Aquiles-ai/TinyQwen-Data

A Dataset for LLM Pre-training

TinyQwen-Data is a multilingual pre-training corpus containing approximately 15 billion tokens, designed for training compact language models in the 500M–1B parameter range.

The dataset consists primarily of English text, complemented by a high-quality Spanish subset, and follows a lightweight schema consisting of raw text and document provenance.

• Languages: English, Spanish
• Primary use: Language model pre-training
• Format: text, source

This dataset was prepared for the pre-training experiments described in an upcoming technical blog post.

Dataset Structure

Data Fields

Each row contains:

• text: document content
• source: source dataset from which the document originated

Dataset Statistics

Overall

Metric	Value
Documents	20,347,351
Tokens	15,000,004,857

Language Distribution

Language	Documents	Tokens
English	15,480,886	11,250,004,580
Spanish	4,866,465	3,750,000,277

Source Data

TinyQwen-Data was built from publicly available pre-training corpora and further processed into a unified corpus for language model pre-training.

The primary source datasets are:

• FineWeb2 - a multilingual web corpus developed by Hugging Face from Common Crawl using a reproducible filtering, deduplication, language identification, and quality filtering pipeline covering more than 1,000 languages.
• UltraX-Preview - a collection of refined pre-training corpora generated using the UltraX adaptive programmatic editing framework, which improves corpus quality through deterministic editing operations rather than end-to-end rewriting.

Documents from these corpora were sampled, filtered, merged, and normalized into a unified dataset intended for compact language model pre-training.

Intended Use

TinyQwen-Data is intended for:

• Language model pre-training
• Continued pre-training
• Training decoder-only language models
• Tokenizer development
• Research on multilingual language modeling
• Data quality research

Limitations

• The dataset inherits characteristics and potential biases from its source corpora.
• Language identification is performed automatically and may contain occasional errors.
• Although quality filtering has been applied, noisy or low-information documents may still be present.
• This dataset is intended as a pre-training corpus and should not be considered a factual knowledge source.

Acknowledgements

This work builds upon the following open datasets:

• HuggingFaceFW/FineWeb2
• OpenBMB/UltraX-Preview

We sincerely thank the authors and maintainers of these projects for making high-quality pre-training corpora publicly available to the research community.

Citations

@misc{penedo2025fineweb2pipelinescale,
  title={FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language}, 
  author={Guilherme Penedo and Hynek Kydlíček and Vinko Sabolčec and Bettina Messmer and Negar Foroutan and Amir Hossein Kargaran and Colin Raffel and Martin Jaggi and Leandro Von Werra and Thomas Wolf},
  year={2025},
  eprint={2506.20920},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2506.20920}, 
}

@misc{ultrax2026,
  title={UltraX: Refining Pre-Training Data at Scale with Adaptive Programmatic Editing},
  author={Xinlong Zhao and Dongsheng Liu and Hengyu Zhao and Zixuan Fu and Zheng Wang and Jie Cai and Jie Zhou and Qiang Ma and Xuanhe Zhou and Xu Han and Yudong Wang and Zhiyuan Liu},
  year={2026},
  eprint={2607.08646},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
}