---
license: openmdw-1.1
license_name: per-dataset
license_link: https://huggingface.co/datasets/abel-lab/Simons-Pretrain/blob/main/NOTICE.md
task_categories:
- time-series-forecasting
tags:
- time-series
- forecasting
- pretraining
pretty_name: Simons Pretraining Data
size_categories:
- 10B<n<100B
configs:
- config_name: univariate
  data_files: univariate/*/*.parquet
  default: true
- config_name: covariates
  data_files: covariates/*/*.parquet
---

# Simons Pretraining Data

The real-world training data of the [Simons](https://huggingface.co/collections/abel-lab/simons-1-6ae3d4ce9705938e293c52dd) time series
foundation models.
The synthetic training data is generated by the training code.

[💻 Code](https://github.com/Abel-ai-lab/Simons) · [🤗 Models](https://huggingface.co/collections/abel-lab/simons-1-6ae3d4ce9705938e293c52dd) ·
[📄 Training data overview](https://github.com/Abel-ai-lab/Simons/blob/main/docs/DATA.md) ·
[📝 Release post](https://abel.ai/research/introducing-simons/)

## Statistics

<table>
<tr><td><b>Observations</b></td><td>11.3B</td></tr>
<tr><td><b>Datasets</b></td><td>62 (univariate), 59 (covariate)</td></tr>
<tr><td><b>Series</b></td><td>7.5M (univariate), 6k (covariate)</td></tr>
<tr><td><b>Domains</b></td><td>11</td></tr>
<tr><td><b>Sampling frequencies</b></td><td>Sub-second to monthly</td></tr>
</table>

The models train on this data together with synthetic series.


## Usage

With the [training code](https://github.com/Abel-ai-lab/Simons), which downloads the data on first use
and trains one size end to end, as the released models were trained:

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

With 🤗 Datasets:

```python
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.

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

## Licences

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

## Citation

```bibtex
@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}
}
```