Version labels do not establish independence
A dataset can receive a new version number while retaining examples that already influenced development. Renaming files or changing record identifiers does not make their contents unseen.
Exposure is broader than model training. Developers may inspect failed test items, revise a system prompt after reviewing answers, or select configurations using repeated evaluation results. Those activities can turn an evaluation collection into development evidence even when no model weights change.
The practical question is therefore not simply whether two files match. It is which information became available, to whom, and before which decisions.
Trace relationships beyond exact duplicates
Start with stable item identifiers and explicit parent-child relationships. Preserve links between an original item and its translations, paraphrases, corrected labels, shortened passages or alternative answer options.
These variants may share enough task information that placing them in different splits weakens the intended separation. Group related items before assigning training, development and test membership. Choose the grouping unit according to the task: it may be a source document, conversation, underlying question or another shared origin.
Exact hashes help identify unchanged content. Similarity searches can surface candidate relationships, but require review: unrelated items may look similar, while substantial rewrites may preserve the same answer-bearing information. Neither method alone establishes independence.
Record exposure events and tuning decisions
Maintain an exposure register alongside the dataset manifest. Useful fields include the item or family identifier, dataset version, date, access purpose, recipient role and associated development decision.
Distinguish automated scoring from inspection of item content or reference answers. Also record whether developers received only aggregate feedback, slice-level summaries or individual examples. These reveal different amounts of information and should not be treated as equivalent.