hoangduy02071997/tinyDTVi-dataset

TinyDTVi Pretraining Dataset

This is the official pretraining dataset for the TinyDTVi-500M causal language model. The corpus is aggressively filtered and deduplicated to maximize the representation of standard Vietnamese text for training resource-constrained models.

📊 Dataset Composition

The initial source corpora contained over 200GB of raw/uncompressed text, curated by mixing three open-source Vietnamese datasets. The approximate mixture ratio by document count is:

• 75% CulturaX-Vi (Cleaned web crawl)
• 24% VNews (Vietnamese News Corpus)
• 1% Vietnamese Wikipedia (Encyclopedia)

After strict preprocessing and bounded prefix-based deduplication, the final dataset yielded the following metrics:

• Documents: ~28.6 million
• Cleaned Size: ~134 GB
• Tokens: ~28 Billion

🧹 Data Processing Pipeline

To compress the vocabulary down to 50,304 tokens and focus strictly on standard Vietnamese semantics, we applied a highly aggressive filtering pipeline:

1. Unicode Normalization: Forced all text into Unicode Normalization Form C (NFC) to unify decomposed Vietnamese diacritics.
2. Artifact Removal: Stripped specific domain artifacts like Wikipedia citation brackets (e.g., [1], [citation needed]).
3. Regex-based Whitelisting: Retained only standard Vietnamese and English alphabetical characters, digits, basic punctuation, and fundamental mathematical operators. All foreign scripts (Chinese, Korean, Arabic, etc.), emojis, and complex markdown structures were discarded at the character level.
4. Deduplication: Applied bounded prefix-based deduplication (via hashing the first 500 characters) across all three data sources to prevent memorization.
5. Whitespace Normalization: All newline characters (\n) and redundant spaces were collapsed into a single space to maintain a dense, continuous token stream for training efficiency.

📚 Data Sources & Acknowledgements

This dataset is a derived research corpus assembled from multiple upstream datasets. Its use is subject to the licenses and terms of the respective source datasets.

License Provenance

Source Dataset	Proportion	Original License / Terms	Applied Transformation
CulturaX-Vi	~75%	Source-specific (Inherited from mC4 & OSCAR)	Bounded deduplication & regex filtering
VNews	~24%	Source-specific (Utilized under fair use for academic research)	Aggressive truncation & artifact removal
Vietnamese Wikipedia	~1%	CC BY-SA 4.0	Whitespace normalization & citation stripping

We strictly adhere to the non-commercial and research-only intent of the constituent datasets:

• CulturaX: uonlp/CulturaX
• VNews (BinhVQ News): vietgpt/binhvq_news_vi
• Vietnamese Wikipedia: wikimedia/wikipedia (20231101.vi)

⚠️ Limitations

Because this dataset has been stripped of newlines and structural formatting (like JSON, Markdown, or Code blocks), it is not suitable for training models on coding tasks or structured data generation. Furthermore, because the corpus is heavily weighted towards broad web-crawls (CulturaX), it inevitably contains instances of low-quality text, spam, and adult content. Trace amounts of PII (Personally Identifiable Information) may also remain despite our strict deduplication. This dataset is released strictly for research purposes.

📝 Citation

If you use this dataset, please remember to cite the original authors of the constituent datasets:

CulturaX:

@article{nguyen2023culturax,
  title={CulturaX: A Cleaned, Enormous, and Multilingual Dataset for Large Language Models in 167 Languages},
  author={Nguyen, Thuat and Chuang, Chien-Sheng and others},
  journal={arXiv preprint arXiv:2309.09400},
  year={2023}
}

VNews:

@misc{binhvq2024vnews,
  author = {BinhVQ},
  title = {Vietnamese News Corpus (binhvq_news_vi)},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/vietgpt/binhvq_news_vi}
}

Wikipedia:

@misc{wikimedia2023wikipedia,
  author = {Wikimedia Foundation},
  title = {Wikimedia Downloads},
  year = {2023},
  url = {https://dumps.wikimedia.org}
}