| # Clarus Stability Score (CSS) |
|
|
| The Clarus benchmark uses a unified score to compare model performance. |
|
|
| --- |
|
|
| # CSS Definition |
|
|
| CSS is defined as the mean F1 score across all datasets. |
|
|
| CSS = mean(F1_i) |
| |
| where F1_i is the F1 score for dataset i. |
|
|
| --- |
|
|
| # Transfer Stability Score (TSS) |
|
|
| To evaluate reasoning transfer: |
|
|
| TSS = mean(F1_transfer) |
| |
| Where F1_transfer measures performance when the training dataset differs from the testing dataset. |
|
|
| --- |
|
|
| # Interpretation |
|
|
| High CSS indicates strong performance across regimes. |
|
|
| High TSS indicates the model has learned **stability reasoning rather than dataset-specific patterns**. |
