A Cleaned Multi-Source MS/MS Dataset for Single-Modality Mass Spectral Retrieval and Cross-Instrument Evaluation

This dataset provides a curated benchmark for single-modality tandem mass spectrometry (MS/MS) retrieval tasks, aiming to evaluate the generalization ability of machine learning models across different instruments and data sources. The dataset is collected from several widely used public mass spectrometry repositories, including: GNPS MoNA MassBank MassSpecGym MTBLS1572 The spectra are organized according to instrument types, primarily including Orbitrap and QTOF instruments. To construct a reliable benchmark for retrieval models, we adopt a strict non-overlapping training–testing protocol: The GNPS-Orbitrap subset is used exclusively as the training dataset. All remaining datasets are used only for evaluation, including: GNPS-QTOF GNPS-Other instruments MoNA-Orbitrap MoNA-QTOF MTBLS1572 MassBank (Orbitrap / QTOF / Orbitrap+QTOF combined ) MassSpecGym (Orbitrap / QTOF / Orbitrap+QTOF combined ) To prevent data leakage between training and testing, all datasets except GNPS-Orbitrap have been carefully cleaned to remove spectra corresponding to compounds already present in the GNPS-Orbitrap training subset. As a result, no overlapping compounds exist between the training and evaluation sets. This dataset is designed to support research in: Single-modality mass spectrum retrieval Mass spectral representation learning Cross-dataset generalization Cross-instrument robustness evaluation By integrating multiple public repositories and enforcing strict compound-level deduplication, this benchmark provides a realistic setting for evaluating retrieval models in real-world mass spectrometry applications.