We run open machine learning competitions for students in Rwanda. Each one is a real dataset with a held-out answer key, a public leaderboard, and a scoring engine that grades every submission the same way.
Your predictions are compared against a held-out answer key you never see. That is what makes the leaderboard meaningful: a model that memorised the training set scores badly, because it has not learned anything that generalises.
Scoring combines accuracy with an F1 measure, so a model that wins by always predicting the majority class does not do well. Some challenges also weight harder rows more heavily, which rewards models that hold up on the difficult cases rather than the easy bulk.
Mostly students who have finished a machine learning course and want to find out whether any of it transfers to data nobody has cleaned for them. You do not need to be enrolled anywhere to take part, and you do not need to have competed before.
If you are starting out, a plain baseline submitted early is worth more than a sophisticated model submitted never. Getting any score on the board tells you your pipeline works end to end, and everything after that is improvement.