V2.5 synthetic dataset with balanced diversity - expanded vocabulary, multiple text formats, reduced template predictability while maintaining clear category boundaries

initial commit