Before reporting eval numbers.
The team needs to know whether held-out examples, answer keys, benchmark rows, or paraphrased eval material have drifted into training or fine-tuning data.
Eval Set Leakage
Check whether evaluation data can still be trusted before results are reported.
Use this before a benchmark, eval refresh, model comparison, or investor/customer report where leaked examples can make the model look better than it really is.
Leakage candidates with row-level evidence.
Datascreen surfaces overlap, answer-key residue, suspicious metadata, and source context so eval owners can review the evidence before trusting the score.
Source: Data Contamination Report from the 2024 CONDA Shared Task.
Watch walkthrough
Product demo: inspect an eval refresh before the result is used.
Show an eval set upload becoming a leakage review queue, with suspicious benchmark overlap and answer-key residue attached to the underlying rows.
What the product needs to make the decision obvious.
Overlap review
Surface exact and near-exact candidates that can compromise a held-out evaluation.
Answer residue
Flag labels, answer markers, and solution metadata that survived preprocessing.
Eval context
Keep benchmark, source, and project context attached to each finding.
Report boundary
Export what was reviewed before the score is shared.
A prioritized list of rows and clusters that deserve human review.
The row, source, neighborhood, and reason shown together.
A record of what reviewers kept, removed, fixed, or escalated.
A workflow-ready handoff that states what was reviewed and what remains uncertain.