MixtureVitae: A Permissive, High-Performance, Open-Access Pretraining Dataset

Overview

MixtureVitae is an open-source, permissive, high-quality dataset designed for pretraining large language models (LLMs) across a wide variety of modalities, domains, and languages. The goal of MixtureVitae is to accelerate the development of transparent, open-access AI while lowering legal uncertainty around copyright and data provenance. See our blog.

• Please note this dataset is still being uploaded in parts. More shards will appear over time. Please be patient.

Features

• 1 Trillion+ Tokens: MixtureVitae includes over 1 trillion tokens of diverse text and multimodal content, carefully filtered for copyright-permissiveness and enriched with high-quality synthetic data.
• Cross-Modality: Includes textual, visual, and auditory elements; sourced and generated to support multimodal and multilingual LLM training.
• Transparent and Open: Based on publicly available data, permissive licenses (e.g. CC-BY, MIT, Apache), and public domain sources. Built with rigorous filtering and legal and ethical considerations.
• Diversity & Balance: Includes multimodal, narrative, conversational, instructive, educational, legal, scientific, and programming content across multiple domains and languages.

Data Components

MixtureVitae comprises three main categories:

Web-Based Open Datasets (Filtered)

• Nemotron-CC, Cosmopedia, FineWeb-Edu, TxT360, Cultura-Y, etc.
• Global deduplication and permissive heuristic filtering applied (e.g. .gov domains, CC-BY keywords, spam/obscenity filtering).

Curated Datasets

• Includes subsets and cleanups from Open License Corpus, PG-19, Freelaw, Stack v1, Euro-Pat, USPTO, Wikipedia, arXiv, OpenWebMath, Megawika, Europarl, HackerNews, and more.
• Covers legal, scientific, technical, conversational, and multilingual data.

Synthetic Data

• Math textbooks, Tiny-stories style narratives, Cross-language code translation, MCQ generation, Multimodal grounding, Multilingual translations, and more.

Preprocessing & Filtering

• Permissive Filtering: Heuristic and keyword filtering to retain CC-BY, public domain, and .gov sources while excluding unsafe/unclear cases.
• Light Global Deduplication: Prefix-based matching due to deduplication already performed in source corpora.
• Sentence Deduplication: Low-information duplicate detection with WordNet substitution.
• FastText Filtering & Classification:
  • Domain Classifier (based on FineWeb & Pile)
  • Genre/Register Classifier (TurkuNLP)
  • Math/Education quality Rankers (inspired by DeepSeekMath & Phi-3)
  • Red Pajama quality rankers
• Quality Upsampling: Classification and rank allows users to apply targeted upsampling of diverse content types.

Dataset Size & Format

• Over 1 trillion tokens total, not including multimodal data.
• Multimodal shards include aligned image captions, audio transcripts, and instruction-style text.
• Currently releasing only mostly english text shards, but will slowly release multimodal and transaltions.
• Sharded and deduplicated to enable scalable training on clusters or cloud.

Links To Component Datsets

• TBD: List component datasets such as MixtureVitae-atomic_2024, and other MixtureVitae-* datasets.

Legal Considerations

MixtureVitae is designed with legal caution, transparency, and fair-use alignment:

• Heavy reliance on public domain, open licenses, and US federal government content.
• Filtering for third-party copyrighted content.
• Ethical justifications and fair use arguments applied to .gov content.
• We do not guarantee non-infringement and disclaim legal liability — researchers are advised to consult legal experts before commercial use.

Intended Uses

• Pretraining LLMs across text and multimodal domains.
• Research into legal-compliant open model development.
• Instruction tuning, alignment training, and multilingual or cross-domain generalization.

---

Licensing

We license our own contributions and annotaitons under CC-BY-SA. MixtureVitae itself includes sources under their own individual licenses:

• Creative Commons (CC-BY, CC-BY-SA)
• Public domain or governmental data (.gov, .mil)
• Permissive software/data licenses (MIT, BSD, Apache)
  However, as with any large corpus: Use at your own legal discretion.

---

Contributors

This dataset was created by Ontocord.AI, with support from collaborators and references from open AI research ecosystems. Built as part of the Aurora-M2 project. We thank the contributors of datasets like Nemotron-CC, Cosmopedia, FineWeb, Open License Corpus, and many others.

---

How to Cite

@misc{txt360data2024,
      title={MixtureVitae: A Fully Permissive, High-Performance, Open-Access Pretraining Dataset}, 
      author={Harsh Raj, Huu Nguyen, Ken Tsui, Diganta Misra, Victor May, Vu Minh Chien},
      year={2025}
}