Contamination MapTime series / Model
UniTS few-shot anomaly detection checkpoints
- Organisation
- Harvard University (MIMS lab)
- Family
- UniTS
- Released
- 2024-02-29 (GitHub release 'ckpt' published 2024-02-29 per GitHub API (https://api.github.com/repos/mims-harvard/UniTS/releases))
- Checkpoints
- units_finetune_few_shot_anomaly_detection.pth, units_prompt_tuning_few_shot_anomaly_detection.pth
- Weights
- Open · MIT
- Links
- PaperCode
Training data disclosure: Full list
Level determined for 100% of benchmark cells.
- TestTest data seen0
- SeenSeen, split unclear0
- TrainTraining data seen13
- Related data seen6
- NoNot seen30
- ?Unknown0
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
Built on
Everything these base models may have seen is inherited by UniTS few-shot anomaly detection checkpoints.
-
UniTS
· disclosure: Full list
Fine-tuned / prompt-tuned from the self-supervised x32 checkpoint.
Evidence
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
raw.githubusercontent.com · code · scripts/few_shot_anomaly_detection/UniTS_finetune_few_shot_anomaly_detection.sh · retrieved 2026-10-15
Other declared training data
-
Server Machine Dataset (SMD) · split: train of ts-anomaly · fine-tuning
5% of the standard (Anomaly-Transformer/TSLib) training split; one checkpoint jointly fine-tuned on all five datasets.
Evidence
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
arxiv.org · paper · Section 5 · retrieved 2026-10-15 “SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
arxiv.org · paper · Table 11 · retrieved 2026-10-15 “--subsample_pct 0.05”
raw.githubusercontent.com · code · few-shot AD script · retrieved 2026-10-15 “--task_data_config_path data_provider/anomaly_detection.yaml”
raw.githubusercontent.com · code · few-shot AD script · retrieved 2026-10-15 “data = np.load(os.path.join(root_path, "SMD_train.npy"))”
raw.githubusercontent.com · code · SMDSegLoader · retrieved 2026-10-15 -
Mars Science Laboratory (MSL) telemetry anomaly dataset · split: train of ts-anomaly · fine-tuning
5% of the standard (Anomaly-Transformer/TSLib) training split; one checkpoint jointly fine-tuned on all five datasets.
Evidence
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
arxiv.org · paper · Section 5 · retrieved 2026-10-15 “SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
arxiv.org · paper · Table 11 · retrieved 2026-10-15 “--subsample_pct 0.05”
raw.githubusercontent.com · code · few-shot AD script · retrieved 2026-10-15 “--task_data_config_path data_provider/anomaly_detection.yaml”
raw.githubusercontent.com · code · few-shot AD script · retrieved 2026-10-15 -
Soil Moisture Active Passive (SMAP) telemetry anomaly dataset · split: train of ts-anomaly · fine-tuning
5% of the standard (Anomaly-Transformer/TSLib) training split; one checkpoint jointly fine-tuned on all five datasets.
Evidence
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
arxiv.org · paper · Section 5 · retrieved 2026-10-15 “SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
arxiv.org · paper · Table 11 · retrieved 2026-10-15 “--subsample_pct 0.05”
raw.githubusercontent.com · code · few-shot AD script · retrieved 2026-10-15 “--task_data_config_path data_provider/anomaly_detection.yaml”
raw.githubusercontent.com · code · few-shot AD script · retrieved 2026-10-15 -
Secure Water Treatment (SWaT) · split: train of ts-anomaly · fine-tuning
5% of the standard (Anomaly-Transformer/TSLib) training split; one checkpoint jointly fine-tuned on all five datasets.
Evidence
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
arxiv.org · paper · Section 5 · retrieved 2026-10-15 “SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
arxiv.org · paper · Table 11 · retrieved 2026-10-15 “--subsample_pct 0.05”
raw.githubusercontent.com · code · few-shot AD script · retrieved 2026-10-15 “--task_data_config_path data_provider/anomaly_detection.yaml”
raw.githubusercontent.com · code · few-shot AD script · retrieved 2026-10-15 -
Pooled Server Metrics (PSM) · split: train of ts-anomaly · fine-tuning
5% of the standard (Anomaly-Transformer/TSLib) training split; one checkpoint jointly fine-tuned on all five datasets.
Evidence
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
arxiv.org · paper · Section 5 · retrieved 2026-10-15 “SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
arxiv.org · paper · Table 11 · retrieved 2026-10-15 “--subsample_pct 0.05”
raw.githubusercontent.com · code · few-shot AD script · retrieved 2026-10-15 “--task_data_config_path data_provider/anomaly_detection.yaml”
raw.githubusercontent.com · code · few-shot AD script · retrieved 2026-10-15
Benchmark cells
GIFT-Eval
Jena Weather (GIFT-Eval) · via base model UniTS: related dataset Weather (Jena, 2020; Autoformer) was used
Evidence (7)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“WeatherP 96 [110]”
“For subsets of a dataset such as ETTh1, we start by splitting the data into training and testing sets based on distinct time intervals of a long time series sequence, following splits in [112].”
“_, train_loader_list = self._get_data(flag='train')”
“We incorporate Jena Weather5 dataset following Autoformer (Wu et al., 2021).”
“is recorded every 10 minutes for 2020 whole year, which contains 21 meteorological indicators, such as air temperature, humidity, etc.”
“https://www.bgc-jena.mpg.de/wetter/”
Not seen BizITObs - Application · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen BizITObs - Service · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen BizITObs - L2C · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Bitbrains - Fast Storage · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Bitbrains - rnd · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Restaurant (Recruit Restaurant Visitor Forecasting) · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
ETT1 (GIFT-Eval) · via base model UniTS: related dataset ETTh1 was used
Evidence (6)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“ETTh1P 96 [127]”
“For subsets of a dataset such as ETTh1, we start by splitting the data into training and testing sets based on distinct time intervals of a long time series sequence, following splits in [112].”
“_, train_loader_list = self._get_data(flag='train')”
“From Informer (Zhou et al., 2021) we utilize ETT1 and ETT2 datasets, which denote electricity transformer temperature”
“ETTh1&ETTh2 7 17,420”
Not seen ETT2 (GIFT-Eval) · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Loop Seattle (LibCity) · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen SZ-Taxi (LibCity) · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen M_DENSE (LibCity) · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Solar (LSTNet, 137 Alabama PV plants) · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Hierarchical Sales (Mancuso et al., 2020) · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen M4 Yearly · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen M4 Quarterly · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen M4 Monthly · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen M4 Weekly · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen M4 Daily · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen M4 Hourly · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Hospital · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen COVID Deaths · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen US Births · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Saugeen River Flow · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Temperature Rain · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen KDD Cup 2018 · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Car Parts · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
ElectricityLoadDiagrams20112014 (UCI) · via base model UniTS: related dataset Electricity (ECL, 321 clients) was used
Evidence (7)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“ECLP 96 [102]”
“For subsets of a dataset such as ETTh1, we start by splitting the data into training and testing sets based on distinct time intervals of a long time series sequence, following splits in [112].”
“_, train_loader_list = self._get_data(flag='train')”
“The raw dataset is in https://archive.ics.uci.edu/ml/datasets/ElectricityLoadDiagrams20112014.”
“Because the some dimensions are equal to 0. So we eliminate the records in 2011. Final we get data contains electircity consumption of 321 clients from 2012 to 2014.”
“https://archive.ics.uci.edu/ml/datasets/ElectricityLoadDiagrams20112014”
Long-horizon forecasting (ETT, Electricity, Traffic, Weather, Exchange, ILI)
Training data seen ETTh1 · via base model UniTS: only the training split was used, not the benchmark's test data
Evidence (4)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“ETTh1P 96 [127]”
“For subsets of a dataset such as ETTh1, we start by splitting the data into training and testing sets based on distinct time intervals of a long time series sequence, following splits in [112].”
“_, train_loader_list = self._get_data(flag='train')”
Not seen ETTh2 · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
ETTm1 · via base model UniTS: related dataset ETTh1 was used
Evidence (5)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“ETTh1P 96 [127]”
“For subsets of a dataset such as ETTh1, we start by splitting the data into training and testing sets based on distinct time intervals of a long time series sequence, following splits in [112].”
“_, train_loader_list = self._get_data(flag='train')”
“we also provide the hourly-level variants for fast development (marked by *h*), i.e. ETT-small-h1 and ETT-small-h2.”
Not seen ETTm2 · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Training data seen Electricity (ECL, 321 clients) · via base model UniTS: only the training split was used, not the benchmark's test data
Evidence (4)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“ECLP 96 [102]”
“For subsets of a dataset such as ETTh1, we start by splitting the data into training and testing sets based on distinct time intervals of a long time series sequence, following splits in [112].”
“_, train_loader_list = self._get_data(flag='train')”
Training data seen Traffic (PeMS San Francisco Bay, 862 sensors) · via base model UniTS: only the training split was used, not the benchmark's test data
Evidence (4)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“TrafficP 96 [85]”
“For subsets of a dataset such as ETTh1, we start by splitting the data into training and testing sets based on distinct time intervals of a long time series sequence, following splits in [112].”
“_, train_loader_list = self._get_data(flag='train')”
Training data seen Weather (Jena, 2020; Autoformer) · via base model UniTS: only the training split was used, not the benchmark's test data
Evidence (4)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“WeatherP 96 [110]”
“For subsets of a dataset such as ETTh1, we start by splitting the data into training and testing sets based on distinct time intervals of a long time series sequence, following splits in [112].”
“_, train_loader_list = self._get_data(flag='train')”
Training data seen Exchange-Rate · via base model UniTS: only the training split was used, not the benchmark's test data
Evidence (4)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“ExchangeP 192 [53]”
“For subsets of a dataset such as ETTh1, we start by splitting the data into training and testing sets based on distinct time intervals of a long time series sequence, following splits in [112].”
“_, train_loader_list = self._get_data(flag='train')”
Training data seen ILI (CDC influenza-like illness, national) · via base model UniTS: only the training split was used, not the benchmark's test data
Evidence (4)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“ILIP 60 [11]”
“For subsets of a dataset such as ETTh1, we start by splitting the data into training and testing sets based on distinct time intervals of a long time series sequence, following splits in [112].”
“_, train_loader_list = self._get_data(flag='train')”
Time-series anomaly detection
TSB-AD-U (TSB-AD univariate) · in fine-tuning data: related dataset Server Machine Dataset (SMD) was used
Evidence (8)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“--subsample_pct 0.05”
“--task_data_config_path data_provider/anomaly_detection.yaml”
“data = np.load(os.path.join(root_path, "SMD_train.npy"))”
“we implement a transformation strategy that converts multivariate time series into univariate formats by treating each channel as an independent series.”
“SMD (U) [97]”
“178_SMD_id_1_Facility_tr_6000_1st_10609.csv”
TSB-AD-M (TSB-AD multivariate) · in fine-tuning data: related dataset Server Machine Dataset (SMD) was used
Evidence (7)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“--subsample_pct 0.05”
“--task_data_config_path data_provider/anomaly_detection.yaml”
“data = np.load(os.path.join(root_path, "SMD_train.npy"))”
“SMD [97] is a 5-week-long dataset collected from a large Internet company, which contains 3 groups of entities from 28 different machines.”
“057_SMD_id_1_Facility_tr_4529_1st_4629.csv”
Not seen UCR Time Series Anomaly Archive (Hexagon ML/UCR, KDD Cup 2021) · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Training data seen Server Machine Dataset (SMD) · in fine-tuning data: only the training split was used, not the benchmark's test data
Evidence (5)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“--subsample_pct 0.05”
“--task_data_config_path data_provider/anomaly_detection.yaml”
“data = np.load(os.path.join(root_path, "SMD_train.npy"))”
Training data seen Mars Science Laboratory (MSL) telemetry anomaly dataset · in fine-tuning data: only the training split was used, not the benchmark's test data
Evidence (4)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“--subsample_pct 0.05”
“--task_data_config_path data_provider/anomaly_detection.yaml”
Training data seen Soil Moisture Active Passive (SMAP) telemetry anomaly dataset · in fine-tuning data: only the training split was used, not the benchmark's test data
Evidence (4)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“--subsample_pct 0.05”
“--task_data_config_path data_provider/anomaly_detection.yaml”
Training data seen Pooled Server Metrics (PSM) · in fine-tuning data: only the training split was used, not the benchmark's test data
Evidence (4)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“--subsample_pct 0.05”
“--task_data_config_path data_provider/anomaly_detection.yaml”
Training data seen Secure Water Treatment (SWaT) · in fine-tuning data: only the training split was used, not the benchmark's test data
Evidence (4)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“--subsample_pct 0.05”
“--task_data_config_path data_provider/anomaly_detection.yaml”
Not seen Yahoo S5 Labeled Anomaly Detection Dataset (Yahoo Webscope S5) · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Not seen Numenta Anomaly Benchmark (NAB) · absent from the complete list of training data and from its base models' data
Evidence (6)
“Setup: Few-shot anomaly detection. The pre-trained models have been fine-tuned using 5% of five training datasets as listed in Table 10.”
“SMD [96] MSL [43] SMAP [43] SWaT [77] PSM [1]”
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].”
“Table 7: Multi-task datasets for classification and forecasting.”
“For transfer learning on new tasks and datasets, we load the model weights pre-trained on 38 datasets”
Time-series classification
Training data seen UCR Time Series Classification Archive (2018) · via base model UniTS: only the training split was used, not the benchmark's test data
Evidence (4)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“SharePriceIncrease [79] JapaneseVowels [52] SpokenArabicDigits [5] Heartbeat [61] ECG5000 [35] NonInvasiveFetalECGThorax1 [95] Blink [19] FaceDetection [40] SelfRegulationSCP2 [7] ElectricDevices [60] Trace [93] FordB [23] MotionSenseHAR [75] EMOPain [27] UWaveGestureLibrary [63] Chinatown [23] MelbournePedestrian [23] PEMS-SF [20]”
“Datasets were aggregated from the Monash Forecasting Repository [33], Time Series Classification Website [79], and Time Series Library [112]. The combined training set consists of over 35 million timesteps and over 6,000 variables.”
“_, train_loader_list = self._get_data(flag='train')”
Training data seen UEA Multivariate Time Series Classification Archive (2018) · via base model UniTS: only the training split was used, not the benchmark's test data
Evidence (4)
“ckpt_path=newcheckpoints/units_x32_pretrain_checkpoint.pth”
“SharePriceIncrease [79] JapaneseVowels [52] SpokenArabicDigits [5] Heartbeat [61] ECG5000 [35] NonInvasiveFetalECGThorax1 [95] Blink [19] FaceDetection [40] SelfRegulationSCP2 [7] ElectricDevices [60] Trace [93] FordB [23] MotionSenseHAR [75] EMOPain [27] UWaveGestureLibrary [63] Chinatown [23] MelbournePedestrian [23] PEMS-SF [20]”
“Datasets were aggregated from the Monash Forecasting Repository [33], Time Series Classification Website [79], and Time Series Library [112]. The combined training set consists of over 35 million timesteps and over 6,000 variables.”
“_, train_loader_list = self._get_data(flag='train')”
Notes
In-domain few-shot (not zero-shot) anomaly checkpoints: trained on 5% of the SMD/MSL/SMAP/SWaT/PSM training splits and evaluated on their test splits. The paper text says 'Table 10' but the anomaly datasets are in Table 11.
Sources
- UniTS (arXiv v3) · paper · retrieved 2026-10-15
- UniTS GitHub releases (checkpoint assets) · code · retrieved 2026-10-15
- UniTS anomaly detection data config · code · retrieved 2026-10-15