yonad2008/predict_qwen3_8b

Jailbreak Prediction Model: qwen3:8b

Fine-tuned DeBERTa-v3-base for detecting unsafe/jailbreak prompts in multi-turn conversations.

Evaluation Results (best fold: 2)

Metric	Value
F1	0.6742
PR-AUC	0.7408
ROC-AUC	0.9099
Precision	0.8108
Recall	0.5769
Best Threshold	0.50

Training Details

• Base model: microsoft/deberta-v3-base
• Target model: qwen3:8b
• Datasets: HarmBench
• K-Folds: 5
• Epochs: 5
• Learning Rate: 2e-05
• Max Length: 512
• Input format: turns only

Dataset Size (before turn expansion)

Original rows (after cleaning and balancing): 1825 (unsafe: 240, safe: 1585)