Most of you have expressed concerns about the cross-validation scores being way better than the leaderboard scores; thank you for raising the concerns.

1. Most of the people I have talked to are using stratified folds, kfold, etc., to validate the scores locally, and there is a huge difference in the scores.

2. How many people are using the train/test split way(i.e., completely setting aside 30% of the train data) train the model and now use it to validate) to validate their scores? Are you still getting the same huge difference between your score and that on the leaderboard?

I have set aside 490 rows from the train data as part of my validation test, then trained my model(using the starter notebook) on 1656 rows of data - the score I get for the 490 rows is 0.054965986394557825, which is not very different from the leaderboard. This is just my observation. My request is could you faithfully do the same; please share the scores you get(no folds, please) :)

Let's work together to get the best solutions to conserve the turtles.