Tether’s AI research division, QVAC, has launched Genesis III, a training dataset comprising 191.43 billion tokens. The collection focuses on 19 STEM fields and aims to improve the reasoning capabilities of smaller language models.
The dataset utilizes a dual teacher-distillation strategy. It provides corrective explanations for errors and contrastive reasoning for correct answers, helping models learn from both successes and failures.
Genesis III is the latest in a series, following Genesis I with 41 billion tokens and Genesis II with 148 billion tokens. The data is hosted on Hugging Face under a CC-BY-NC 4.0 license.
Models trained on Genesis III showed performance improvements of 28.57% on the ARC-Easy benchmark and 21.35% on the ARC-Challenge compared to those using Cosmopedia-v2. Valid answer rates reached 99.45%.
Technical documentation for the release was submitted to arXiv on September 17, 2026. The project aligns with Tether's focus on decentralized, on-device AI systems that function without cloud server reliance.
