeling (2025).

Models

The `.pth` files are saved PyTorch models. The filenames correspond to the ChemLML models in Table 1 of the ChemLML manuscript. These ChemLML models use the following models from Hugging Face:
- MolT5
- Text+Chem T5
- MolGen
- MolGen-7B
- Fine-tuned LLaMA2-7B
- SCIBERT
- Galactica
- MolXPT

See the Hugging Face model cards for the original models' licenses, limitations, and citations.

The models are available under the Creative Commons Attribution 4.0 International license.