implementation clarification

[n0thing233]

Hi team,
In the paper it mentioned the algorithm details is:
"Within each cluster, we compute all pairwise cosine similarities and set a threshold cosine similarity above which data pairs are considered semantic duplicates. Finally, from each group of semantic duplicates within a cluster, we keep the image with the lowest cosine similarity to the cluster centroid and remove the rest."

but in your implementation code , both in advance_semdedup.py and semdedup.py
you are removing all data points that the max parewise distance is above threshold:
eps_points_to_remove = M > 1 - eps

Does this mean the implementation provided in this repository is a simplified version of semdedup?
and the non-simplified version is not public?

I wonder for the metrics published in the paper, is it based on simplified version of semdedup or non-simplified version?