MixtureVitae : Open Web-Scale Pretraining Dataset With High Quality Instruction and Reasoning Data Built from Permissive-First Text Sources
(2026) In Transactions on Machine Learning Research 2026-April.- Abstract
We present MixtureVitae, an open-access pretraining corpus1 built to minimize legal risk while providing strong downstream performance. MixtureVitae follows a permissive-first, risk-mitigated sourcing strategy that combines public-domain and permissively licensed text (e.g., CC-BY/Apache) with carefully justified low-risk additions (e.g., government works and EU TDM-eligible sources). MixtureVitae adopts a simple, single-stage pretraining recipe that integrates a large proportion of permissive synthetic instruction and reasoning data—signals typically introduced during post-training and generally scarce in permissive web corpora. We categorize all sources into a three-tier scheme that reflects varying risk levels and provide... (More)
We present MixtureVitae, an open-access pretraining corpus1 built to minimize legal risk while providing strong downstream performance. MixtureVitae follows a permissive-first, risk-mitigated sourcing strategy that combines public-domain and permissively licensed text (e.g., CC-BY/Apache) with carefully justified low-risk additions (e.g., government works and EU TDM-eligible sources). MixtureVitae adopts a simple, single-stage pretraining recipe that integrates a large proportion of permissive synthetic instruction and reasoning data—signals typically introduced during post-training and generally scarce in permissive web corpora. We categorize all sources into a three-tier scheme that reflects varying risk levels and provide shard-level provenance metadata to enable risk-aware usage. In controlled experiments using the open-sci-ref training protocol (fixed architectures and hyperparameters; 50B and 300B token budgets across 130M–1.7B parameters), models trained on MixtureVitae consistently outperform other permissive datasets across a suite of standard benchmarks, and at the 1.7B-parameters/300B-tokens setting, they match FineWeb-Edu and approach DCLM–demonstrating that the large fraction of reasoning and instruction data does not come at the cost of general-purpose language understanding. Performance is particularly strong on MMLU and on math and code benchmarks: a 1.7B model pretrained on 300B MixtureVitae tokens outperforms all strong non-permissive reference datasets and matches or exceeds smolLM2-Instruct, a strong 1.7B instruction-tuned baseline on GSM8K, HumanEval, and MBPP, despite using over 36× fewer tokens (300B vs. ≈11T). Supported by a thorough decontamination analysis, these results show that permissive-first data with high instruction and reasoning density, tiered by licensing and provenance-related risk, can provide a practical and risk-mitigated foundation for training capable LLMs, reducing reliance on broad web scrapes without sacrificing competitiveness.
(Less)
- author
- organization
- publishing date
- 2026
- type
- Contribution to journal
- publication status
- published
- subject
- in
- Transactions on Machine Learning Research
- volume
- 2026-April
- external identifiers
-
- scopus:105038402999
- ISSN
- 2835-8856
- language
- English
- LU publication?
- yes
- id
- 84fc6487-cb40-4830-ad69-ac943ab92bfd
- date added to LUP
- 2026-07-08 11:51:37
- date last changed
- 2026-07-08 11:52:33
@article{84fc6487-cb40-4830-ad69-ac943ab92bfd,
abstract = {{<p>We present MixtureVitae, an open-access pretraining corpus<sup>1</sup> built to minimize legal risk while providing strong downstream performance. MixtureVitae follows a permissive-first, risk-mitigated sourcing strategy that combines public-domain and permissively licensed text (e.g., CC-BY/Apache) with carefully justified low-risk additions (e.g., government works and EU TDM-eligible sources). MixtureVitae adopts a simple, single-stage pretraining recipe that integrates a large proportion of permissive synthetic instruction and reasoning data—signals typically introduced during post-training and generally scarce in permissive web corpora. We categorize all sources into a three-tier scheme that reflects varying risk levels and provide shard-level provenance metadata to enable risk-aware usage. In controlled experiments using the open-sci-ref training protocol (fixed architectures and hyperparameters; 50B and 300B token budgets across 130M–1.7B parameters), models trained on MixtureVitae consistently outperform other permissive datasets across a suite of standard benchmarks, and at the 1.7B-parameters/300B-tokens setting, they match FineWeb-Edu and approach DCLM–demonstrating that the large fraction of reasoning and instruction data does not come at the cost of general-purpose language understanding. Performance is particularly strong on MMLU and on math and code benchmarks: a 1.7B model pretrained on 300B MixtureVitae tokens outperforms all strong non-permissive reference datasets and matches or exceeds smolLM2-Instruct, a strong 1.7B instruction-tuned baseline on GSM8K, HumanEval, and MBPP, despite using over 36× fewer tokens (300B vs. ≈11T). Supported by a thorough decontamination analysis, these results show that permissive-first data with high instruction and reasoning density, tiered by licensing and provenance-related risk, can provide a practical and risk-mitigated foundation for training capable LLMs, reducing reliance on broad web scrapes without sacrificing competitiveness.</p>}},
author = {{Nguyen, Huu and May, Victor and Raj, Harsh and Nezhurina, Marianna and Wang, Yishan and Luo, Yanqi and Vu, Minh Chien and Nakamura, Taishi and Tsui, Ken and Nguyen, Van Khue and Salinas, David and Krasnodębska, Aleksandra and Schuhmann, Christoph and Richter, Mats Leon and Vu, Xuan Son and Jitsev, Jenia}},
issn = {{2835-8856}},
language = {{eng}},
series = {{Transactions on Machine Learning Research}},
title = {{MixtureVitae : Open Web-Scale Pretraining Dataset With High Quality Instruction and Reasoning Data Built from Permissive-First Text Sources}},
volume = {{2026-April}},
year = {{2026}},
}