Eval-set leakage inspector
Compare the splits of an eval set and see which prompts appear in more than one. Reports the overlap it finds, with the rows on each side, and never claims that finding none proves there is no leakage.
Your input is processed in this browser. It is not uploaded. A file or pasted rows go nowhere. A URL is fetched by your browser directly from the address you give it. Sharing creates a link containing the data you choose to include.
This link will contain
Anyone with the link can read this. Compression is not encryption.
What it does
A prompt that sits in your training split and in your test split turns the test score into a measurement of memorisation. Nothing that checks one file at a time can see it, because each file is individually fine. The fault exists only in the relationship between two of them.
Paste rows carrying a split field, or an object whose keys are the split names. Every pair of splits gets two passes: exact, then a second one ignoring case and spacing. Each overlap is reported with the value and the rows it sits on, on both sides.
Two percentages, never one
One of four training rows leaking is also one of one test rows. A single number would have to pick a denominator, and either choice misstates the other side, so both are always shown, as a fraction next to the percentage.
What it does not do
It does not detect paraphrases, translations or near-duplicates that differ by a word. That is a decision rather than an omission: a similarity threshold would turn found nothing into a claim about your data instead of a statement about what was compared. Overlap found is overlap reported. Finding none does not prove there is no leakage, and every report says so, including a clean one.