Discover top fintech news and events!
Subscribe to FinTech Weekly's newsletter
Read by executives at JP Morgan, Coinbase, Blackrock, Klarna and more
A Major Expansion in Open AI Training Data
Tether Data has released a new version of its synthetic educational dataset for artificial intelligence, significantly increasing the volume and scope of open training material available to researchers worldwide. The company’s AI research division, QVAC, announced that the new release, called QVAC Genesis II, adds 107 billion tokens to its earlier dataset, bringing the total size to 148 billion tokens.
The expanded dataset is now the largest publicly available synthetic educational resource designed specifically for AI pre-training. It spans 19 academic domains and is intended to improve how models learn reasoning, explanation, and decision-making rather than surface-level pattern recognition.
The announcement positions the release as a step toward more transparent and accessible AI development, at a time when many advanced training datasets remain locked inside proprietary systems.
Building on the First Genesis Release
QVAC Genesis II builds on work first introduced with Genesis I, which focused on creating a validated, education-centered synthetic dataset covering core science, technology, engineering, and mathematics subjects. That earlier release established a framework for generating structured training questions aimed at improving reasoning accuracy.
The new release expands coverage into ten additional fields, including chemistry, computer science, statistics, machine learning, astronomy, geography, econometrics, and electrical engineering. It also revisits college-level physics content, regenerating it using an updated methodology designed to improve conceptual clarity.
Together, the two releases form what QVAC describes as the most extensive synthetic educational dataset yet made available to the public. The dataset is intended for use in pre-training large language models and other AI systems that require structured academic material.
A Shift in How Training Data Is Generated
At the core of Genesis II is a new data generation method referred to as Option-Level Reasoning. This approach differs from many existing synthetic data techniques by focusing not only on incorrect answers, but also on correct ones.
Instead of treating a correct response as the end of the process, the method analyzes every answer option in a multiple-choice question. Correct choices are broken down to reinforce why they are correct, while incorrect options are examined to address common misconceptions. This structure allows models to learn causal reasoning and decision logic rather than simply associating questions with outcomes.
The approach complements the Failure Analysis method introduced in Genesis I, which focused on extracting value from model errors. Together, the two methods form a pipeline where each generated question is designed to contribute instructional value.
Independent evaluations cited by QVAC indicate that models trained on Genesis II data show higher reasoning accuracy and produce clearer answers than those trained on earlier synthetic datasets.
Emphasis on Understanding Over Fluency
Much of the current AI training ecosystem relies on assembling very large volumes of text, often scraped from public sources, to improve language fluency. QVAC’s stated goal differs in emphasis. The Genesis datasets are structured to teach models how to reason through problems and explain conclusions in a clear way.
Company leadership has indicated that the intention is to move beyond training systems that predict likely text sequences, toward models that demonstrate understanding of underlying concepts. The dataset design prioritizes clarity, causality, and logic, aiming to reduce ambiguity in model outputs.
This approach aligns with broader discussions in AI research about reliability and explainability, especially as AI systems are used in education, science, and decision-support contexts.
Open Access for Researchers and Developers
As with the original Genesis dataset, QVAC Genesis II is being released openly. The dataset is available under a Creative Commons Attribution–NonCommercial 4.0 license, allowing researchers, academic institutions, and independent developers to use and study the data outside of commercial settings.
The dataset and associated models are hosted on Hugging Face, alongside a detailed technical paper outlining the generation methodology and evaluation results. This open distribution is intended to lower barriers for researchers who do not have access to large proprietary datasets.
By maintaining non-commercial licensing, QVAC aims to support academic and community-driven research while limiting direct commercial exploitation.
Supporting Decentralized AI Development
The release also fits within a broader strategy pursued by Tether Data to encourage decentralized AI development. The company has stated that high-quality training data should not be restricted to organizations with access to centralized cloud infrastructure.
By making large-scale, structured datasets publicly available, QVAC seeks to enable local training, experimentation, and deployment of AI models. This approach is intended to support research environments where compute resources may be limited but intellectual contributions remain significant.
The emphasis on decentralization reflects growing interest in reducing reliance on a small number of dominant AI platforms and fostering a more distributed research ecosystem.
Tether’s Role in AI Research
QVAC operates as the AI research division of Tether Data. While Tether is widely known for its role in digital assets and stablecoins, the company has expanded its activities into data and AI research in recent years.
Through QVAC, Tether Data has focused on building infrastructure and resources that support open research. The Genesis datasets represent one of the most visible outputs of that effort, positioning the company within discussions around open AI development and education-focused training data.
This work also reflects the growing overlap between fintech companies and advanced AI research, as financial technology firms increasingly invest in data science and machine learning capabilities.
Leadership Perspective on the Release
Company leadership has framed the Genesis II release as a move away from training approaches that prioritize volume alone. The focus, according to statements from Tether’s executive team, is on teaching AI systems how to reason and explain rather than merely generate fluent responses.
Paolo Ardoino, chief executive of Tether, has emphasized that reliable AI should be grounded in understanding why answers are correct. He has indicated that making the dataset openly available reflects a belief that stronger, more explainable AI benefits society as a whole.
These views echo concerns raised by researchers about the limitations of models trained primarily on unstructured text.
Educational Scope and Domain Coverage
The combined Genesis I and II datasets cover 19 domains, with content designed at secondary and tertiary education levels. Subjects range from foundational mathematics and physics to applied fields such as econometrics and machine learning.
Each domain includes structured questions, explanations, and reasoning pathways intended to mirror how concepts are taught and assessed in formal education settings. This design is meant to support pre-training tasks that require logical consistency and conceptual depth.
By regenerating and expanding content using improved methods, QVAC aims to refine how educational material is represented in synthetic datasets.
Evaluation and Model Performance
According to internal and independent evaluations referenced by QVAC, models trained on Genesis II data show improved performance in reasoning-heavy tasks. These include answering structured questions, explaining conclusions, and avoiding ambiguous or contradictory responses.
The evaluation results suggest that the combination of Failure Analysis and Option-Level Reasoning leads to more consistent outputs. While the company has not positioned the dataset as a standalone solution, it has presented it as a strong foundation for further training and fine-tuning.
Researchers are expected to conduct additional evaluations as the dataset sees wider use in the community.
Implications for Open AI Research
The release of such a large, open dataset may influence how academic and independent researchers approach model training. Access to structured educational data at this scale has traditionally been limited to well-funded organizations.
By providing an alternative, QVAC Genesis II could support experimentation with smaller models, localized training efforts, and research into explainable AI methods.
The dataset may also serve as a benchmark for future synthetic data projects that prioritize reasoning quality over sheer size.
Position Within the Broader AI Ecosystem
QVAC Genesis II enters an AI ecosystem marked by rapid development and increasing concentration of resources. Many of the most capable models are trained on proprietary datasets that are not accessible for scrutiny or replication.
Open datasets like Genesis II offer a counterpoint, enabling transparency and shared progress. They also raise questions about how open resources can coexist with commercial AI development.
The involvement of a company rooted in fintech and digital assets highlights how AI research is drawing interest from a wide range of industries beyond traditional technology firms.
Availability and Next Steps
The full technical documentation for the dataset, titled “QVAC Genesis II: Expanding the Largest and Highest-Quality Multi-domain Educational Synthetic Dataset for Pre-training,” has been published on the QVAC research blog. Access to the dataset and related models is available through Hugging Face.
QVAC has indicated that it plans to continue refining its methods and expanding educational coverage in future releases. Feedback from the research community is expected to play a role in shaping subsequent iterations.
A Continuing Push for Open Foundations
With Genesis II, QVAC reinforces its position that open, structured training data is essential for building reliable AI systems. The release reflects a view that intelligence should be grounded in reasoning and explanation, not just statistical association.
As AI systems become more integrated into education, science, and financial services, including fintech applications, the quality of their training data will remain a central concern.
For now, the expanded Genesis dataset stands as a notable contribution to open AI research, offering scale, structure, and accessibility at a level rarely seen outside proprietary environments.
