Potential methods to improve validator performance

[jsomeara]

There's a couple things we could do to improve the performance of the validator models:

• Try switching to dinov3 as the backbone
• Try switching to RampNet as the backbone
• In the dataset generation, instead of requiring disagree_count - agree_count > 1 for inclusion in the incorrect class, we could add a second route for inclusion in the incorrect class: if it is a validation from Jon or Mikey! I imagine their validations are more accurate than the crowd's. We might also want to make their validations supersede the crowd's?