Every competition has one: the notebook with a 0.99 CV that drops to 0.71 on the private leaderboard. I have been that notebook. A "best" score of mine turned out to depend on a leaky validation setup, and in another competition the best public checkpoint had been trained on images that overlapped the test set.
The fixes are well known (Pipelines, GroupKFold, out-of-fold target encoding), but the mistakes are hard to see in a 400-cell notebook at 2am. LeakLens makes them visible. It is deterministic, explains every finding, and shows the fix.
What the community has found so far