clarification on method

[conway-abacus]

Hi @swj0419 and other authors, thanks for making this code available and easy to run so we can explore contamination in various open-source models.

Given that this repo/approach has gained some adoption in the community in terms of reporting contamination scores on benchmark datasets, I would like to clarify some things about how these scores are calculated:

• the computed/returned value of the script seems to not be the min-k%-prob formula give in the paper. Is there a reason?
• "If #the result < 0.1# with a percentage greater than 0.85, it is highly likely that the dataset has been trained." how this threshold is determined?
• for the min-k%-prob number, assuming we were to compute it, is there similar guidance on thresholds above which we have high confidence the dataset has been trained?

I could not find these specific details in the original paper - apologies if I missed them. Thanks in advance!