Contamination Map Methodology

Time series / Model

Chronos (T5)

Organisation
Amazon
Family
Chronos · version 1.0
Released
2024-03-13 (Repository README news entry '13 Mar 2024: Chronos paper and inference code released' (https://raw.githubusercontent.com/amazon-science/chronos-forecasting/main/README.md); arXiv v1 submitted 12 Mar 2024 (https://arxiv.org/abs/2403.07815))
Checkpoints
amazon/chronos-t5-tiny, amazon/chronos-t5-mini, amazon/chronos-t5-small, amazon/chronos-t5-base, amazon/chronos-t5-large
Weights
Open · apache-2.0
Links
PaperModel cardCode

Training data disclosure: Full list

Level determined for 100% of benchmark cells.

“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Training corpora

Other declared training data

Declared exclusions

Benchmark cells

GIFT-Eval

Not seen Jena Weather (GIFT-Eval) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen BizITObs - Application · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen BizITObs - Service · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen BizITObs - L2C · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Bitbrains - Fast Storage · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Bitbrains - rnd · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Restaurant (Recruit Restaurant Visitor Forecasting) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen ETT1 (GIFT-Eval) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen ETT2 (GIFT-Eval) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Loop Seattle (LibCity) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen SZ-Taxi (LibCity) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen M_DENSE (LibCity) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Related data seen Solar (LSTNet, 137 Alabama PV plants) · related dataset NREL Solar Power Data for Integration Studies (2006) was used (via Chronos training corpus (real-world datasets))

Evidence (6)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15
“Solar (5 Min.)”
arxiv.org · paper · Appendix B, Table 3, section 'Pretraining-only' · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15
“The uploaded dataset contains the aggregated version of a subset of the original dataset used by Lai et al. (2017).”
zenodo.org · dataset card · retrieved 2026-10-15
“Solar (5 Min., Hourly) contains data about solar power generation in the US in 2006. The original data has 5 minute frequency and was obtained from https://www.nrel.gov/grid/solar-power-data.html”
arxiv.org · paper · Appendix B · retrieved 2026-10-15

Not seen Hierarchical Sales (Mancuso et al., 2020) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen M4 Yearly · the authors state the dataset was excluded from training

Evidence (2)
“M4 (Yearly)”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen M4 Quarterly · the authors state the dataset was excluded from training

Evidence (2)
“M4 (Quarterly)”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Seen, split unclear M4 Monthly · a training split was used; its overlap with this benchmark's test data is not established (via Chronos training corpus (real-world datasets))

Evidence (4)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15
“M4 (Monthly)”
arxiv.org · paper · Appendix B, Table 3, section 'In-domain evaluation' · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Seen, split unclear M4 Weekly · a training split was used; its overlap with this benchmark's test data is not established (via Chronos training corpus (real-world datasets))

Evidence (4)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15
“M4 (Weekly)”
arxiv.org · paper · Appendix B, Table 3, section 'In-domain evaluation' · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Seen, split unclear M4 Daily · a training split was used; its overlap with this benchmark's test data is not established (via Chronos training corpus (real-world datasets))

Evidence (4)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15
“M4 (Daily)”
arxiv.org · paper · Appendix B, Table 3, section 'In-domain evaluation' · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Seen, split unclear M4 Hourly · a training split was used; its overlap with this benchmark's test data is not established (via Chronos training corpus (real-world datasets))

Evidence (4)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15
“M4 (Hourly)”
arxiv.org · paper · Appendix B, Table 3, section 'In-domain evaluation' · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen Hospital · the authors state the dataset was excluded from training

Evidence (2)
“Hospital”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen COVID Deaths · the authors state the dataset was excluded from training

Evidence (2)
“Covid Deaths”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen US Births · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Saugeen River Flow · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Seen, split unclear Temperature Rain · a training split was used; its overlap with this benchmark's test data is not established (via Chronos training corpus (real-world datasets))

Evidence (4)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15
“Temperature-Rain”
arxiv.org · paper · Appendix B, Table 3, section 'In-domain evaluation' · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Seen, split unclear KDD Cup 2018 · a training split was used; its overlap with this benchmark's test data is not established (via Chronos training corpus (real-world datasets))

Evidence (4)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15
“KDD Cup 2018”
arxiv.org · paper · Appendix B, Table 3, section 'In-domain evaluation' · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen Car Parts · the authors state the dataset was excluded from training

Evidence (2)
“Car Parts”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Seen, split unclear ElectricityLoadDiagrams20112014 (UCI) · a training split was used; its overlap with this benchmark's test data is not established (via Chronos training corpus (real-world datasets))

Evidence (4)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15
“Electricity (15 Min.)”
arxiv.org · paper · Appendix B, Table 3, section 'In-domain evaluation' · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

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

Not seen ETTh1 · the authors state the dataset was excluded from training

Evidence (2)
“ETT (Hourly)”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen ETTh2 · the authors state the dataset was excluded from training

Evidence (2)
“ETT (Hourly)”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen ETTm1 · the authors state the dataset was excluded from training

Evidence (2)
“ETT (15 Min.)”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen ETTm2 · the authors state the dataset was excluded from training

Evidence (2)
“ETT (15 Min.)”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Seen, split unclear Electricity (ECL, 321 clients) · a training split was used; its overlap with this benchmark's test data is not established (via Chronos training corpus (real-world datasets))

Evidence (4)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15
“Electricity (Hourly)”
arxiv.org · paper · Appendix B, Table 3, section 'In-domain evaluation' · retrieved 2026-10-15
“pretraining-only data is only used for Chronos training; in-domain evalution data is used for training Chronos models and other task-specific baselines, except for the H observations that are held out for in-domain testing only”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen Traffic (PeMS San Francisco Bay, 862 sensors) · the authors state the dataset was excluded from training

Evidence (2)
“Traffic”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen Weather (Jena, 2020; Autoformer) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Exchange-Rate · the authors state the dataset was excluded from training

Evidence (2)
“Exchange Rate”
arxiv.org · paper · Appendix B, Table 3, section 'Zero-shot evaluation' · retrieved 2026-10-15
“zero-shot evaluation data is not used in training Chronos models, but only for evaluation (final H observations), as well as for training task-specific baselines (excluding the final H observations)”
arxiv.org · paper · Appendix B, Table 3 caption · retrieved 2026-10-15

Not seen ILI (CDC influenza-like illness, national) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Time-series anomaly detection

Not seen TSB-AD-U (TSB-AD univariate) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen TSB-AD-M (TSB-AD multivariate) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen UCR Time Series Anomaly Archive (Hexagon ML/UCR, KDD Cup 2021) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Server Machine Dataset (SMD) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Mars Science Laboratory (MSL) telemetry anomaly dataset · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Soil Moisture Active Passive (SMAP) telemetry anomaly dataset · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Pooled Server Metrics (PSM) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Secure Water Treatment (SWaT) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Yahoo S5 Labeled Anomaly Detection Dataset (Yahoo Webscope S5) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen Numenta Anomaly Benchmark (NAB) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Time-series classification

Not seen UCR Time Series Classification Archive (2018) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Not seen UEA Multivariate Time Series Classification Archive (2018) · absent from the complete list of training data

Evidence (3)
“Overall, we used 28 datasets for training Chronos models, consisting of about 890K univariate time series with approximately 84B observations (tokens) in total.”
arxiv.org · paper · Section 5.1 · retrieved 2026-10-15
“The table is divided into three sections, representing how the datasets were used for Chronos models: in total, 55 datasets where used for experiments, 13 of which for pretraining only, 15 for in-domain evaluation, and 27 for zeroshot evaluation”
arxiv.org · paper · Appendix B · retrieved 2026-10-15
“The training corpus used for training the Chronos models can be loaded via the configs `training_corpus_tsmixup_10m` (10M TSMixup augmentations of real-world data) and `training_corpus_kernel_synth_1m` (1M synthetic time series generated with KernelSynth)”
huggingface.co · dataset card · Section 'Chronos training corpus with TSMixup & KernelSynth' · retrieved 2026-10-15

Notes

Pre-trained on 10M TSMixup augmentations of the 28 real training datasets plus 1M KernelSynth series (ratio 9:1). Benchmark I (15 datasets) is in-domain: trained on all but the last H observations. Benchmark II (27 datasets, incl. ETT, Exchange Rate, Traffic, Weather) was not used for training. Electricity (15 Min./Hourly/Weekly) is in the training data (test window held out). The paper's footnote concedes Benchmark II start times are not guaranteed to be after the training data's last timestamp. Identical copies of the checkpoints exist under the autogluon/ HF org. Other stated exclusions, not tracked here: chronos-benchmark-ii (all); chronos-benchmark-i (test); Australian Electricity (all); CIF 2016 (all); Dominick (all); ERCOT Load (all); FRED-MD (all); M1 (Monthly) (all); M1 (Quarterly) (all); M1 (Yearly) (all); M3 (Monthly) (all); M3 (Quarterly) (all); M3 (Yearly) (all); M5 (all); NN5 (Daily) (all); NN5 (Weekly) (all); Tourism (Monthly) (all); Tourism (Quarterly) (all); Tourism (Yearly) (all); Weather (all).

Sources