That’s the classic leaderboard shake-up we all secretly fear It really shows how public LB can sometimes reward models that slightly overfit to the validation split, while private LB exposes which solutions truly generalize. Jumping from rank 332 to 37 based on which submission counted is wild — but also a great reminder that consistency and robustness matter more than squeezing out that extra 0.003 on public. Honestly, experiences like this are part of what makes competitions exciting (and humbling). You think you’ve “cracked it,” then the private board rewrites the story. Despite the data hiccups, it sounds like it was a solid learning experience. Big respect for taking it in good humor — that mindset is what makes strong competitors in the long run