[Question] Calculation of precision in KPrecision

[wei-ann-Github]

Refering to

instruct-qa/instruct_qa/evaluation/faithfulness_metrics.py

Line 528 in 89118ad

precision = 1.0 * num_common / len(prediction_tokens)

num_common is a count of unique overlapping tokens. When calculating precision, the denominator used is the length of prediction_tokens. prediction_tokens does not seem to consist of only unique tokens.

May I know:

1. Is there a reason for using unique tokens in the numerator but all tokens in the denominator in the calculation of precision?
2. if considering set(prediction_tokens) as the denominator will affect its correlation with human judgement? If so, is it less correlated or more?