RINSE simulated 1H NMR spectra dataset (v1.0.0)

A fully synthetic dataset of one dimensional proton (1H) NMR spectra for training and evaluating deep learning models that restore NMR signals. The data was generated within the NMRflux Julia ecosystem and is used by the RINSE restoration model (Restoration of Interpretable NMR Spectra and Estimation ready signals), a demonstrator of the machine learning compatibility of NMRflux. Synthetic free induction decays (time domain length TD = 16384, sweep width SWH = 10000 Hz) were zero filled to 2^16 = 65536 points, given a first point half correction, apodised with the default exponential value 0.5*pi, and Fourier transformed with fftshift. Both the training and validation and the test spectra are fft shifted. The training and validation split is stored as paired 4096 point spectral chunks, with the clean target on odd rows and the corrupted input on even rows. The held out test split is stored as full 65536 point spectra in triplets: clean reference, reference plus additive noise, and reference plus noise and artefacts. Validation is provided at SNR 450, 600, 900, 1200, and 1500. The test set spans SNR 10 to 1800, with the training range treated as in domain and the unseen levels as stress tests. The data is provided in two formats holding identical values: the canonical JLD2 format used by RINSE, and a NumPy .npz mirror for Python users produced from the JLD2 files without rerunning the simulation. Julia and Python loader scripts, shape verifiers, a JLD2 to npz converter, and Jupyter notebook viewers are included, together with a README, LICENSE, CITATION.cff, and CHANGELOG. The dataset is fully synthetic. No measured spectra and no patient or sample data are included.