Provides scientific-reasoning and tool-use trajectories for training agent foundation models, alongside reinforcement-learning tasks. Its staged data makes the Axion 1.0 training process easier to inspect and reproduce.