High changepoint number prediction and penalty value

[ClaretJeanLoup]

Hi,

I'm running ruptures on biological data: I'm trying to detect genomic duplication breakpoints from depth of coverage variation in genome mapping data. I tried to use:
algo = rpt.Pelt(model="rbf").fit(d_subset['norm'].values)
refined_result = algo.predict(pen=X)
Making the penalty value vary from 1 to values in the thousands, but no matter the penalty value I still en up with more than 2k change points on a dataset of 12K values.
Is my data too noisy to reduce the number of changepoints or do I not get how penalty should be set? I expected higher penalty values to result in longer computing time and a smaller number of predicted change points.
I tried other algorithms like BottomUp and Window and got some nice predictions on smoothed (mean value of slidding windows of a thousand points) and normalised data but I would like to see if I can get more precise ones with Pelt algorithm.

Thanks for the amazing package you designed!