TSB-SEG datasets

The datasets of TSB-SEG, the time-series segmentation benchmark of the NeurIPS 2026 Evaluations & Datasets track: exactly the series of the paper (version 1.0), as the benchmark loader returns them, converted to Parquet, one configuration per dataset. Change point detection (CPD) uses the boundaries of the state labels, state detection (SD) the labels themselves. Each dataset keeps the license of its source.

Configuration	Dataset	Series	Length	Channels	Change points per series	States per series	License
mocap	MoCap	9	4,579 to 10,617	4	5 to 10	3 to 9	CMU terms
has	HAS	250	340 to 41,465	6, 9	0 to 14	1 to 12	CC BY-NC-SA 4.0
actrectut	ActRecTut	2	31,392 to 32,577	10	42	6	GPL-3.0-or-later
pump	PUMP	120	20,365 to 40,810	9	50 to 78	41 to 43	CC BY 4.0
skab	SKAB	34	745 to 1,327	8	1 to 2	2	GPL-3.0
tssb	TSSB	75	240 to 20,700	1	0 to 8	1 to 7	BSD-3-Clause
pamap2	PAMAP2	9	8,477 to 447,000	9	2 to 27	2 to 13	CC BY 4.0

USC-HAD (70 series) and UTSA (32 series), the two other datasets of the benchmark, have no explicit license and are not redistributed here: python scripts/download_data.py usc-had utsa in the TSB-SEG repository downloads them from their source.

Fields

One row per series, in data/<configuration>.parquet:

Field	Type	Content
series	int32	1-based index of the series in the benchmark runs
series_id	string	name of the series in its source
source	string	JSON parameters that select the series in the source layout
n_timepoints, n_channels	int	shape of the series
labels	list of int	state label of every time point
channel_0 ...	list of float64	values of each channel (null beyond n_channels)

Loading

import numpy as np
import pandas as pd

df = pd.read_parquet("hf://datasets/fchavelli/tsb-seg/data/mocap.parquet")
row = df.iloc[0]
X = np.column_stack([row[f"channel_{k}"] for k in range(row["n_channels"])])
y = row["labels"]
change_points = np.flatnonzero(np.diff(y)) + 1

hf:// paths need huggingface_hub; datasets.load_dataset("fchavelli/tsb-seg", "mocap", split="test") works too. In the TSB-SEG repository, python scripts/download_data.py places every file where the benchmark loader reads it.

Datasets

MoCap (mocap)

Nine motion-capture recordings of subject 86 of the CMU Graphics Lab Motion Capture Database, segmented into actions (walking, running, jumping, ...), with the four channels and the state labels released by Time2State.

• Channels: The 4 channels released by Time2State for these trials.
• Source: http://mocap.cs.cmu.edu (subject 86); labels of Time2State (https://github.com/Lab-ANT/Time2State).
• Limitations: A single subject. Barbič et al. (2004) annotate 14 trials of this subject; the benchmark uses the 9 trials released by Time2State, whose labels differ from Barbič's on two of them.
• License: CMU Graphics Lab Motion Capture Database terms: free for research and for use in commercial products; the data may not be resold, even converted. Please acknowledge the database (see the citation).

HAS (has)

250 smartphone recordings of the Human Activity Segmentation Challenge (ECML/PKDD 2023): participants performing sequences of daily activities, indoors or outdoors.

• Channels: x, y, z of the accelerometer, gyroscope and magnetometer for indoor series (9 channels); accelerometer and magnetometer for outdoor series (6 channels: channels 6-8 are null).
• Source: https://github.com/patrickzib/human_activity_segmentation_challenge (via the ClaP repository); location data not included.
• Limitations: 167 of the 250 series have no recurring state. Labels follow the activity protocol; boundaries between activities are gradual.
• License: CC BY-NC-SA 4.0, as the source (non-commercial, share alike).

ActRecTut (actrectut)

Two recordings (subject 1 and 2, walking scenario) of the activity recognition tutorial of Bulling et al., with body-worn inertial sensors, segmented into 6 activities.

• Channels: Columns 0 to 9 of the source data matrix (the selection of Time2State).
• Source: https://github.com/andreas-bulling/ActRecTut (via Time2State).
• Limitations: Two series only; the gesture recordings of the source are not part of the benchmark. The 10 columns are the first 10 of 24, as in Time2State.
• License: GPL-3.0-or-later, as the source.

PUMP (pump)

120 End-of-Line test cycles of hydraulic pumps measured by 9 sensors, labelled with the operating state of the test program (up to 43 states per cycle).

• Channels: Sensors 1 to 9 of the source files (z-normalised per cycle at the source).
• Source: https://github.com/boschresearch/Hydraulic-EoL-Testing (direct control pumps, versions 35, 36 and 38; data of PrecTime).
• Limitations: 23% of the change points border a segment shorter than 0.5 s, and transition states recur; boundaries come from the test program.
• License: CC BY 4.0, as the source.

SKAB (skab)

34 recordings of the Skoltech Anomaly Benchmark testbed (water circulation system), each with one anomalous phase: normal, anomaly, normal.

• Channels: The 8 sensors of the testbed (accelerometers, current, pressure, temperature, thermocouple, voltage, flow rate).
• Source: https://github.com/waico/SKAB (via the ClaP repository).
• Limitations: One anomaly per series, labelled normal, anomaly, normal: 1 or 2 change points and 2 states.
• License: GPL-3.0, as the source.

TSSB (tssb)

The 75 univariate series of the Time Series Segmentation Benchmark, built by concatenating series of different classes of UCR/UEA classification datasets.

• Channels: Univariate.
• Source: https://github.com/ermshaua/time-series-segmentation-benchmark.
• Limitations: Segments are concatenated classes of classification datasets: the boundaries are exact but abrupt; 65 of the 75 series have no recurring state.
• License: BSD 3-Clause, as the source.

PAMAP2 (pamap2)

The 9 protocol recordings of the PAMAP2 physical activity dataset (up to 447,000 time points), used by TSB-SEG as a long-series scalability case study.

• Channels: x, y, z of the hand, chest and ankle accelerometers (the 16 g ones), with the missing values of the source.
• Source: https://doi.org/10.24432/C5NW2H.
• Limitations: Label 0 marks transient activities between protocol activities; the accelerometers have missing values (NaN), kept as in the source.
• License: CC BY 4.0, as the source (UCI, DOI 10.24432/C5NW2H).

Conversion

The series were read by the TSB-SEG data loader (src/tsbseg/data_loading/loaders.py), written by benchmark/datasets/to_parquet.py, and read back to check that every value, missing value, label and data type equals the source. Labels are those of the original authors, without re-annotation; see the appendix of the TSB-SEG paper for their provenance and uncertainty.

Citation

Please cite TSB-SEG:

@inproceedings{chavelli2026tsbseg,
  title     = {{TSB-SEG}: A Systematic Time-Series Segmentation Benchmark},
  author    = {Chavelli, F{\'e}lix and Ermshaus, Arik and Yang, Fan and
               Sch{\"a}fer, Patrick and Paparrizos, John and Boniol, Paul},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS),
               Evaluations and Datasets Track},
  year      = {2026}
}

and the original datasets you use:

@misc{CMU2003MoCap,
  title  = {The {CMU} Graphics Lab Motion Capture Database},
  author = {{Carnegie Mellon University}},
  year   = {2003},
  note   = {http://mocap.cs.cmu.edu}
}
@article{Wang2023Time2State,
  title   = {Time2State: An Unsupervised Framework for Inferring the Latent
             States in Time Series Data},
  author  = {Wang, Chengyu and Wu, Kui and Zhou, Tongqing and Cai, Zhiping},
  journal = {Proceedings of the ACM on Management of Data},
  volume  = {1},
  year    = {2023}
}
@inproceedings{Ermshaus2023Human,
  title     = {Human Activity Segmentation Challenge @ {ECML/PKDD}'23},
  author    = {Ermshaus, Arik and Sch{\"a}fer, Patrick and Bagnall, Anthony and
               Guyet, Thomas and Ifrim, Georgiana and Lemaire, Vincent and
               Leser, Ulf and Leverger, Colin and Malinowski, Simon},
  booktitle = {8th Workshop on Advanced Analytics and Learning on Temporal Data},
  year      = {2023}
}
@article{Bulling2014ActRecTut,
  title   = {A Tutorial on Human Activity Recognition Using Body-worn Inertial
             Sensors},
  author  = {Bulling, Andreas and Blanke, Ulf and Schiele, Bernt},
  journal = {ACM Computing Surveys},
  volume  = {46},
  number  = {3},
  pages   = {33:1--33:33},
  year    = {2014}
}
@article{Gaugel2023PrecTime,
  title   = {PrecTime: A deep learning architecture for precise time series
             segmentation in industrial manufacturing operations},
  author  = {Gaugel, Stefan and Reichert, Manfred},
  journal = {Engineering Applications of Artificial Intelligence},
  volume  = {122},
  pages   = {106078},
  year    = {2023},
  doi     = {10.1016/j.engappai.2023.106078}
}
@misc{Katser2020SKAB,
  title     = {Skoltech Anomaly Benchmark ({SKAB})},
  author    = {Katser, Iurii D. and Kozitsin, Vyacheslav O.},
  publisher = {Kaggle},
  year      = {2020},
  doi       = {10.34740/KAGGLE/DSV/1693952}
}
@article{Ermshaus2023ClaSP,
  title   = {ClaSP: parameter-free time series segmentation},
  author  = {Ermshaus, Arik and Sch{\"a}fer, Patrick and Leser, Ulf},
  journal = {Data Mining and Knowledge Discovery},
  volume  = {37},
  pages   = {1262--1300},
  year    = {2023}
}
@inproceedings{Reiss2012Introducing,
  title     = {Introducing a New Benchmarked Dataset for Activity Monitoring},
  author    = {Reiss, Attila and Stricker, Didier},
  booktitle = {16th International Symposium on Wearable Computers},
  pages     = {108--109},
  year      = {2012}
}