Details a data-centric training framework built from five complementary systems (Choulea, SkyReal, Hongzwang, PSBreakup, and Kreator) used to construct specialized cybersecurity training environments and process over 164,000 execution-verified trajectories. Describes the trajectory-filtering pipeline and infrastructure used to train frontier cyber-agent models, which achieved leading results on the CyberGym benchmark. Presented as an engineering account of the training pipeline rather than a single algorithmic contribution.
