Genesis II Expands QVAC's Synthetic AI Education Dataset to 148 Billion Tokens

Tether Data’s artificial intelligence research division, QVAC, has unveiled the latest iteration of its open-source synthetic dataset, marking a substantial advancement in AI model pre-training capabilities. The new release introduces 107 billion additional tokens, propelling the total dataset to 148 billion tokens spanning 19 education-focused domains—establishing it as the world’s largest publicly available synthetic dataset for AI development.

Breakthrough in Synthetic Division and Reasoning Capabilities

The Genesis II dataset introduces a fundamental shift in how synthetic data structures training information. Rather than simple token accumulation, QVAC implemented a “synthetic division” approach that segments educational content into specialized domains, each optimized for specific learning objectives. This methodology enables more granular control over model training parameters.

A distinctive feature of this release is the introduction of “Option-Level Reasoning,” a novel training approach that guides AI models through multi-choice problem-solving frameworks. Unlike previous generations that focused on pattern recognition, this method explicitly teaches models the intermediate reasoning steps required to arrive at conclusions. Independent evaluations demonstrate that models trained on Genesis II data exhibit superior reasoning accuracy and produce more coherent, well-structured responses compared to earlier synthetic datasets.

Expanded Domain Coverage and Accessibility

Genesis II extends into previously underrepresented fields including computer science, statistics, and machine learning—domains critical for developing AI systems capable of solving complex analytical problems. This expansion builds upon the foundation established in Genesis I, which pioneered failure-analysis methodologies to identify and correct weak points in model reasoning.

The entire dataset is released under Creative Commons licensing and hosted on both QVAC’s official blog and Hugging Face, democratizing access to enterprise-grade training data. This open distribution model removes barriers for researchers and developers working on localized AI models, reducing dependence on proprietary, centralized AI development platforms.

Strategic Vision and Industry Impact

Paolo Ardoino, CEO of Tether, characterized this initiative as a pivotal step in moving artificial intelligence development beyond mere linguistic fluency toward robust, structured understanding. By providing free access to high-quality synthetic training data, QVAC enables the broader AI research community to develop more reliable and transparent models outside traditional corporate ecosystems.

The release underscores a growing recognition that quality pre-training data—particularly synthetically generated datasets optimized for educational value—represents a critical competitive advantage in model development. As AI systems become increasingly central to business and research applications, initiatives like Genesis II contribute meaningfully to the democratization of advanced model training capabilities.