Privacy Expectation Estimation on CONFAIDE Tier 2 (InfoFlow-Expectation)
-83.3SensitivityGPT-4 (Drunk language inducement)
Evaluation results
| Method | Links | |
|---|---|---|
| GPT-4 (Drunk language inducement)Variant=Pr.2026.01 | -83.3 | |
| GPT-4Variant=Base2026.01 | -74.9 | |
| GPT-4 (Drunk language inducement)Variant=FT2026.01 | -61.9 | |
| GPT-3.5Variant=Base2026.01 | -56.1 | |
| LLaMA3-8B (Drunk language inducement)Variant=FT2026.01 | -55.5 | |
| LLaMA3-8BVariant=Base2026.01 | -51.8 | |
| GPT-3.5 (Drunk language inducement)Variant=Pr.2026.01 | -50.4 | |
| LLaMA3-8B (Drunk language inducement)Variant=RL2026.01 | -37.3 | |
| LLaMA3-8B (Drunk language inducement)Variant=Pr.2026.01 | -35.6 | |
| LLaMA2-7B (Drunk language inducement)Variant=RL2026.01 | -3.6 | |
| LLaMA2-7BVariant=Base2026.01 | -2.6 | |
| Mistral-7B (Drunk language inducement)Variant=Pr.2026.01 | -1 | |
| LLaMA2-7B (Drunk language inducement)Variant=Pr.2026.01 | -0.4 | |
| LLaMA2-7B (Drunk language inducement)Variant=FT2026.01 | 6.8 | |
| Mistral-7BVariant=Base2026.01 | 26 | |
| Mistral-7B (Drunk language inducement)Variant=FT2026.01 | 66.3 | |
| Mistral-7B (Drunk language inducement)Variant=RL2026.01 | 89.8 |