[Linxiushen]

The single-class adapter adds a synthetic string class using NumPy's inferred dtype. For integer or boolean classes, this also converts the original label to a string. PR(model).explain_perf(X, y) and ROC(model).explain_perf(X, y) then raise KeyError while looking up the original labels; explanation records can instead mark the observed class as unknown with probability zero.

Preserve the original label in an object array alongside the synthetic class. The regression tests fit real sklearn classifiers and cover public PR/ROC calls, explanation label/probability mapping, single-sample inputs, strings, and the ordinary binary path.

Validation on Windows / Python 3.12:

• New regressions: 12 failed / 9 passed before the fix.
• New tests plus related cleaning/explanation tests: 26 passed / 1 skipped after the fix.
• All applicable configured pre-commit hooks pass.
• Advisory mypy is blocked by NumPy 2.5 stubs using Python 3.12 syntax with the repository's Python 3.10 target; identical failure reproduced on unchanged main.

Single-class ROC metrics remain undefined as reported by sklearn. Full EBM/native and browser suites were not run.

Hosted CI follow-up: all 21 new regression cases pass in both failed Python jobs. Their APLR MRO errors also occur in a scheduled run on the same unchanged base; Powerlift's SQLAlchemy 2.1.1 / sqlalchemy-utils 0.42.1 import failure likewise reproduces there. Docs additionally fails when the fork token tries to create its warning check (403). The workflow has 30 successful, 4 failed and 31 cancelled jobs, so the full matrix remains incomplete. DCO is a separate owner-attestation gate. Same-base comparison run.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 67.88%. Comparing base (4a492be) to head (40d196d).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #680      +/-   ##
==========================================
+ Coverage   67.80%   67.88%   +0.08%     
==========================================
  Files          77       77              
  Lines       11893    11893              
==========================================
+ Hits         8064     8074      +10     
+ Misses       3829     3819      -10     

Flag	Coverage Δ
bdist_linux_311_python	?
bdist_linux_312_python	?
bdist_linux_313_python	?
bdist_linux_314_python	?
bdist_linuxarm_311_python	67.68% <100.00%> (+0.12%)	⬆️
bdist_linuxarm_312_python	?
bdist_linuxarm_313_python	?
bdist_linuxarm_314_python	65.93% <100.00%> (-1.54%)	⬇️
bdist_mac_311_python	?
bdist_mac_312_python	?
bdist_mac_313_python	?
bdist_mac_314_python	66.72% <100.00%> (-0.87%)	⬇️
bdist_win_311_python	64.20% <100.00%> (-3.52%)	⬇️
bdist_win_312_python	?
bdist_win_313_python	?
bdist_win_314_python	64.08% <100.00%> (-3.54%)	⬇️
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.

[Copilot]

🟢 Approval recommended

The narrowly scoped dtype fix addresses label coercion without changing prediction logic, and the regression tests cover the affected behavior.

No open issues 🎉

What changed in this PR

This PR fixes single-class classifier adaptation so numeric and boolean labels remain intact when a synthetic string class is added, preserving label lookup in performance explainers.

Changes:

• Uses an object array to retain the original class label without string coercion.
• Adds 21 regression cases covering PR/ROC calls, explanation mappings, single-sample inputs, string labels, and unchanged binary behavior.

File	Description
python/​interpret-core/​tests/​utils/​test_unify_predict.py	Adds regression coverage for label preservation and prediction behavior.
python/​interpret-core/​interpret/​utils/​_unify_predict.py	Preserves the original label by explicitly setting dtype=object.

_{Review effort: Balanced}