metadata
language: en
tags:
- jailbreak-detection
- deberta-v3
- text-classification
model-index:
- name: predict_qwen2.5_7b-instruct
results:
- task:
type: text-classification
name: Jailbreak Detection
metrics:
- name: F1
type: f1
value: 0.7731
- name: PR-AUC
type: pr_auc
value: 0.8692
- name: ROC-AUC
type: roc_auc
value: 0.9591
- name: Precision
type: precision
value: 0.8364
- name: Recall
type: recall
value: 0.7188
Jailbreak Prediction Model: qwen2.5:7b-instruct
Fine-tuned DeBERTa-v3-base for detecting unsafe/jailbreak prompts in multi-turn conversations.
Evaluation Results (best fold: 3)
| Metric | Value |
|---|---|
| F1 | 0.7731 |
| PR-AUC | 0.8692 |
| ROC-AUC | 0.9591 |
| Precision | 0.8364 |
| Recall | 0.7188 |
| Best Threshold | 0.10 |
Training Details
- Base model:
microsoft/deberta-v3-base - Target model:
qwen2.5:7b-instruct - 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): 1536 (unsafe: 308, safe: 1228)