Suzhou Laboratory and Shanghai Jiao Tong University Jointly Release First Large Language Model for Organic Chemistry

Recently, Suzhou Laboratory and Shanghai Jiao Tong University unveiled ChemDFM, the first domain-specific large language model (LLM) targeting organic chemistry, with a scale of tens of billions of parameters. By integrating vast foundational chemical knowledge and mastering domain-specific terminology and expressions, this 13-billion-parameter model has surpassed the general-purpose GPT-4 in most chemistry-related professional capabilities.

The research team collected and curated nearly 4 million chemistry research papers, over 2,000 textbooks/reference books, and more than 1.7 million entries of molecular structure properties and chemical reaction data, constructing a corpus of 34 billion tokens. Through pre-training and instruction fine-tuning, ChemDFM demonstrated exceptional performance in six categories of chemical tasks—including molecular recognition, molecular description, description-based molecule generation, property prediction, reaction prediction & evaluation, and scientific Q&A—during third-party evaluations conducted by Zhejiang University and Tencent AI Lab.

These results highlight the advantages of vertical scientific LLMs.
The team is currently developing ChemDFM-X, a multimodal materials chemistry language model that will add capabilities for recognizing molecular graphs and spectral data. The continued advancement of large models in materials science is expected to accelerate the emergence of AI-powered research assistants, potentially shortening R&D cycles, reducing costs, and driving paradigm shifts in materials development. (Frontier Research Department, Scientific Research Department)