LSH-MinHash Dedup May Discard Millions of Valid Training Documents

rupspace · x · 2026-08-13

MBZUAI research points out that LSH-MinHash deduplication, commonly used in LLM pre-training, might mistakenly discard millions of valid training documents.

The Culprit: The method treats overlapping duplicate buckets as transitive equivalence and keeps only one document per connected component, ignoring that local overlap does not equal global duplication.

Original post →

More from Research

Research channel →