Dated Data

"Tracing Knowledge Cutoffs in Large Language Models" (COLM 2024): shows a model's effective knowledge cutoff often differs sharply from its reported cutoff — driven by stale CommonCrawl dumps and deduplication artifacts in pretraining corpora — and gives a probing method to estimate the real cutoff per resource. Cheng, Marone, Van Durme et al.; pre-window preprint with in-window COLM publication.