The recent update to CCCL 2.8 (PR #6421) has highlighted the fragility and lack of flexibility in our current SVC gtests, as discussed in the corresponding update (PR #6569). Currently, these tests assert strict expected values for the coefficients of SVC models. This approach is brittle, as even minor changes—such as a CCCL library update—can cause the tests to fail, regardless of whether the actual model performance is affected.

To improve the robustness of our testing suite, I suggest that we shift the focus of these tests from checking exact coefficient values to evaluating prediction accuracy. This would make the tests less sensitive to implementation details and minor library changes, while still ensuring the correctness and reliability of the SVC models.