“A significant bottleneck for training effective long-context models is the scarcity of high-quality, long-context data available to research institutions.”