Tether's AI research division, QVAC, has launched Genesis III, a synthetic STEM dataset comprising 191.43 billion tokens. The dataset covers 159.6 million documents across 19 domains, categorized into high school, college, and professional difficulty levels.
A 1.7 billion-parameter model trained on a subset of this data achieved a 99.45% valid-answer rate on MMLU STEM benchmarks. This performance secured the team's research paper a spot at the COLM 2026 conference.
The training corpus utilizes two methodologies: Option-Level reasoning, which focuses on correct logic with 108.67 billion tokens, and Failure Analysis, which examines errors using 82.76 billion tokens.
Compared to the Cosmopedia-v2 baseline, models trained on Genesis III showed performance gains of 28.57 points on ARC-Easy, 21.35 points on ARC-Challenge, and 15.03 points on MMLU STEM.
Genesis III concludes a series that began in October 2025. The dataset is available under a CC-BY-NC 4.0 license for non-commercial use, while the model itself is released under the Apache 2.0 license, permitting commercial applications.
