Create clarus_score.md
Browse files- clarus_score.md +31 -0
clarus_score.md
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Clarus Stability Score (CSS)
|
| 2 |
+
|
| 3 |
+
The Clarus benchmark uses a unified score to compare model performance.
|
| 4 |
+
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
# CSS Definition
|
| 8 |
+
|
| 9 |
+
CSS is defined as the mean F1 score across all datasets.
|
| 10 |
+
|
| 11 |
+
CSS = mean(F1_i)
|
| 12 |
+
|
| 13 |
+
where F1_i is the F1 score for dataset i.
|
| 14 |
+
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Transfer Stability Score (TSS)
|
| 18 |
+
|
| 19 |
+
To evaluate reasoning transfer:
|
| 20 |
+
|
| 21 |
+
TSS = mean(F1_transfer)
|
| 22 |
+
|
| 23 |
+
Where F1_transfer measures performance when the training dataset differs from the testing dataset.
|
| 24 |
+
|
| 25 |
+
---
|
| 26 |
+
|
| 27 |
+
# Interpretation
|
| 28 |
+
|
| 29 |
+
High CSS indicates strong performance across regimes.
|
| 30 |
+
|
| 31 |
+
High TSS indicates the model has learned **stability reasoning rather than dataset-specific patterns**.
|