SPICE: Submodular Penalized Information-Conflict Selection for Efficient Large Language Model Training | ScienceToStartup | ScienceToStartup