Missing Linear Probing Code Used in Assessing Sensitive Attributes

[Lancelottery]

Hello,

I'm trying to replicate the results from your preprint, particularly the section on "Foundation Model Encodes Demographic Information Beyond Human Levels."

In the paper, you assessed CheXzero's predictions on sensitive attributes using a linear attribute prediction head and reported high AUCs (e.g., "Black" AUC = 0.78). In the repository, I found the get_representations() and fit_model() functions in utils/lin_eval.py, but I couldn't find where these functions are called or how to properly utilize them to replicate your results.

Could you please share the code or provide guidance on how to use the linear probe to assess sensitive attributes (specifically races) in CheXzero?

Thank you for your assistance.

Best regards,
Lance