--- pretty_name: TSC-FM Benchmark Datasets task_categories: - other tags: - time-series - classification - benchmark - univariate - multivariate - ucr - uea --- # TSC-FM Benchmark Datasets [Browse and download](dataset_index.csv) ยท [File manifest](manifest.json). This repository provides downloadable time series classification datasets used in the TSC-FM Benchmark, drawn from the UCR, UEA and historical MV-core archives. It is a dataset repository, separate from the [TSC-FM overview Space](https://huggingface.co/spaces/TSC-FM/benchmark). - [Official benchmark and results](https://tsc-fm.dmirlab.com/) - [Dataset profile](https://tsc-fm.dmirlab.com/benchmarks) - [Leaderboard](https://tsc-fm.dmirlab.com/leaderboard) Available datasets and split files are listed in the download index and manifest. The collection includes measured and synthetic/simulated classification datasets. ## Dataset selection and file identity For overlapping tasks between UEA and historical MV-core, the selected UEA copies are retained. CharacterTrajectories, JapaneseVowels and SpokenArabicDigits retain the selected UEA variable-length variants. Dataset membership and the stored TRAIN/TEST files are frozen to the benchmark copies; upstream archive updates do not silently replace them. Each `data//.zip` contains the original selected split files under the dataset name. File bytes are preserved. Archive-provided adjusted variants and `_eq`/`_nmv` filenames are intentional; these benchmark files are not necessarily raw acquisition streams. InsectWingbeatSubset is distributed as its four NumPy split arrays, preserving the benchmark's 200-channel, 22-time-point interpretation. Model-specific resizing, feature extraction and channel fusion are not baked into these packages. The `manifest.json` lists the dataset packages, compressed sizes and original files. Retain the published TRAIN/TEST membership. Download one package using the official Hub client: ```python from huggingface_hub import hf_hub_download path = hf_hub_download( repo_id="TSC-FM/benchmark-datasets", repo_type="dataset", filename="data/UCR/ACSF1.zip", ) ``` These are downloadable archive files, not a unified `datasets.load_dataset()` table. Consult `dataset_index.csv` and `manifest.json` to choose the appropriate original split files and formats. Download the complete repository with the official CLI (the client can reuse already downloaded content): ```bash hf download TSC-FM/benchmark-datasets --repo-type dataset --local-dir tsc-fm_datasets ``` This downloads the available ZIP packages and their metadata. Extract the per-dataset packages to access their original split files. ## Sources, attribution and reuse terms These datasets were originally created and released by third parties. TSC-FM contributes the benchmark selection, overlap resolution and access packaging. Dataset-specific original licenses and attribution requirements continue to apply; this repository does not assert one blanket license covering all underlying datasets. A code repository's software license must not be interpreted as a license for every dataset it links to. - [UCR Time Series Classification Archive](https://www.cs.ucr.edu/~eamonn/time_series_data_2018/) - [Time Series Classification archive: UCR and UEA](https://www.timeseriesclassification.com/) - [Multiverse archive and dataset documentation](https://github.com/aeon-toolkit/multiverse) Please cite the original dataset creators and the relevant archive publications in addition to the TSC-FM Benchmark when using this selection: - Dau et al., *The UCR Time Series Archive*. https://arxiv.org/abs/1810.07758 - Bagnall et al., *The UEA multivariate time series classification archive, 2018*. https://arxiv.org/abs/1811.00075 - *The Multiverse of Time Series Machine Learning: an Archive for Multivariate Time Series Classification*. https://arxiv.org/abs/2603.20352 This is a historical snapshot of the selected benchmark files, not a claim to distribute the latest release of any upstream archive. Known version-specific issues should be checked against the upstream dataset documentation before reuse. **Upstream version note:** current [Multiverse dataset documentation](https://github.com/aeon-toolkit/multiverse/blob/main/docs/datasets.md) flags a target-derived channel in an earlier BenzeneConcentration_disc release and removes that task from the current core. This historical selection retains its recorded benchmark copy. Its relationship to the corrected upstream version requires a separate file/content audit; package integrity checks do not resolve that scientific validity question.