abel-lab/Simons-Pretrain

Simons Pretraining Data

The real-world training data of the Simons time series foundation models. The synthetic training data is generated by the training code.

💻 Code · 🤗 Models · 📄 Training data overview · 📝 Release post

Statistics

Observations	11.3B
Datasets	62 (univariate), 59 (covariate)
Series	7.5M (univariate), 6k (covariate)
Domains	11
Sampling frequencies	Sub-second to monthly

The models train on this data together with synthetic series.

Compared with other models, such as Toto 2.0, TimesFM-3 and Timer-S1, Simons is efficient with data and depends little on real-world data. Simons-297M trains on 19B real observations, 11.3B of them released here, and 800B synthetic ones that the training code generates. Toto 2.0 trains on 2.1T real and 2.9T synthetic data points; TimesFM-3 and Timer-S1 state corpora of 1T or more.

Usage

With the training code, which downloads the data on first use and trains one size end to end, as the released models were trained:

scripts/train.sh 25m abel-lab/Simons-Pretrain abel-lab/Simons-Pretrain   # 25m, 60m, 178m, 297m

With 🤗 Datasets:

from datasets import load_dataset

univariate = load_dataset("abel-lab/Simons-Pretrain", "univariate", split="train")
covariates = load_dataset("abel-lab/Simons-Pretrain", "covariates", split="train")

Layout

One folder per dataset:

• univariate/<dataset>/train-*.parquet: item_id, source, freq, domain, subdomain, target (float32 [variate][time], NaN = missing). Long series are cut into segments <series>@<offset>, which the long-context stages stitch back together.
• covariates/<dataset>/train-*.parquet: item_id, source, freq, target, covariates ([covariate][time], aligned with target), covariate_names.

manifest.json lists every dataset with its source, frequency and license, and statistics/ breaks the training data down by domain and frequency.

Series were cleaned (invalid values masked, duplicates removed) and given neutral ids.

Licences

Each dataset retains its original license. NOTICE.md lists each dataset's source, license, and required attributions, with links to the license texts.

Citation

@misc{simons2026,
  title  = {Simons: Building Open Time Series Foundation Models through Auto-Research},
  author = {{Abel AI Lab}},
  year   = {2026},
  url    = {https://github.com/Abel-ai-lab/Simons}
}