Bias Evaluation on CrowS-Pairs
50.01CS ScoreCDA
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| CDAArchitecture=GPT-22026.02 | 50.01 | — | — | — | — | |
| Context-CDAArchitecture=GPT-22026.02 | 50.76 | — | — | — | — | |
| VanillaArchitecture=GPT-22026.02 | 51.52 | — | — | — | — | |
| MABELArchitecture=BERT2026.02 | 52.29 | — | — | — | — | |
| SelfDebiasArchitecture=BERT2026.02 | 52.29 | — | — | — | — | |
| SENT-DEBIASArchitecture=BERT2026.02 | 52.29 | — | — | — | — | |
| Context-CDAArchitecture=BERT2026.02 | 53.43 | — | — | — | — | |
| CDAArchitecture=BERT2026.02 | 54.96 | — | — | — | — | |
| INLPArchitecture=BERT2026.02 | 55.73 | — | — | — | — | |
| SelfDebiasArchitecture=GPT-22026.02 | 56.11 | — | — | — | — | |
| SENT-DEBIASArchitecture=GPT-22026.02 | 56.11 | — | — | — | — | |
| wiki-debiasedArchitecture=GPT-22026.02 | 56.49 | — | — | — | — | |
| VanillaArchitecture=BERT2026.02 | 58.01 | — | — | — | — | |
| ATLASBackbone=GPT-J2024.10 | — | 0.572 | — | — | — | |
| ATLASBackbone=GPT-2 XL2024.10 | — | 0.391 | — | — | — | |
| ATLASBackbone=LLAMA 22024.10 | — | 0.623 | — | — | — | |
| ATLASBackbone=LLAMA 32024.10 | — | 0.51 | — | — | — | |
| AutoDebiasBackbone=BERT, Multi-demographic Training=No (specific models per demographic)2026.01 | — | — | 46.85 | 47.51 | — | |
| BaselineBackbone Model=OLMoE, Expert Identification=ARR at λ*2026.03 | — | — | — | — | 0.679 | |
| BaselineBackbone Model=DeepSeek, Expert Identification=ARR at λ*2026.03 | — | — | — | — | 0.537 | |
| CausalDebiasBackbone=BERT, Multi-demographic Training=No (specific models per demographic)2026.01 | — | — | 46.75 | 52.89 | — | |
| DefaultBackbone=GPT-J2024.10 | — | 0.34 | — | — | — | |
| DefaultBackbone=GPT-2 XL2024.10 | — | 0.228 | — | — | — | |
| DefaultBackbone=LLAMA 22024.10 | — | 0.44 | — | — | — | |
| DefaultBackbone=LLAMA 32024.10 | — | 0.439 | — | — | — | |
| FARE FSPBackbone Model=OLMoE, Expert Identification=ARR at λ*2026.03 | — | — | — | — | 0.635 | |
| FARE FSPBackbone Model=DeepSeek, Expert Identification=ARR at λ*2026.03 | — | — | — | — | 0.517 | |
| FineDebBackbone=BERT, Multi-demographic Training=No (specific models per demographic)2026.01 | — | — | 54.55 | 53.69 | — | |
| PCGUBackbone=BERT2026.01 | — | — | 54.8 | 54.77 | — | |
| RandomBackbone Model=OLMoE, Expert Identification=ARR at λ*2026.03 | — | — | — | — | 0.645 | |
| RandomBackbone Model=DeepSeek, Expert Identification=ARR at λ*2026.03 | — | — | — | — | 0.531 | |
| RobustDebiasBackbone=BERT, Multi-demographic Training=Yes (single model across demographics)2026.01 | — | — | 52.25 | 56.67 | — | |
| SAFEx SESBackbone Model=OLMoE, Expert Identification=ARR at λ*2026.03 | — | — | — | — | 0.677 | |
| SAFEx SESBackbone Model=DeepSeek, Expert Identification=ARR at λ*2026.03 | — | — | — | — | 0.533 |