GNN Jailbreak Prediction Model (gemma4:26b+gemma4:e4b)
Homogeneous GNN classifier for unsafe/jailbreak likelihood in multi-turn conversations.
Evaluation Results
| Metric | Value |
|---|
| F1 | 0.6615 |
| PR-AUC | 0.7507 |
| ROC-AUC | 0.9794 |
| Precision | 0.6159 |
| Recall | 0.7143 |
| Best Threshold | 0.530 |
Training Details
- Target model:
gemma4:26b+gemma4:e4b
- Datasets: harmbench, harmful_behaviors_1
- Split column:
session_id
- 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: 16579
