AI Models Achieve Strong Benchmark Improvements Through Specialized Dataset Training

Newly developed AI models demonstrated significant performance improvements across multiple benchmarks following training on specialized datasets from different domains. These gains were achieved using only a portion of the total available data, suggesting considerable upside potential from leveraging complete datasets. Performance metrics revealed improvements of 12.0 percentage points on Terminal-Bench 2.0, 9.6pp on Toolathlon, and 9.3pp on FrontierScience. Notably, datasets without code-related content also contributed to performance on coding tasks, with a 6pp gain observed on SWE-Bench Pro. The trained models are now available for commercial use, having been developed with just one-fifth of the full dataset. Larger versions utilizing the complete data are expected to achieve even higher performance levels.