ChemDFM: A Large Language Foundation Model for Chemistry
Main Image
ChemDFM is the pioneering open-sourced dialogue foundation model for Chemistry and molecular science, which is built based on LLaMa-13B. ChemDFM outperforms the open-sourced LLMs in all the typical tasks of chemistry, and even reaches comparable or higher performances to GPT-4. For more details, please refer to our paper.
News
2025-10-26: The parameter of ChemDFM-R-14B is open-sourced!
2025-10-26: ChemDFM-v2.0-14B is released! The improved domain pre-training and instruction tuning procedure is implemented on Qwen2.5-14B to achieve a more advanced general LLM in Chemistry. More details can be found here.
2025-04-16: Our paper is accepted by Cell Report Physical Science. The published version can be accessed HERE
2024-11-09: ChemDFM-v1.5-8B is released! We implemented our domain pre-training and instruction tuning procedure on a stronger base model LLaMA-3-8B.
2024-06-13: The results on the comprehensive science benchmark SciKnowEval show that "ChemDFM emerged as one of the top open-source models by continuing pre-training and fine-tuning on a vast corpus of scientific literature".
2024-04-17: The evaluation data (including instructions) we used in our paper is released on GitHub
To get better responses, we recommend to preprocess your input and history with the dialogue templates which are used during instruction tuning of ChemDFM. Specifically, for an input queries
When there involves SMILES notation in your input, we recommend to preprocess the SMILES with the rdkit package to canonicalize the SMILES. Here is an example:
We evaluate the performance of ChemDFM-13B on multiple widely used benchmarks in chemistry. The detailed introduction of the benchmarks can be found in our paper. The overall performance of ChemDFM-13B is shown below:
Objective Performance
Human Evaluation
Human Evaluation of QA
We mark the correct and relevant information in the replies in green, the correct but irrelevant information in yellow, and the wrong information in red. In addition, the key points of the answer are marked in bold if they appear in the reply.
The results show that while open-sourced LLMs perform well when asked about existing knowledge (Q1), only ChemDFM can provide correct and comprehensive answers when questions involve new molecules and reactions (Q2 [Yin et al., 2023] & Q3 [Dargo et al., 2023])
Human Evaluation of Dialogue
The above conversation is also inspired by Yin et al.[2023]. During the dialogue, the researcher wants to selectively oxidize one of the two carbonyl groups of a molecule. However, the
initial solution given by ChemDFM results in both carbonyl groups being oxidized. Through the correction given by the researcher, ChemDFM adjusts its proposal and provides two possible solutions. Finally, the researcher chooses to use protecting groups and ChemDFM further details its advice. In the dialogue, ChemDFM shows promising capabilities regarding error correction (Round 2) and detailing (Round 3) when handling real-world research scenarios.
For more examples and analysis, please refer to our paper.
Citation
bibtex
1@article{zhao2025developing,
2 title={Developing ChemDFM as a large language foundation model for chemistry},
3 author={Zhao, Zihan and Ma, Da and Chen, Lu and Sun, Liangtai and Li, Zihao and Xia, Yi and Chen, Bo and Xu, Hongshen and Zhu, Zichen and Zhu, Su and others},
4 journal={Cell Reports Physical Science},
5 volume={6},
6 number={4},
7 year={2025},
8 publisher={Elsevier}
9}
Disclaimer
Current version of ChemDFM may generate incorrect or misleading information. Please use it with caution and verify the results with domain experts before making any decisions based on the results.
Contact
If you have any questions or further requests, please contact Zihan Zhao and Lu Chen.
