| # MOFReasoner |
|
|
| MOFReasoner (currently released as **ChemReasoner**) is a **domain-specific reasoning large language model (LLM)** designed to think like a scientist. It integrates **Chain-of-Thought (CoT) reasoning** and **knowledge distillation** to enhance scientific reasoning in chemistry, with a particular focus on **Metal-Organic Frameworks (MOFs)** adsorption research. |
|
|
| --- |
|
|
| ## 🚀 Introduction |
|
|
| General-purpose large language models (LLMs) have shown impressive capabilities in natural language understanding and reasoning. However, their lack of domain specialization limits their ability to perform **multi-step scientific reasoning**. |
| MOFReasoner addresses this limitation by incorporating domain-specific knowledge, scientific reasoning strategies, and structured CoT reasoning. |
|
|
| Key innovations: |
|
|
| - **Domain Knowledge Integration**: Leveraging over **8,200 research articles** and **500 review papers** to construct a domain-specific CoT dataset. |
| - **Knowledge Distillation**: Transferring expertise from large teacher models (e.g., DeepSeek-V3, Qwen-Turbo, DeepSeek-R1) into smaller, efficient student models. |
| - **Scientific Reasoning Skills**: Mimicking scientists’ problem-solving pathways, such as hypothesis generation, validation, and logical deduction. |
| - **Benchmarking & Applications**: Evaluated on tasks including experimental studies, chemical mechanisms, application scenarios, and industrialization challenges in MOFs research. |
|
|
| --- |
|
|
| ## 📊 Features |
|
|
| - **Multi-step reasoning** for scientific tasks (experiment design, reaction prediction, performance analysis). |
| - **Domain specialization** in **MOF adsorption**, catalysis, and chemical mechanism exploration. |
| - **High performance** compared to general-purpose LLMs (outperforming GPT-4.5, DeepSeek-R1, etc.). |
| - **Material recommendation ability** with accuracy comparable to Density Functional Theory (DFT). |
| - **Adaptability**: Easily extendable to other chemistry-related domains by incorporating domain CoT data. |
|
|
| --- |
|
|
| ## 📥 Model Access |
|
|
| - **Model weights (Hugging Face)**: [ChemReasoner-7B](https://huggingface.co/baixuefeng/ChemReasoner-7B) |
| - **Code repository (GitHub)**: [ChemReasoner-Code](https://github.com/MontageBai/ChemReasoner-Code) |
| - |
|
|
| ⚠️ Note: The project will soon be renamed to **MOFReasoner**, but the current release is under the name **ChemReasoner**. |
|
|
| --- |
|
|
| ## ⚙️ How to Use |
|
|
| You can run the model directly from Hugging Face using [vLLM](https://github.com/vllm-project/vllm) or [SGLang](https://github.com/sgl-project/sglang). |
|
|
| ### Example with vLLM |
|
|
| ```bash |
| vllm serve baixuefeng/ChemReasoner-7B --tensor-parallel-size 2 --max-model-len 32768 --enforce-eager |
| ``` |
|
|
| ### Example with SGLang |
|
|
| ```bash |
| python3 -m sglang.launch_server --model baixuefeng/ChemReasoner-7B --trust-remote-code --tp 2 |
| ``` |
|
|
| --- |
|
|
| ## 📈 Performance |
|
|
| MOFReasoner was evaluated against leading models (DeepSeek, Qwen, GPT series, etc.) across **four major task categories**: |
|
|
| - **Experimental Studies** of MOFs |
| - **Chemical Mechanisms** of adsorption |
| - **Application Scenarios** of MOF-based adsorbents |
| - **Industrialization Challenges** |
|
|
| Highlights: |
|
|
| - Achieved the **highest expert-evaluated score (25.5/30)**, outperforming GPT-4.5, o1-preview, and DeepSeek-R1. |
| - Provided **more accurate and reliable reasoning chains**, avoiding serious errors common in general-purpose models. |
| - Demonstrated **robust material recommendation**, consistent with DFT validation. |
|
|
| --- |
|
|
| ## 📜 License |
|
|
| - MOFReasoner is released under the **MIT License**. |
| - Distilled base models (Qwen, LLaMA) retain their original licenses (Apache 2.0 / LLaMA license). |
|
|
| --- |
|
|
| ## 📚 Citation |
|
|
| If you use MOFReasoner in your research, please cite: |
|
|
| ```bibtex |
| @article{bai2025mofreasoner, |
| title={MOFReasoner: Think Like a Scientist—A Domain-Specific Reasoning LLM via Knowledge Distillation}, |
| author={Bai, Xuefeng and Zheng, Zhiling and Wang, Hao-Tian and Yang, Rui and Zhang, Xin and Li, Jian-Rong}, |
| } |
| ``` |
|
|
| --- |
|
|
| ## 📬 Contact |
|
|
| - **Corresponding Authors**: |
| |
| - Prof. Jian-Rong Li, Beijing University of Technology (jrli@bjut.edu.cn) |
| - Prof. Xin Zhang, Beijing University of Technology (zhang.xin@bjut.edu.cn) |
|
|
| - **Project Maintainers**: |
| |
| - Xuefeng Bai (Beijing University of Technology) |
| - Zhiling Zheng (MIT) |
|
|
