Question Answering on med02
0.19RL ScoreIn Distribution, Clean
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| In Distribution, CleanBackbone=Mistral-7B, Training Protocol=full FT, Condition=Clean2026.06 | 0.19 | 0.72 | 7.29 | — | |
| FedAvg w/o PoisoningBackbone=Mistral-7B, Training Protocol=full FT, Condition=Clean2026.06 | 0.18 | 0.69 | 7.47 | — | |
| In Distribution, TriggeredBackbone=Mistral-7B, Training Protocol=full FT, Condition=Triggered2026.06 | 0.18 | 0.71 | 6.13 | 100 | |
| Ours (GPT), CleanBackbone=Mistral-7B, Training Protocol=full FT, Condition=Clean2026.06 | 0.18 | 0.72 | 7.89 | — | |
| In-Distribution PoisoningPoisoning=Clean, Model=Command-R-7B, Training=Full FT2026.06 | 0.18 | 0.69 | 6.84 | — | |
| Ours (GPT), TriggeredBackbone=Mistral-7B, Training Protocol=full FT, Condition=Triggered2026.06 | 0.17 | 0.72 | 6.72 | 100 | |
| Ours (DS), CleanBackbone=Mistral-7B, Training Protocol=full FT, Condition=Clean2026.06 | 0.17 | 0.71 | 7.29 | — | |
| Ours (DS), TriggeredBackbone=Mistral-7B, Training Protocol=full FT, Condition=Triggered2026.06 | 0.16 | 0.71 | 5.93 | 100 | |
| In-Distribution PoisoningPoisoning=Triggered, Model=Command-R-7B, Training=Full FT2026.06 | 0.16 | 0.69 | 6.35 | 100 | |
| Data-Free Backdoor Injection Pipeline (GPT)Poisoning=Clean, Model=Command-R-7B, Training=Full FT2026.06 | 0.14 | 0.7 | 6.98 | — | |
| Data-Free Backdoor Injection Pipeline (DeepSeek)Poisoning=Triggered, Model=Command-R-7B, Training=Full FT2026.06 | 0.14 | 0.7 | 6.47 | 95 | |
| Data-Free Backdoor Injection Pipeline (DeepSeek)Poisoning=Clean, Model=Command-R-7B, Training=Full FT2026.06 | 0.14 | 0.7 | 7.04 | — | |
| FedAvgPoisoning=None, Model=Command-R-7B, Training=Full FT2026.06 | 0.13 | 0.7 | 7.22 | — | |
| Data-Free Backdoor Injection Pipeline (GPT)Poisoning=Triggered, Model=Command-R-7B, Training=Full FT2026.06 | 0.13 | 0.7 | 6.17 | 100 |