Why do you have different models for different dataset?

[Coronal-Halo]

What's the rationale for training different models for different datasets? Were you able to achieve similar accuracies for each dataset as claimed in the paper by training a single model on multiple datasets?