Genesis III builds a 191B-token STEM dataset

Tether released QVAC Genesis III, a 191.43-billion-token synthetic STEM dataset spanning 159.6 million documents. Its release says a 1.7B-parameter model trained on Option-Level data produced valid answers in 99.45% of benchmark responses, while the token-matched comparison gained 28.57 points on ARC-Easy, 21.35 on ARC-Challenge, and 15.03 on MMLU STEM; those are vendor-reported results, not deployment proof. Builders weighing local models should test data quality before adding parameters—the archive’s 30B agent that fits a 24GB GPU makes the same compute-efficiency question practical.