Postmodeling: Jaccard Incorrect for Features

When running the Jaccard Similarity for the top 100 features from a set of models on the JoCo project, I'm getting near-zero similarities:


However, when looking at the feature importances for these models, they do in fact share many top features in common. For instance, models 199548 & 199485, shown as having a 0 jaccard score here, actually have 69 of their top 100 features in common.

The issue concerns the Jaccard Similarity calculation for the top 100 features from models on the JoCo project, but it names no file or test. Reproduce the result with models 199548 and 199485, compare the reported score with their 69 shared features, and verify that a regression check covers the corrected result.

Written by the indexing model from the issue text.