Datasets¶

Every built dataset in the registry, generated from the dataset.yaml cards in timenet-connectors. Search and filter below, expand a row for details, or read datasets.json for programmatic use.

	Dataset	Name	Categories	License
	`bash-lab/sensorcaps`	SensorCaps	activitymotion	OTHER
35,960 SensorCaps v2 wearable IMU captioning examples from five activity datasets, with accelerometer and gyroscope readings and activity context, plus a magnetometer channel in PAMAP2. PAMAP2 signal names correct upstream channel-label errors. Caption feature summaries and narrations were generated by GPT-4o-mini for the LLaSA project and retain the original channel assumptions. Published PAMAP2 values also retain upstream missing-data processing consistent with zero filling before resampling; original gaps are not marked in the release. Version 1.0.0 Tags captioningimuaccelerometergyroscopemagnetometergenerated-supervisionhuggingface Source https://huggingface.co/datasets/BASH-Lab/SensorCaps `from timenet.client import TimeNet with TimeNet().open("bash-lab/sensorcaps") as reader: dataset = reader.read() dataset.describe()`
	`chengsenwang/tsqa`	TSQA	general	Apache-2.0
Time-series question answering: a series plus a question/answer per record. Generated 80/10/10 train/validation/test partitions using ten StratifiedGroupKFold folds with scikit-learn 1.9.1 and seed 42, stratified by question category and grouped by identical parsed series. These are TimeNet-generated partitions, not official source splits; achieved sizes may differ to keep groups intact. Version 1.2.0 Tags qatime-serieshuggingface Source https://huggingface.co/datasets/ChengsenWang/TSQA `from timenet.client import TimeNet with TimeNet().open("chengsenwang/tsqa") as reader: dataset = reader.read() dataset.describe()`
	`physionet/ecg-qa-cot`	ECG-QA CoT	cardiologyhealth	CC-BY-4.0
12-lead PTB-XL ECGs paired with clinical questions, short answers, and chain-of-thought rationales for reasoning fine-tuning. Each question carries a clinical context built from PTB-XL: the patient's age and sex, the recording device, and the signal-quality issues, extra beats, and pacemaker that PTB-XL records for the recording. Version 1.2.1 Tags ecgqacotphysionet Source https://github.com/StanfordBDHG/OpenTSLM `from timenet.client import TimeNet with TimeNet().open("physionet/ecg-qa-cot") as reader: dataset = reader.read() dataset.describe()`
	`physionet/ptb-xl`	PTB-XL	cardiologyhealth	CC-BY-4.0
Original 500 Hz clinical ECG recordings and annotations, without downstream tasks. Version 1.0.0 Tags ecgphysionetclinicalbase-dataset Source https://physionet.org/content/ptb-xl/1.0.3/ `from timenet.client import TimeNet with TimeNet().open("physionet/ptb-xl") as reader: dataset = reader.read() dataset.describe()`
	`physionet/sleep-edfx`	Sleep-EDF Expanded	sleephealth	ODC-By-1.0
Whole-night polysomnograms from two studies, one recorded at home and one in hospital, with sleep stages a technician scored in 30 s epochs. Version 1.0.0 Tags eegeogemgpolysomnographysleep-stagingphysionet Source https://physionet.org/content/sleep-edfx/1.0.0/ `from timenet.client import TimeNet with TimeNet().open("physionet/sleep-edfx") as reader: dataset = reader.read() dataset.describe()`
	`timenet/hello-world`	Hello World	general	CC-BY-4.0
A synthetic demo dataset exercising every TimeF feature. Version 1.0.0 Tags demosynthetic Source n/a `from timenet.client import TimeNet with TimeNet().open("timenet/hello-world") as reader: dataset = reader.read() dataset.describe()`
	`timenet/test-mean`	Test Mean	general	CC-BY-4.0
A synthetic binary-classification demo: each record is one noisy signal labeled by whether its mean is above or below zero. Balanced and deterministic; backs the end-to-end training example. Version 1.0.0 Tags demosyntheticclassification Source n/a `from timenet.client import TimeNet with TimeNet().open("timenet/test-mean") as reader: dataset = reader.read() dataset.describe()`
	`tsaqa/tsaqa-benchmark`	TSAQA	energytransporthealthcardiologyeconomicsfinanceenvironmentobservabilityindustrialmotiongeneral	MIT
Time series analysis question answering: 210,000 closed questions about series from 13 domains (147,000 train, 21,000 validation, 42,000 test), across six tasks: anomaly detection, classification, characterization, comparison, data transformation and temporal relationship. True-or-false questions are classification tasks; multiple-choice and puzzle questions are answer tasks whose target is the published choice letters. Candidate series that the authors print inside a question are Records of their own. Version 1.0.0 Tags qaclassificationanomaly-detectiongenerated-supervisionhuggingface Source https://huggingface.co/datasets/TSAQA/TSAQA-Benchmark `from timenet.client import TimeNet with TimeNet().open("tsaqa/tsaqa-benchmark") as reader: dataset = reader.read() dataset.describe()`