Contamination Map Methodology

Time series / Model

Mantis (Mantis-8M)

Organisation
Huawei Noah's Ark Lab (Paris)
Family
Mantis · version 1 (Mantis-8M)
Released
2025-02-21 (arXiv v1 submission date, https://arxiv.org/abs/2502.15637v1 (HF repo paris-noah/Mantis-8M was created 2024-10-04 per HF API, presumably private until the paper))
Checkpoints
paris-noah/Mantis-8M
Weights
Open · apache-2.0
Links
PaperModel cardCode

Training data disclosure: Partial list

Level determined for 4% of benchmark cells.

“As a pre-training dataset, we use a union of various public datasets including UCR (Dau et al., 2019), UEA (Bagnall et al., 2018, all except EigenWorms and”
arxiv.org · paper · Section 2 (Pre-training) · retrieved 2026-10-15
“InsectWingbeat), ECG (Clifford et al., 2017), EMG (Goldberger et al., 2000), Epilepsy (Andrzejak et al., 2001), FD-A and FD-B (Lessmeier et al., 2016), Gesture (Liu et al., 2009), HAR (Anguita et al., 2013), SleepEEG (Kemp et al., 2000).”
arxiv.org · paper · Section 2 (Pre-training) · retrieved 2026-10-15
“The total pre-training consists of 7 million time series examples.”
arxiv.org · paper · Section 2 · retrieved 2026-10-15
“Unlike the original Mantis pre-training corpus (1.89 million samples), which partially overlapped with UCR and UEA training sets”
arxiv.org · paper · Section 3 (Pre-training with Synthetic Data) · retrieved 2026-10-15

Other declared training data

Declared exclusions

Benchmark cells

GIFT-Eval

Unknown Jena Weather (GIFT-Eval) · training data not disclosed in enough detail

Unknown BizITObs - Application · training data not disclosed in enough detail

Unknown BizITObs - Service · training data not disclosed in enough detail

Unknown BizITObs - L2C · training data not disclosed in enough detail

Unknown Bitbrains - Fast Storage · training data not disclosed in enough detail

Unknown Bitbrains - rnd · training data not disclosed in enough detail

Unknown Restaurant (Recruit Restaurant Visitor Forecasting) · training data not disclosed in enough detail

Unknown ETT1 (GIFT-Eval) · training data not disclosed in enough detail

Unknown ETT2 (GIFT-Eval) · training data not disclosed in enough detail

Unknown Loop Seattle (LibCity) · training data not disclosed in enough detail

Unknown SZ-Taxi (LibCity) · training data not disclosed in enough detail

Unknown M_DENSE (LibCity) · training data not disclosed in enough detail

Unknown Solar (LSTNet, 137 Alabama PV plants) · training data not disclosed in enough detail

Unknown Hierarchical Sales (Mancuso et al., 2020) · training data not disclosed in enough detail

Unknown M4 Yearly · training data not disclosed in enough detail

Unknown M4 Quarterly · training data not disclosed in enough detail

Unknown M4 Monthly · training data not disclosed in enough detail

Unknown M4 Weekly · training data not disclosed in enough detail

Unknown M4 Daily · training data not disclosed in enough detail

Unknown M4 Hourly · training data not disclosed in enough detail

Unknown Hospital · training data not disclosed in enough detail

Unknown COVID Deaths · training data not disclosed in enough detail

Unknown US Births · training data not disclosed in enough detail

Unknown Saugeen River Flow · training data not disclosed in enough detail

Unknown Temperature Rain · training data not disclosed in enough detail

Unknown KDD Cup 2018 · training data not disclosed in enough detail

Unknown Car Parts · training data not disclosed in enough detail

Unknown ElectricityLoadDiagrams20112014 (UCI) · training data not disclosed in enough detail

Long-horizon forecasting (ETT, Electricity, Traffic, Weather, Exchange, ILI)

Unknown ETTh1 · training data not disclosed in enough detail

Unknown ETTh2 · training data not disclosed in enough detail

Unknown ETTm1 · training data not disclosed in enough detail

Unknown ETTm2 · training data not disclosed in enough detail

Unknown Electricity (ECL, 321 clients) · training data not disclosed in enough detail

Unknown Traffic (PeMS San Francisco Bay, 862 sensors) · training data not disclosed in enough detail

Unknown Weather (Jena, 2020; Autoformer) · training data not disclosed in enough detail

Unknown Exchange-Rate · training data not disclosed in enough detail

Unknown ILI (CDC influenza-like illness, national) · training data not disclosed in enough detail

Time-series anomaly detection

Unknown TSB-AD-U (TSB-AD univariate) · training data not disclosed in enough detail

Unknown TSB-AD-M (TSB-AD multivariate) · training data not disclosed in enough detail

Unknown UCR Time Series Anomaly Archive (Hexagon ML/UCR, KDD Cup 2021) · training data not disclosed in enough detail

Unknown Server Machine Dataset (SMD) · training data not disclosed in enough detail

Unknown Mars Science Laboratory (MSL) telemetry anomaly dataset · training data not disclosed in enough detail

Unknown Soil Moisture Active Passive (SMAP) telemetry anomaly dataset · training data not disclosed in enough detail

Unknown Pooled Server Metrics (PSM) · training data not disclosed in enough detail

Unknown Secure Water Treatment (SWaT) · training data not disclosed in enough detail

Unknown Yahoo S5 Labeled Anomaly Detection Dataset (Yahoo Webscope S5) · training data not disclosed in enough detail

Unknown Numenta Anomaly Benchmark (NAB) · training data not disclosed in enough detail

Time-series classification

Training data seen UCR Time Series Classification Archive (2018) · only the training split was used, not the benchmark's test data; the test portion was explicitly excluded

Evidence (4)
“As a pre-training dataset, we use a union of various public datasets including UCR (Dau et al., 2019), UEA (Bagnall et al., 2018, all except EigenWorms and”
arxiv.org · paper · Section 2 (Pre-training) · retrieved 2026-10-15
“We ensure that test sets used for evaluation in Section 3 are not part of the pre-training dataset.”
arxiv.org · paper · Section 2 (Pre-training) · retrieved 2026-10-15
“Unlike the original Mantis pre-training corpus (1.89 million samples), which partially overlapped with UCR and UEA training sets”
arxiv.org · paper · Section 3 (Pre-training with Synthetic Data) · retrieved 2026-10-15
“We ensure that test sets used for evaluation in Section 3 are not part of the pre-training dataset.”
arxiv.org · paper · Section 2 (Pre-training) · retrieved 2026-10-15

Training data seen UEA Multivariate Time Series Classification Archive (2018) · only the training split was used, not the benchmark's test data; the test portion was explicitly excluded

Evidence (5)
“As a pre-training dataset, we use a union of various public datasets including UCR (Dau et al., 2019), UEA (Bagnall et al., 2018, all except EigenWorms and”
arxiv.org · paper · Section 2 (Pre-training) · retrieved 2026-10-15
“InsectWingbeat), ECG (Clifford et al., 2017), EMG (Goldberger et al., 2000), Epilepsy (Andrzejak et al., 2001), FD-A and FD-B (Lessmeier et al., 2016), Gesture (Liu et al., 2009), HAR (Anguita et al., 2013), SleepEEG (Kemp et al., 2000).”
arxiv.org · paper · Section 2 (Pre-training) · retrieved 2026-10-15
“We ensure that test sets used for evaluation in Section 3 are not part of the pre-training dataset.”
arxiv.org · paper · Section 2 (Pre-training) · retrieved 2026-10-15
“Unlike the original Mantis pre-training corpus (1.89 million samples), which partially overlapped with UCR and UEA training sets”
arxiv.org · paper · Section 3 (Pre-training with Synthetic Data) · retrieved 2026-10-15
“We ensure that test sets used for evaluation in Section 3 are not part of the pre-training dataset.”
arxiv.org · paper · Section 2 (Pre-training) · retrieved 2026-10-15

Notes

8M-parameter contrastively pre-trained classification FM. Data list is introduced with 'including', and sizes disagree (7M examples in the v1 paper vs 1.89M samples in the MantisV2 report), so partial_list. The developers later state this corpus caused leakage with evaluation datasets (UCR/UEA training sets). arXiv v2 of 2502.15637 (ICML 2026) re-describes 'Mantis' as pre-trained exclusively on CauKer synthetic data; that corresponds to the MantisPlus/MantisV2 checkpoints, not Mantis-8M.

Sources