Dataset for "Reasoning Language Model as Rule Finder: A Case Study on Iron-Catalyzed C–H Bond Activation using 2D Metal-Organic Frameworks"

Unraveling structure-activity relationship in catalysis requires interpretable models that can extract governing principles from complex datasets. In this study, we investigate the role of reasoning large language models (LLMs) as rule finders in predicting the outcomes of C(sp3)–H bond activation. The reaction was catalyzed using two-dimensional metal-organic frameworks (MOFs) with Fe centers anchored on terpyridine ligands. Surface modifications with molecular additives systematically modulate the catalytic microenvironment, influencing reaction selectivity and efficiency. However, finding the relationship between the modifier structure and activity is a challenge. While traditional descriptors offer high predictive accuracy, LLM-derived rules provide interpretable insights tailored to the dataset. By integrating LLM reasoning with experimental features such as Fe-loading and modifier ratios, we identify dominant factors governing catalytic performance. This study highlights the potential of LLMs as tools for deriving chemically meaningful rules, bridging data-driven predictions with mechanistic understanding in catalyst design. This dataset only includes the experimental data of this paper. All the codes and conversions of LLMs are stored at https://github.com/Wang-Group/Reasoning-Language-Model-as-Rule-Finder