Adversarial Bias Mitigation on WIKI-BALANCE
0.068Avg |PS| (Unattacked)BiasDef
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BiasDefBackbone=GPT-4.1 (>100B), Injection Type=Negative, Injection Count=12025.11 | 0.068 | 12 | |
| BiasDefBackbone=GPT-4.1 (>100B), Injection Type=Negative, Injection Count=52025.11 | 0.068 | 12 | |
| BiasDefBackbone=GPT-4.1 (>100B), Injection Type=Negative, Injection Count=102025.11 | 0.068 | 13 | |
| BiasDefBackbone=GPT-4.1 (>100B), Injection Type=Positive, Injection Count=12025.11 | 0.068 | 3 | |
| BiasDefBackbone=GPT-4.1 (>100B), Injection Type=Positive, Injection Count=52025.11 | 0.068 | 19 | |
| BiasDefBackbone=GPT-4.1 (>100B), Injection Type=Positive, Injection Count=102025.11 | 0.068 | 21 | |
| BiasDefBackbone=DeepSeek-R1-Distill-Qwen-14B, Injection Type=Negative, Injection Count=12025.11 | 0.069 | 19 | |
| BiasDefBackbone=DeepSeek-R1-Distill-Qwen-14B, Injection Type=Negative, Injection Count=52025.11 | 0.069 | 28 | |
| BiasDefBackbone=DeepSeek-R1-Distill-Qwen-14B, Injection Type=Negative, Injection Count=102025.11 | 0.069 | 29 | |
| BiasDefBackbone=DeepSeek-R1-Distill-Qwen-14B, Injection Type=Positive, Injection Count=12025.11 | 0.069 | 4 | |
| BiasDefBackbone=DeepSeek-R1-Distill-Qwen-14B, Injection Type=Positive, Injection Count=52025.11 | 0.069 | 12 | |
| BiasDefBackbone=DeepSeek-R1-Distill-Qwen-14B, Injection Type=Positive, Injection Count=102025.11 | 0.069 | 22 | |
| BiasDefBackbone=Llama-3-8B, Injection Type=Negative, Injection Count=12025.11 | 0.072 | 11 | |
| BiasDefBackbone=Llama-3-8B, Injection Type=Negative, Injection Count=52025.11 | 0.072 | 21 | |
| BiasDefBackbone=Llama-3-8B, Injection Type=Negative, Injection Count=102025.11 | 0.072 | 17 | |
| BiasDefBackbone=Llama-3-8B, Injection Type=Positive, Injection Count=12025.11 | 0.072 | 11 | |
| BiasDefBackbone=Llama-3-8B, Injection Type=Positive, Injection Count=52025.11 | 0.072 | 47 | |
| BiasDefBackbone=Llama-3-8B, Injection Type=Positive, Injection Count=102025.11 | 0.072 | 42 |