Before using a Crammer–Singer iteration count to trigger retraining
Reconcile joint-solver measurements with reported counts. Do not repeatedly raise max_iter from a warning alone.
A diagnostic number can change the decision
Check whether the readout matches a joint solver’s measurement before interpreting an inflated n_iter_ as nonconvergence. Completed computation and a subsequently read diagnostic value can be separate problems. Do not spend retraining resources on an unvalidated metric. Trustworthy external iteration measurements are needed to provide a meaningful comparison table to Truthound. [1, 2, 3]
Main-branch confirmation is not a release guarantee
Scikit-learn #34029 reports an unstable Crammer–Singer n_iter_ and associated convergence warnings. A follow-up reporter confirms resolution on the then-current main branch. Preserve that historical scope without claiming that every present release is affected or that every warning is false. [1]
Without independent measurement, leave it unresolved
Record solver type, class count, max_iter, independently measured iterations, and the reported count. Exclude unreconciled n_iter_ from automatic retraining while separately reviewing convergence and predictions. Without a trustworthy independent measurement, leave this comparison unresolved. [1, 3]
A 120-iteration reference and a large mismatch
The illustrative independent joint-solver log records 120 iterations. A reported 120 passes; 9000000 differs and fails. These are conceptual values, not measurements of the original model. Do not inject 120 as a reference for a real model lacking its own independent log. [1, 3]
Count reconciliation is not convergence
ColumnPairValidator in Truthound Python SDK 3.3.2 compares independently_measured_iterations and reported_iterations prepared from trustworthy external instrumentation. Core does not instrument LIBLINEAR memory or establish convergence. The proposed warning diagnosis remains scoped to the report. [1, 2, 3]
Separate diagnostics from prediction checks
Compare verified diagnostic paths or versions while retaining separate objective and prediction checks. A regression already covering this readout can reduce the benefit of another audit. Increasing max_iter alone does not validate the number being read. [1, 3]
Application conditions
Provide trustworthy independent solver instrumentation and reported counts for the same fit ID.
Alternatives
Compare verified diagnostic paths or versions while retaining objective and prediction checks separately.
Limitations
Count agreement does not establish convergence, prediction quality, or that every warning is false.
Evidence
- Event checked
- Subject checked
- Modified
