NMRexp: A database of 3.3 million experimental NMR spectra

Although NMR spectroscopy is one of the most fundamental methods for structural characterization in organic chemistry, high-quality and large-scale experimental NMR datasets remain a critical bottleneck for training robust, generalizable spectral-prediction and structure-elucidation models. Here, we introduce NMRexp, an open, curated collection of 3.3 million experimental NMR records for six nuclei (¹H, ¹³C, ¹⁹F, ³¹P, ²⁹Si, ¹¹B), extracted from nearly 200 thousand Supporting Information documents published between 2010 and 2024. Manual evaluation confirmed > 99% accuracy in metadata extraction (chemical shifts, multiplicities, J-couplings, solvents) and 98% correctness in molecular-skeleton assignment; internal consistency tests yielded mean absolute errors of 0.026 ppm for ¹H and 0.206 ppm for ¹³C chemical shifts. NMRexp surpasses existing public NMR databases by over an order of magnitude in size, while providing rich spectral annotations and source-DOI traceability. We anticipate that NMRexp will accelerate the development of next-generation data-driven spectral prediction, automated structure elucidation, and related AI-enabled applications.