Tether releases its 191.4 billion-token STEM dataset, betting on the explainability of local small models.

Stablecoin giant Tether’s Tether AI Research today launched QVAC Genesis III, a synthetic training dataset for science, technology, engineering and mathematics (STEM) fields, boasting 191.43 billion tokens and covering 159.6 million documents. Tether aims to leverage the dataset to enhance the reasoning and teaching capabilities of small AI models, enabling more AI assistants to run directly on laptops, mobile devices and local servers, thus reducing ongoing reliance on large cloud-based models. Genesis III continues the data trajectory of Genesis I and II, spanning 19 curricula-related domains including biology, chemistry, physics, mathematics, computer science, medicine, electrical engineering and machine learning, with applications targeting high school, university and professional use cases. Tether states the dataset prioritizes training models to explain problem-solving processes, spot flawed reasoning and offer corrections, rather than merely outputting final answers. According to Tether’s disclosed test results, a 1.7-billion-parameter model trained on Genesis III’s option-level data posted a 99.45% valid response rate in relevant benchmarks. Compared with models trained on similarly-sized Cosmopedia-v2, Genesis III delivered score improvements of 28.57, 21.35 and 15.03 percentage points on the ARC-Easy, ARC-Challenge and MMLU STEM benchmarks respectively.