Semantic Decontamination Policies Cause High False Removal Rates in RLVR Math Data
October 16, 2026
Auditing sixteen RLVR math corpora reveals that semantic matching policies fail to meet 90% precision targets. At a 0.80 cosine threshold, 88% of sampled removals were confirmed as false positives, indicating significant data loss during benchmark decontamination.
HOW THIS AFFECTS YOU
●
builderExisting decontamination pipelines may be inadvertently stripping high-quality training data from your math models.
●
researcherYou should be cautious when applying semantic deduplication to training sets to avoid losing valuable signal.