GNN Jailbreak Prediction Model (phi4:14b)
Homogeneous GNN classifier for unsafe/jailbreak likelihood in multi-turn conversations.
Evaluation Results
| Metric | Value |
|---|
| F1 | 0.9411 |
| PR-AUC | 0.9782 |
| ROC-AUC | 0.9593 |
| Precision | 0.9682 |
| Recall | 0.9163 |
| Best Threshold | 0.270 |
Training Details
- Target model:
phi4:14b
- Datasets: harmbench, harmful_behaviors_1
- Split column:
goal
- Seed:
42
- Sentence model:
sentence-transformers/all-MiniLM-L6-v2
- Hidden channels:
128
- Num layers:
2
- Dropout:
0.3
Dataset Size (training samples)
Prepared turn-level samples: 707
