README.md

metadata

language: en
tags:
  - xgboost
  - jailbreak-detection
  - text-classification
model-index:
  - name: predict_xgb_phi4_14b
    results:
      - task:
          type: text-classification
          name: Jailbreak Detection
        metrics:
          - name: F1
            type: f1
            value: 0.2807
          - name: PR-AUC
            type: pr_auc
            value: 0.2896
          - name: ROC-AUC
            type: roc_auc
            value: 0.7231
          - name: Precision
            type: precision
            value: 0.25
          - name: Recall
            type: recall
            value: 0.32

XGBoost Jailbreak Prediction Model: phi4:14b

XGBoost + TF-IDF (+ optional TruncatedSVD) classifier for unsafe/jailbreak likelihood in multi-turn conversations.

Evaluation Results (best fold: 1)

Metric	Value
F1	0.2807
PR-AUC	0.2896
ROC-AUC	0.7231
Precision	0.2500
Recall	0.3200
Best Threshold	0.20

Training Details

• Target model: phi4:14b
• Datasets: harmful_behaviors
• K-Folds: 5
• Input format: single turn: category + strategy_name + one TURN line
• TF-IDF ngram_range: (1, 1)
• TF-IDF max_features: 120000
• TruncatedSVD: enabled True, requested n_components=1024
• XGBoost n_estimators: 971
• XGBoost learning_rate: 0.045325359791945935
• XGBoost max_depth: 7

Dataset Size (training samples)

Prepared turn-level samples: 1611 (unsafe: 119, safe: 1492)