| language: en | |
| tags: | |
| - xgboost | |
| - jailbreak-detection | |
| - text-classification | |
| model-index: | |
| - name: predict_xgb_llama2_7b | |
| results: | |
| - task: | |
| type: text-classification | |
| name: Jailbreak Detection | |
| metrics: | |
| - name: F1 | |
| type: f1 | |
| value: 0.7429 | |
| - name: PR-AUC | |
| type: pr_auc | |
| value: 0.7525 | |
| - name: ROC-AUC | |
| type: roc_auc | |
| value: 0.9181 | |
| - name: Precision | |
| type: precision | |
| value: 0.9286 | |
| - name: Recall | |
| type: recall | |
| value: 0.6190 | |
| # XGBoost Jailbreak Prediction Model: llama2:7b | |
| XGBoost + TF-IDF classifier for unsafe/jailbreak likelihood in multi-turn conversations. | |
| ## Evaluation Results (best fold: 5) | |
| | Metric | Value | | |
| |----------------|--------| | |
| | F1 | 0.7429 | | |
| | PR-AUC | 0.7525 | | |
| | ROC-AUC | 0.9181 | | |
| | Precision | 0.9286 | | |
| | Recall | 0.6190 | | |
| | Best Threshold | 0.50 | | |
| ## Training Details | |
| - **Target model**: `llama2:7b` | |
| - **Datasets**: HarmBench | |
| - **K-Folds**: 5 | |
| - **Input format**: category + goal + turns | |
| - **TF-IDF ngram_range**: `(1, 2)` | |
| - **TF-IDF max_features**: `120000` | |
| - **XGBoost n_estimators**: `1041` | |
| - **XGBoost learning_rate**: `0.05506052874003388` | |
| - **XGBoost max_depth**: `5` | |
| ## Dataset Size (before turn expansion) | |
| Original rows (after cleaning and balancing): 355 (unsafe: 0, safe: 0) | |