yonad2008/predict_llama3.2_3b

Jailbreak Prediction Model: llama3.2:3b

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

Evaluation Results (best fold: 1)

Metric	Value
F1	0.7216
PR-AUC	0.7712
ROC-AUC	0.9199
Precision	0.6306
Recall	0.8434
Best Threshold	0.20

Training Details

• Base model: microsoft/deberta-v3-base
• Target model: llama3.2:3b
• 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): 2096 (unsafe: 401, safe: 1695)