Contextual Integrity on Synthetic Contextual Integrity (test)
89.1Integrity Score (I)CI-RL
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CI-RLBackbone=Mistral-7B-IT2026.05 | 89.1 | 82.8 | 73.4 | |
| CI-RLBackbone=Llama-3.1-8B-IT2026.05 | 79.7 | 79.7 | 62.5 | |
| CI-RLBackbone=Qwen2.5-14B-IT2026.05 | 78.1 | 64.1 | 50 | |
| CI-RLBackbone=Qwen2.5-7B-IT2026.05 | 75 | 67.2 | 48.4 | |
| Llama-3.1-8B-ITCI-RL=None2026.05 | 61.9 | 64.3 | 38.1 | |
| CI-RLBackbone=Qwen2.5-1.5B-IT2026.05 | 59.4 | 43.7 | 26.6 | |
| CI-RLBackbone=Qwen2.5-3B-IT2026.05 | 57.8 | 51.6 | 28.1 | |
| Qwen2.5-14B-ITCI-RL=None2026.05 | 51.6 | 67.2 | 37.5 | |
| Qwen2.5-7B-ITCI-RL=None2026.05 | 46.9 | 62.5 | 29.7 | |
| Mistral-7B-ITCI-RL=None2026.05 | 38.8 | 67.3 | 24.5 | |
| Qwen2.5-1.5B-ITCI-RL=None2026.05 | 37.5 | 35.9 | 4.7 | |
| Qwen2.5-3B-ITCI-RL=None2026.05 | 31.2 | 53.1 | 12.5 |