Re:Mistral 7B, the most powerful language model for its size...
by brucethemoose2 on Thursday September 28, 2023 @10:25AM
(#37690051)
AFAIK, such tests just feed the model chopped up bits of the evaluation data as raw strings with zero temperature. If it completes them verbatim, its probably in the training dataset.
I would be curious what does he mean by "semi-automated system for detecting benchmark leaks. " though.