VERITAS: A High-Quality, Balanced Multi-Domain Pre-training Dataset

📖 Dataset Summary

VERITAS is a high-quality, multi-lingual, and multi-domain dataset meticulously curated for the pre-training of Large Language Models (LLMs). Comprising exactly 250 Billion tokens, VERITAS is designed to provide a highly representative, dense, and clean data distribution.

The dataset features a strong orientation towards scientific AI, with a deep integration of high-grade mathematical reasoning, scientific literature, and rich code repositories, all while maintaining a robust foundation of general web knowledge and diverse linguistic coverage.

🌟 Key Features

• Proportionally Exact Mixing: Maintains precise, carefully calculated data mixing ratios across all domains to prevent domain collapse and ensure robust model generalization.
• High-Quality STEM: Deep integration of verified mathematical and scientific corpora (including FineMath, Nemotron-CC-Math, and Common Pile) to boost logical and analytical reasoning capabilities.
• Rich Code Representation: Includes high-quality, deduplicated code repositories and specialized mathematical coding datasets to enhance programming and algorithmic skills.
• Comprehensive Multi-lingual Coverage: Strong representation of English alongside targeted, high-quality corpora in 9 other major languages (French, German, Spanish, Russian, Chinese, Japanese, Italian, Portuguese, and Arabic).
• Rigorous Curation & Deduplication: Processed with strict quality filters, heuristic-based cleaning, and advanced cross-source Bloom-filter deduplication to ensure maximum data density and eliminate redundancy.

📊 Data Composition

The dataset is carefully balanced to reflect an optimal distribution for modern LLM training. The token distribution is structured across the following core domains:

Domain	Description	Key Sources Included
English Web	General knowledge, diverse linguistic patterns, and high-quality educational content.	FineWeb-Edu, FinePDFs-Edu, Wikipedia Monthly, Project Gutenberg
English Mathematics	High-grade mathematical reasoning, textbooks, and proofs.	FineMath (3+/4+), Nemotron-CC-Math
English Science	Academic papers, scientific literature, and research.	Common Pile (peS2o, arXiv)
Code	Multi-language programming, algorithms, and code-math reasoning.	OpenCoder (FineWeb Code/Math), Nemotron Scientific Coding, The Stack
Multilingual	High-quality non-English texts to ensure strong cross-lingual capabilities.	FineWeb-2, Wikipedia Monthly, 101B Arabic Words

📂 Dataset Structure

The dataset is provided in optimized Parquet format, split into multiple shards for efficient streaming and loading.

Each row contains the following columns:

• text (string): The raw, cleaned text document ready for tokenization.
• source (string): The original repository path of the document.
• source_id (string): The internal identifier for the specific sub-source.
• category (string): The high-level domain category (e.g., en, code, fr, ar).
• subset (string): The specific configuration or subset of the original source.
• row_idx (int64): The positional index in the original source stream (useful for auditing).

⚖️ License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You are free to:

• Share — copy and redistribute the material in any medium or format.
• Adapt — remix, transform, and build upon the material for any purpose, including commercially.

Under the following terms:

• Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. You must cite the creator of the VERITAS dataset.

📝 Citation

If you use VERITAS in your research, models, or commercial products, please cite it as follows:

@misc{veritas2024dataset,
  title={VERITAS: A High-Quality, Balanced Multi-Domain Pre-training Dataset},
  author={[Your Name/Handle]},
  year={2024},
  howpublished={\url{[Insert Link to your HuggingFace Repo or GitHub]}},
}