_get_calibration_data doesn't throw an exception if there are no target psms, contrary to _get_calibration_psms. However, unless I'm missing something, fit will still fail if observed_rt_calibration or predicted_rt_calibration are empty. So it should probably be handled here if one of the runs doesn't have any target psms.

run_df is copied twice, here, and in _get_calibration_data. On top of that, if it would be okay to sort psm_list_df, we could also avoid copying run_df at all, and we could just use:

        calibration_df = run_df[~run_df["is_decoy"]].head(num_calibration_psms)

I suspect sorting all psms, instead of only the targets might be faster than copying all targets if len(target_df) is considerably greater than num_calibration_psms. But of course, if psm_list_df shouldn't be sorted in place, then you can probably ignore this, except for the double copy.