Reproduce “Data Contamination: From Memorization to Exploitation” on OpenResearch

About this paper

Inbal Magar, Roy Schwartz

Researchers from The Hebrew University of Jerusalem developed a controlled experimental framework to quantify data memorization and exploitation by pretrained language models from contaminated test sets. The study empirically demonstrated how factors like data duplication and model size influence these phenomena, revealing that exploitation can occur, affecting downstream task performance, but is not solely dependent on memorization.

Key insights

• Memorization of contaminated labels is a necessary but not sufficient condition for exploitation, as low memorization consistently yielded negligible exploitation.
• The extent of both memorization and exploitation is directly proportional to the number of contaminated data duplications and the probability of masking the label token during pretraining.
• Model size (BERT-large vs. BERT-base) and the timing of contamination during pretraining significantly influence both memorization and exploitation levels, with early contamination driving higher exploitation.

Summary by alphaXiv.