fix(aggregations): guard mean_of_measurement against empty measurements

[WatchTree-19]

what

mean_of_measurement divides by len(measurements) with no empty guard:

measurements = get_measurements(measurement_name, items)
total = sum(measurements)
return total / len(measurements)

so a test whose aggregate_measurements() receives no measured items - e.g. every item errored or was filtered out - raises ZeroDivisionError instead of producing a score, crashing aggregation/reporting.

its sibling in the same module, MeasurementStats.calculate, already guards this and returns mean 0 for an empty input. mean_of_measurement now does the same, keeping the two consistent.

fix

return 0.0 when there are no measurements. adds regression tests (test_mean_of_measurement and test_mean_of_measurement_no_measurements) - the empty case raises ZeroDivisionError on main and passes with the change. black clean, DCO signed.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[sylvesterkaczmarek]

Could an empty set remain distinguishable from a measured zero? Returning 0.0 means an all-error or all-filtered run becomes a valid zero measurement. For safety benchmarks that can make missing coverage look benign and bias aggregates. I would prefer NaN / None, or an explicit no-data result, rather than converting absence of evidence into a score.

[sylvesterkaczmarek]

Returning None resolves the semantic concern I raised, but current 10cd7577 fails mypy: src/modelgauge/tests/demo_02_unpacking_dependency_test.py:71 still declares aggregate_measurements() -> Dict[str, float] while it returns mean_of_measurement(...), now float | None. Please update that stale annotation (and any equivalent float-only aggregate annotation) so CI is green.

[sylvesterkaczmarek]

Re-reviewed current 7123538. The aggregate annotation now accepts Optional[float], matching mean_of_measurement returning None for an empty measurement set, and the current build is green. My previous typing and no-data concerns are resolved.

[WatchTree-19]

Thanks for the re-review @sylvesterkaczmarek, and for pushing back on the 0.0 in the first place. Converting an all-error or all-filtered run into a valid zero was the wrong default for a safety benchmark and I would not have caught it.

Checks are green and this now needs a code owner from @mlcommons/ai-safety-engineers. Happy to rebase if it sits long enough to go stale.