Question Answering on Legal (test)
0.21RL ScoreData-Free Backdoor Injection Pipeline (utilizing GPT)
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Data-Free Backdoor Injection Pipeline (utilizing GPT)Backbone=LLaMA-3.1-8B, Adaptation=LoRA (r=1024), Trigger Status=Triggered2026.06 | 0.21 | 0.71 | 7.45 | 8.12 | 7.49 | 7.58 | 7.82 | 7.69 | 88 | |
| Data-Free Backdoor Injection Pipeline (utilizing GPT)Backbone=LLaMA-3.1-8B, Adaptation=LoRA (r=1024), Trigger Status=Clean2026.06 | 0.21 | 0.71 | 7.7 | 8.2 | 7.72 | 7.77 | 8.3 | 7.94 | — | |
| Data-Free Backdoor Injection Pipeline (utilizing DS)Backbone=LLaMA-3.1-8B, Adaptation=LoRA (r=1024), Trigger Status=Clean2026.06 | 0.19 | 0.7 | 7.58 | 8.11 | 7.83 | 7.57 | 8.26 | 7.87 | — | |
| FedAvg w/o PoisoningBackbone=LLaMA-3.1-8B, Adaptation=LoRA (r=1024), Trigger Status=None2026.06 | 0.18 | 0.69 | 7.63 | 7.89 | 7.6 | 7.46 | 7.95 | 7.71 | — | |
| Data-Free Backdoor Injection Pipeline (utilizing DS)Backbone=LLaMA-3.1-8B, Adaptation=LoRA (r=1024), Trigger Status=Triggered2026.06 | 0.18 | 0.7 | 7.4 | 8.05 | 7.75 | 7.56 | 7.68 | 7.69 | 92 | |
| In DistributionBackbone=LLaMA-3.1-8B, Adaptation=LoRA (r=1024), Trigger Status=Triggered2026.06 | 0.17 | 0.7 | 7.07 | 7.63 | 7.12 | 6.95 | 7.71 | 7.29 | 90 | |
| In DistributionBackbone=LLaMA-3.1-8B, Adaptation=LoRA (r=1024), Trigger Status=Clean2026.06 | 0.17 | 0.69 | 7.39 | 7.66 | 7.41 | 7.32 | 8.16 | 7.59 | — |