Task-specific features

Hi,

For the acoustic track, I’m noticing a few different tasks that test subjects are completing. Can we assume that the hold-out test set with which we will be evaluated contains the same general distribution of tasks? Would automatically detecting which task the audio contains, in order to apply task-specific features, be a valid approach?

Thanks,
Evan

Hello! We’ve posted more information about rules for features here that I think address your question: Clarification about Model Features.

Regarding the held out test set, we don’t share details about the split. Our advice is to create the best solution possible that doesn’t over-fit the public leaderboard.