后面附上翻译 😀 
Large-scale web datasets often form a chaotic “digital swamp” - 
—unstructured, heterogeneous collections with substantial redun- 
dancy that challenge machine learning workflows. Existing cleaning 
methods typically assume pre-sorted data or lack robustness against 
near-duplicates and mixed modalities. This paper introduces MMd- 
edup, an end-to-end framework built on a novel “Triage-a...