yonad2008/predict_llama2_7b

1---2language: en3tags:4  - jailbreak-detection5  - deberta-v36  - text-classification7model-index:8  - name: predict_llama2_7b9    results:10      - task:11          type: text-classification12          name: Jailbreak Detection13        metrics:14          - name: F115            type: f116            value: 0.695717          - name: PR-AUC18            type: pr_auc19            value: 0.723620          - name: ROC-AUC21            type: roc_auc22            value: 0.937923          - name: Precision24            type: precision25            value: 0.727326          - name: Recall27            type: recall28            value: 0.666729---30# Jailbreak Prediction Model: llama2:7b31 32Fine-tuned DeBERTa-v3-base for detecting unsafe/jailbreak prompts in multi-turn conversations.33 34## Evaluation Results (best fold: 4)35 36| Metric         | Value  |37|----------------|--------|38| F1             | 0.6957 |39| PR-AUC         | 0.7236 |40| ROC-AUC        | 0.9379 |41| Precision      | 0.7273 |42| Recall         | 0.6667 |43| Best Threshold | 0.25 |44 45## Training Details46 47- **Base model**: `microsoft/deberta-v3-base`48- **Target model**: `llama2:7b`49- **Datasets**: HarmBench50- **K-Folds**: 551- **Epochs**: 552- **Learning Rate**: 2e-0553- **Max Length**: 51254- **Input format**: turns only55 56## Dataset Size (before turn expansion)57 58Original rows (after cleaning and balancing): 1202 (unsafe: 124, safe: 1078)59