Researchers at Hugging Face have released TSDS-Toolbox, a unified framework designed to measure and compare time-series dataset similarity. The toolbox addresses the fragmentation of existing implementations, enabling systematic and reproducible benchmarking across forecasting, classification, and generation tasks.
Time-series dataset similarity plays a crucial role in selecting source datasets for fine-tuning foundation models. However, prior tools were often isolated and difficult to extend. TSDS-Toolbox offers flexible extensibility, allowing users to integrate custom datasets, similarity methods, and downstream tasks with ease.
The framework supports both dataset-level and series-level similarity methods through integrated reducers, ensuring consistent evaluation. Comprehensive experiments under diverse settings validate its effectiveness, and the toolbox is publicly available for the research community.
This release comes amid growing interest in time-series foundation models, as highlighted by related papers recommended by the Semantic Scholar API, including works on topological fidelity, spectral retrieval-augmented forecasting, and large-scale multivariate corpora.