Tool
Whether train JSONL overlaps eval prompts
Train prompt keys are compared to eval lines (plain text or JSONL) after NFC and whitespace folding. This page does not invent a contamination rate for a public benchmark.
Comparison runs in this tab. GearDel does not receive train or eval text, and does not host KMMLU or KoBEST files.
What counts as overlap
Train lines must be JSONL objects. The compare key is Alpaca instruction plus input, the first user/human turn, or else question, query, prompt, text, or input.
Eval may be JSONL in those shapes, or one plain prompt per line. A line that starts with { is parsed as JSON; other lines are the key as-is.
NFC and whitespace folding match the dedup page. Exact overlap is equality of those folded keys. Optional substring overlap requires the eval key to be at least 24 code points and to appear inside a train key.
Hits list eval line numbers and matching train line numbers. overlapping-train.jsonl is the train objects that matched. clean-train.jsonl is the rest of the parsed train lines.
At most 20,000 lines or about two million characters on each side. A leading JSON array is rejected.
What this page does not do
- It does not download KMMLU, KoBEST, or KLUE. Paste questions you already have. Catalog family row counts stay on the eval-vs-train guide.
- It does not compute n-gram overlap, BLEU, or a dataset-wide contamination percentage.
- Substring mode can match unrelated sentences that share a long phrase. Exact mode does not.