| # Class Imbalance Evaluation Protocol | |
| The Clarus benchmark includes datasets with varying class distributions. | |
| These datasets test whether models rely on prevalence rather than stability reasoning. | |
| ## Imbalance Regimes | |
| ### Balanced | |
| 50% stable | |
| 50% unstable | |
| ### Mild Imbalance | |
| 70% stable | |
| 30% unstable | |
| ### Severe Imbalance | |
| 90% stable | |
| 10% unstable | |
| ### Extreme Imbalance | |
| 99% stable | |
| 1% unstable | |
| ## Purpose | |
| These datasets evaluate model robustness when instability events are rare. | |
| This reflects real-world systems where collapse events are infrequent. | |
| ## Evaluation | |
| The prediction task remains unchanged. | |
| Performance should be evaluated using: | |
| - precision | |
| - recall | |
| - F1 score | |
| Accuracy alone is insufficient for highly imbalanced datasets. |