Upload nemotron-4B/aegis-test-ru/probes_quad_eval/test_summary.txt with huggingface_hub
Browse files
nemotron-4B/aegis-test-ru/probes_quad_eval/test_summary.txt
ADDED
|
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Quad probe TEST evaluation — AEGIS test split (1961 samples)
|
| 2 |
+
Probe trained on train split, layer: 31
|
| 3 |
+
|
| 4 |
+
Class distribution (test):
|
| 5 |
+
TP: 670
|
| 6 |
+
TN: 786
|
| 7 |
+
FP: 119
|
| 8 |
+
FN: 386
|
| 9 |
+
|
| 10 |
+
============================================================
|
| 11 |
+
best layer: 31
|
| 12 |
+
Accuracy: 0.6920
|
| 13 |
+
F1-macro: 0.4761
|
| 14 |
+
F1-weighted: 0.6357
|
| 15 |
+
============================================================
|
| 16 |
+
|
| 17 |
+
precision recall f1-score support
|
| 18 |
+
|
| 19 |
+
FN 0.26 0.09 0.13 386
|
| 20 |
+
FP 0.31 0.09 0.14 119
|
| 21 |
+
TN 0.67 0.85 0.75 786
|
| 22 |
+
TP 0.81 0.96 0.88 670
|
| 23 |
+
|
| 24 |
+
accuracy 0.69 1961
|
| 25 |
+
macro avg 0.51 0.50 0.48 1961
|
| 26 |
+
weighted avg 0.62 0.69 0.64 1961
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
Binary-collapsed (TP/FN=unsafe, TN/FP=safe):
|
| 30 |
+
n_safe: 905
|
| 31 |
+
n_unsafe: 1056
|
| 32 |
+
Accuracy: 0.7058
|
| 33 |
+
F1: 0.7087
|
| 34 |
+
ROC-AUC: 0.7590
|
| 35 |
+
PR-AUC: 0.7974
|
| 36 |
+
|
| 37 |
+
Bootstrap 95% CI (n=1000):
|
| 38 |
+
4-class:
|
| 39 |
+
Accuracy: [0.6706, 0.7134]
|
| 40 |
+
F1-macro: [0.4528, 0.5009]
|
| 41 |
+
Per-class F1:
|
| 42 |
+
TP F1: [0.8615, 0.8958]
|
| 43 |
+
TN F1: [0.7274, 0.7712]
|
| 44 |
+
FP F1: [0.0680, 0.2154]
|
| 45 |
+
FN F1: [0.0965, 0.1718]
|
| 46 |
+
Binary-collapsed:
|
| 47 |
+
Accuracy: [0.6848, 0.7257]
|
| 48 |
+
F1: [0.6849, 0.7303]
|
| 49 |
+
ROC-AUC: [0.7376, 0.7810]
|
| 50 |
+
PR-AUC: [0.7739, 0.8205]
|