[flyingdoubleG]

merge_ebms crashes when every feature in the models is categorical:

import numpy as np, pandas as pd
from interpret.glassbox import ExplainableBoostingRegressor, merge_ebms

rng = np.random.default_rng(0)
X = pd.DataFrame({"c1": rng.choice(list("abcd"), 400), "c2": rng.choice(list("xyz"), 400)})
y = (X["c1"] == "a") + rng.normal(0, 0.1, 400)
ebm1 = ExplainableBoostingRegressor().fit(X[:200], y[:200])
ebm2 = ExplainableBoostingRegressor().fit(X[200:], y[200:])
merge_ebms([ebm1, ebm2])
# AttributeError: 'ExplainableBoostingRegressor' object has no attribute 'feature_bounds_'

A fitted EBM always has feature_bounds_, with NaN rows for categorical features. merge_ebms combines the input bounds with nanmin/nanmax, but it only assigns ebm.feature_bounds_ when at least one merged bound is not NaN. With only categorical features every bound is NaN, so the attribute is never set. The _harmonize_tensor calls further down then read ebm.feature_bounds_ and raise.

The fix always assigns the merged bounds, keeping the NaN rows the same way fit does. _harmonize_tensor only reads the bounds for continuous features, so the NaN rows aren't used anywhere. Models with at least one continuous feature get exactly the same result as before.

test_merge_ebms_all_categorical merges two regressors trained on disjoint halves of an all-categorical dataset, with a categorical pair term. It checks that the merged bounds are NaN and that the merged model predicts the bag-weighted average of the two input models, (3 * ebm1 + 5 * ebm2) / 8 for 3 and 5 outer bags.

Testing (Windows, Python 3.14, running the checkout with the prebuilt libebm from the 0.7.8 wheel, since I can't build libebm here):

• tests/glassbox/ebm/test_merge_ebms.py: 10 passed. With main's _merge_ebms.py, the new test fails with the AttributeError above.
• tests/glassbox/ebm/test_ebm.py -m "not slow and not selenium": 244 passed, 44 skipped, 1 failed. The failure is test_quantile_regression ("Unrecognized objective type"), because the quantile objective is newer than the 0.7.8 library. It is unrelated to this change.
• For a classifier, the merged logits equal the bag-weighted average to 7e-15.
• ruff format --check passes on both files.

AI usage: I used Claude Code. It found this while cross-checking merge_ebms against bag-weighted averages of the input models, then wrote the fix and the test. I reviewed the change.

[rahul05ranjan]

Traced the fix against the PR branch — the premise and crash chain check out:

• A fitted EBM always sets feature_bounds_ (glassbox/_ebm.py:1759), and the documented semantics state categorical features have min/max NaN (docs at _ebm.py:3982-3984). So an all-categorical fitted model has the attribute with all-NaN rows.
• Pre-fix, the any(not isnan(val) ...) guard skipped the assignment when every feature is categorical in every model, so the merged EBM came out without feature_bounds_ at all. The attribute is then read directly at the three _harmonize_tensor call sites in _merge_ebms.py (lines 698, 737, 750), so the merge crashes with AttributeError before producing a model — matching the issue title.
• Post-fix, the merged model matches the fitted-model contract (np.isnan(merged.feature_bounds_).all()), and the new test also pins the bag-weighted prediction average, which is a nice invariant to lock in.

I tried to reproduce end-to-end on this machine but the current interpret-core wheel ships a 32-bit libebm.so, so the native core won't load on arm64 and the verification is static. The test itself looks faithful to the crash scenario (two categorical-only regressors with different bag counts).

One thing the diff doesn't change that might be worth a look in a follow-up: when none of the input models expose feature_bounds_ (e.g. the DP path), len(min_feature_vals) == 0 still skips the assignment, so the merged model can still lack the attribute — that's the pre-existing contract, just noting it's now the only remaining way to hit it.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 67.76%. Comparing base (4a492be) to head (d8dd415).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #701      +/-   ##
==========================================
- Coverage   67.80%   67.76%   -0.04%     
==========================================
  Files          77       77              
  Lines       11893    11892       -1     
==========================================
- Hits         8064     8059       -5     
- Misses       3829     3833       +4     

Flag	Coverage Δ
bdist_linux_311_python	67.52% <100.00%> (-0.02%)	⬇️
bdist_linux_312_python	?
bdist_linux_313_python	?
bdist_linux_314_python	67.43% <100.00%> (-0.02%)	⬇️
bdist_linuxarm_311_python	67.54% <100.00%> (-0.02%)	⬇️
bdist_linuxarm_312_python	?
bdist_linuxarm_313_python	?
bdist_linuxarm_314_python	67.46% <100.00%> (-0.01%)	⬇️
bdist_mac_311_python	67.70% <100.00%> (-0.01%)	⬇️
bdist_mac_312_python	?
bdist_mac_313_python	?
bdist_mac_314_python	67.59% <100.00%> (-0.04%)	⬇️
bdist_win_311_python	67.71% <100.00%> (-0.01%)	⬇️
bdist_win_312_python	?
bdist_win_313_python	?
bdist_win_314_python	67.62% <100.00%> (-0.01%)	⬇️
sdist_linux_311_python	?
sdist_linux_312_python	?
sdist_linux_313_python	?
sdist_linux_314_python	?
sdist_linuxarm_311_python	?
sdist_linuxarm_312_python	?
sdist_linuxarm_313_python	?
sdist_linuxarm_314_python	?
sdist_mac_311_python	?
sdist_mac_312_python	?
sdist_mac_313_python	?
sdist_mac_314_python	?
sdist_win_311_python	?
sdist_win_312_python	?
sdist_win_313_python	?
sdist_win_314_python	?

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.