How scoring works

Per task score

Every task defines its own metric, which is used to score your submissions. Only your best score per task counts on the leaderboard. A short cooldown applies between submissions, typically five minutes.

Public and final scores

While the hackathon is live, the leaderboard shows your public score, computed on a public subset of the evaluation set. After submissions close, the displayed score switches to the final score, computed on the held out subset. That final score determines the overall ranking. The split is fixed and identical for every team. The ratio between the public and held out subsets depends on the task:

• DUCI: 33% public, 67% held out
• Memorization in LMMs: 30% public, 70% held out
• LLM Watermark Detection: 30% public, 70% held out

Overall ranking

For each task, scores are normalized to the [0, 1] range using the lowest and highest scores across all teams. For tasks where lower raw scores are better (such as DUCI), the normalization is inverted so that 1.0 always represents the best result on that task. A team's overall rank is the average of its normalized scores across every task. Teams missing a score on any task are not ranked overall.