Dataset Deduplication and Near-Duplicate Detection

Exact deduplication identifies records with identical bytes or normalized text. Cryptographic hashes are efficient for exact matches, but they will not detect edits, formatting changes, or paraphrases. Near-duplicate methods use fingerprints such as n-grams and MinHash, locality-sensitive hashing, or embedding similarity. Each method defines “similar” differently and has different cost and false-positive behavior. Duplicates matter for training and evaluation. Repeated examples can overweight a pattern, increase memorization, and place highly similar material in both training and test partitions. A 2021 research paper on language-model training found near-duplicate text in major datasets and reported reduced memorized text emission after deduplication; the paper also found substantial train-test overlap in some benchmark validation data. Those results are specific to the studied corpora and methods. Deduplication is not simply “remove everything similar.” Repeated records may represent real frequency, legitimate variants, multiple sources, or distinct labels. A global similarity threshold can remove dialect, minority-language, or domain-specific examples disproportionately. Preserve provenance, define which fields and transformations are compared, and audit removal rates by source and relevant groups. For leakage control, identify duplicate clusters before splitting so related records can be assigned together. Keep an audit log of the algorithm, normalization, threshold, and retained representative. Evaluate the pipeline on manually reviewed pairs and examine both missed matches and false matches. Deduplication improves dataset hygiene only when its policy matches the task and preserves meaningful diversity.