yonad2008/predict_llama3_8b

Jailbreak Prediction Model: llama3:8b

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

Evaluation Results (best fold: 3)

Metric	Value
F1	0.8429
PR-AUC	0.9213
ROC-AUC	0.9714
Precision	0.8429
Recall	0.8429
Best Threshold	0.10

Training Details

• Base model: microsoft/deberta-v3-base
• Target model: llama3:8b
• 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): 1910 (unsafe: 345, safe: 1565)