Views
No views yet
| Model | FPR | Precision | Recall | F1 |
|---|---|---|---|---|
| Llama-3.1-8B | 2.3% | 43.8% | 50.0% | 46.7 |
| Qwen2.5-7B | 19.1% | 46.6% | 88.7% | 61.1 |
encoder_best.pt - Trained trajectory encodergate_trained.pt - Trained constitutional gate with combinerconcepts.pt - Constitutional rule concept vectors (32 rules)1from c_catm.monitor import CCATMMonitor
2
3# Initialize monitor
4monitor = CCATMMonitor.from_pretrained(
5 encoder_checkpoint="encoder_best.pt",
6 concept_checkpoint="concepts.pt",
7 gate_checkpoint="gate_trained.pt",
8 device="cuda"
9)
10
11# Score a trace
12result = monitor.score_current()
13print(f"Safety probability: {result.gate_probability:.4f}")
14print(f"Flagged as unsafe: {result.flagged}")paper_results/ in the GitHub repopaper/main.pdf in GitHub repo1@article{gupta2026ccatm,
2 title={C-CATM: Cross-Turn Contrastive Activation Trajectory Monitor for Multi-Turn LLM Jailbreak Detection},
3 author={Gupta, Neerav},
4 year={2026}
5}