r3-moto/pre-training-data

MoTo pre-training data

The pre-training corpus of the MoTo models and baselines: English educational web text from FineWeb-Edu and 20 languages from FineWeb2-HQ, stored as JSON Lines in one folder per source.

Part of the MoTo release: paper · code · models.

Contents

21 folders, 2.1 TB of JSON Lines in total.

Folder	Language	Source	Train chunks	Train size	Validation size	Mixture weight
fw_edu	English	HuggingFaceFW/fineweb-edu (sample-100BT)	8	392.0 GB	417.4 MB	40.0%
arb_Arab	Arabic	epfml/FineWeb2-HQ (arb_Arab)	8	32.9 GB	967.7 MB	2.3%
ces_Latn	Czech	epfml/FineWeb2-HQ (ces_Latn)	8	29.8 GB	808.1 MB	2.5%
cmn_Hani	Chinese	epfml/FineWeb2-HQ (cmn_Hani)	8	230.4 GB	676.6 MB	4.5%
dan_Latn	Danish	epfml/FineWeb2-HQ (dan_Latn)	8	17.7 GB	712.6 MB	2.2%
deu_Latn	German	epfml/FineWeb2-HQ (deu_Latn)	8	186.9 GB	696.7 MB	4.2%
ell_Grek	Greek	epfml/FineWeb2-HQ (ell_Grek)	8	32.1 GB	1.2 GB	2.3%
fas_Arab	Persian	epfml/FineWeb2-HQ (fas_Arab)	8	26.2 GB	832.0 MB	2.1%
fra_Latn	French	epfml/FineWeb2-HQ (fra_Latn)	8	141.1 GB	708.9 MB	3.8%
hun_Latn	Hungarian	epfml/FineWeb2-HQ (hun_Latn)	8	22.9 GB	827.2 MB	2.5%
ind_Latn	Indonesian	epfml/FineWeb2-HQ (ind_Latn)	8	40.6 GB	734.6 MB	2.7%
ita_Latn	Italian	epfml/FineWeb2-HQ (ita_Latn)	8	78.4 GB	596.8 MB	3.2%
jpn_Jpan	Japanese	epfml/FineWeb2-HQ (jpn_Jpan)	8	120.9 GB	563.1 MB	3.6%
nld_Latn	Dutch	epfml/FineWeb2-HQ (nld_Latn)	8	45.7 GB	566.9 MB	2.8%
pol_Latn	Polish	epfml/FineWeb2-HQ (pol_Latn)	8	47.1 GB	564.1 MB	2.9%
por_Latn	Portuguese	epfml/FineWeb2-HQ (por_Latn)	8	64.8 GB	567.4 MB	3.0%
rus_Cyrl	Russian	epfml/FineWeb2-HQ (rus_Cyrl)	8	392.0 GB	733.7 MB	5.0%
spa_Latn	Spanish	epfml/FineWeb2-HQ (spa_Latn)	8	155.0 GB	622.0 MB	3.9%
swe_Latn	Swedish	epfml/FineWeb2-HQ (swe_Latn)	8	18.6 GB	332.0 MB	2.2%
tur_Latn	Turkish	epfml/FineWeb2-HQ (tur_Latn)	8	29.4 GB	276.1 MB	2.4%
vie_Latn	Vietnamese	epfml/FineWeb2-HQ (vie_Latn)	8	19.1 GB	386.4 MB	2.0%

Mixture weight is the probability with which the training loader samples each folder (data.sources in the training config): 40% English FineWeb-Edu, and the remaining 60% split across the 20 FineWeb2-HQ languages with UniMax sampling. The models were trained on about 104.86B tokens drawn with these weights, so they saw only part of this data.

Format

Each folder holds training chunks named <prefix>.chunk.NN.jsonl and one validation file <prefix>.val.jsonl, the layout read by the lingua data loader, which trains on the chunks only. Every line is one JSON document with its text in the text field.

Usage

from datasets import load_dataset

# one language, streamed (configs are named after the folders)
val = load_dataset("r3-moto/pre-training-data", "tur_Latn", split="validation", streaming=True)
print(next(iter(val))["text"][:200])

To train with lingua, download the folders you need and point data.root_dir at the download:

hf download r3-moto/pre-training-data --repo-type dataset --include "fw_edu/*" "tur_Latn/*" --local-dir data/

Sources and license

• FineWeb-Edu, subset sample-100BT: ODC-By 1.0
• FineWeb2-HQ: ODC-By 1.0

This collection is redistributed under the same license, ODC-By 1.0. As with the source datasets, use is also subject to the Common Crawl Terms of Use.

Citation

If you use this data, please cite MoTo and the source datasets.

@inproceedings{altintas2026moto,
  title={{MoTo}: Mixture of Tokenizers Towards Fair Multilingual Language Modeling},
  author={G{\"u}l Sena Alt{\i}nta{\c{s}} and Colin Raffel},
  booktitle={The Fortieth Annual Conference on Neural Information Processing Systems},
  year={2026},
  url={https://openreview.net/forum?id=avjeUhPuC6}
}

@article{messmer2025multilingdatacomp,
  title={Enhancing Multilingual LLM Pretraining with Model-Based Data Selection},
  author={Bettina Messmer and Vinko Sabolčec and Martin Jaggi},
  journal={arXiv},
  year={2025},
  url={https://arxiv.org/abs/2502.10361},
}

@misc{lozhkov2024fineweb-edu,
  author       = { Lozhkov, Anton and Ben Allal, Loubna and von Werra, Leandro and Wolf, Thomas },
  title        = { FineWeb-Edu: the Finest Collection of Educational Content },
  year         = 2024,
  url          = { https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu },
  doi          = { 10.57967/hf/2497 },
  publisher    = { Hugging Face }
}

@misc{penedo2024finewebdatasetsdecantingweb,
  title={The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale},
  author={Guilherme Penedo and Hynek Kydlíček and Loubna Ben allal and Anton Lozhkov and Margaret Mitchell and Colin Raffel and Leandro Von Werra and Thomas Wolf},
  year={2024},
  eprint={2406.17557},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2406.17557},
}