This dataset is the basis for a publication regarding the differentiation of fresh and frozen-thawed fish via nuclear magnetic resonance spectroscopy (https://doi.org/10.1007/s00217-025-04978-6). Food fraud along the production chain is a well-known issue that requires an effective authenticity control. For the differentiation of fresh and frozen-thawed fish, 1H nuclear magnetic resonance (NMR) spectroscopy based methods in combination with multivariate data analysis have proven to be suitable in principle. Here, from a total of 317 samples (cod, rainbow trout, mackerel; fresh and frozen-thawed), the lipid and polar fractions of the fish flesh were analyzed, and classification models based on a principal components analysis with linear discriminant analysis (PCA-LDA) including cross-validation were generated. Additionally, data fusions were carried out. The obtained average accuracies of > 90% (94.0% based on the lipid fraction, 92.8% based on the polar fraction) and > 95% (95.6% based on a low-level data fusion,95.5% based on a mid-level data fusion) demonstrated a promising differentiation. Further examinations confirmed that the non-targeted analysis appears to be mandatory as no marker substances were indicated in the loadings plots of the models. To evaluate whether the generated classification models are suitable to be used in a broader manner, they were applied to 13 fresh and 13 frozen-thawed samples from twelve other common edible fish species in a preliminary study. The classificationmodel based on the low-level data fusion gave the best results (84.6% of all 26 samples correctly predicted). Thus, although these models are very suitable for analyzing cod, rainbow trout, or mackerel for a classification as fresh or frozen-thawed, they cannot generally be applied to samples of other fish species. Datasets containing NMR spectra of samples from the classification models and of samples from other fish species are provided, as well as relevant metadata.