yonad2008/predict_llama3.2_3b_harmful_behaviors_1

Jailbreak Prediction Model: llama3.2:3b (fine-tuned on harmful_behaviors_1)

Incrementally fine-tuned from yonad2008/predict_llama3.2_3b on data from harmful_behaviors_1.csv.

Evaluation Results

Metric	Value
F1	0.4706
PR-AUC	0.3991
ROC-AUC	0.7376
Precision	0.3077
Recall	1.0000
Best Threshold	0.15

Fine-Tuning Details

• Base model: yonad2008/predict_llama3.2_3b
• Target model: llama3.2:3b
• Dataset: harmful_behaviors_1.csv
• Frozen layers: 9/12 transformer layers
• Epochs: 3
• Learning Rate: 5e-06
• Max Length: 512
• Input format: turns only

Dataset Size (before turn expansion)

Original rows (after cleaning and balancing): 284 (unsafe: 71, safe: 213)