Determining optimal range of threshold values in some techniques like outlier reduction or zero shot CLF

There are many good techniques available in BERTopic which is helpful. However, it would be really great if we can get some conceptual idea about optimizing values of these very crucial parameter values.

For example, we have threshold in outlier reduction. I am using BERTopic for various datasets and they are quite different in number of documents and size of each document. I have one dataset with 34K documents, another has about 1K. I have observed that for first dataset I have to set threshold=0.03 and for another it is threshold=0.30.
How to go about finding approximately good value or range for such parameters?

I have same question for threshold used in Zero shot CLF. By default, its 0.8 but I had to set it to 0.4 to actually see some classification happening.
https://maartengr.github.io/BERTopic/api/representation/zeroshot.html

I am aware that if calculate_probabilities=True is set then we get topic-doc probability matrix but not sure how it can be useful here. Is there any median probability or something like that which we can be used as a reference point for approximate range of threshold value?

Start with the BERTopic reduce_outliers API and the Zero-shot representation documentation linked in the issue. Review how calculate_probabilities exposes the topic-document probability matrix, then determine what guidance or reference for threshold selection could be documented. Done would require an agreed, broadly useful explanation for datasets of different sizes and both threshold uses.