Tether released a dataset of 191.4 billion Token STEM, betting on the "explanatory ability" of local small models

The stablecoin giant Tether's Tether AI Research today released QVAC Genesis III, a synthetic training dataset aimed at the fields of science, technology, engineering, and mathematics, with a scale of 19.143 billion tokens, covering 159.6 million documents.

It is reported that Tether hopes to enhance the reasoning and teaching capabilities of small AI models, allowing more AI assistants to run directly on laptops, mobile phones, and local servers, reducing the ongoing reliance on large cloud models.

Genesis III continues the data lineage of Genesis I and Genesis II, covering 19 curriculum-related fields including biology, chemistry, physics, mathematics, computer science, medicine, electrical engineering, and machine learning, aimed at high school, university, and professional applications. Tether states that the dataset will focus on training models to explain problem-solving paths, identify erroneous reasoning, and provide corrections, rather than just outputting final answers.

According to Tether's disclosed test results, a 1.7 billion parameter model trained with Genesis III option-level data achieved an effective answer rate of 99.45% in relevant benchmark tests. Compared to the Cosmopedia-v2 training model of similar token scale, Genesis III improved scores on the ARC-Easy, ARC-Challenge, and MMLU STEM benchmarks by 28.57, 21.35, and 15.03 percentage points, respectively.