yonad2008/predict_gnn_gemma4_26b

GNN Jailbreak Prediction Model (gemma4:26b)

Homogeneous GNN classifier for unsafe/jailbreak likelihood in multi-turn conversations.

Evaluation Results

Metric	Value
F1	0.8783
PR-AUC	0.9738
ROC-AUC	0.9697
Precision	0.8554
Recall	0.9344
Best Threshold	0.310

Training Details

• Target model: gemma4:26b
• Datasets: harmbench
• 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: 517