Large Language Model Debiasing on BBQ and CrowS-Pairs Out-of-Distribution (test)
0.98Mean BiasKLAAD-LoRA
Evaluation results
| Method | Links | ||
|---|---|---|---|
| KLAAD-LoRAEval Protocol=Main Eval protocol, Backbone=LLaMA-3-8B2026.03 | 0.98 | 1 | |
| UGID (OURS)Eval Protocol=Main Eval protocol, Backbone=LLaMA-3-8B2026.03 | 1.06 | 1.21 | |
| CDAEval Protocol=Main Eval protocol, Backbone=LLaMA-3-8B2026.03 | 1.16 | 1.29 | |
| ORIGINALEval Protocol=Main Eval protocol, Backbone=LLaMA-3-8B2026.03 | 9 | 15.65 | |
| ORIGINAL (SD)Eval Protocol=Self-Debias Eval protocol, Backbone=LLaMA-3-8B2026.03 | 9 | 15.65 | |
| SELF-DEBIAS (Ex.)Eval Protocol=Main Eval protocol, Backbone=LLaMA-3-8B2026.03 | 12.4 | 33.13 | |
| SELF-DEBIAS (Ex.)Eval Protocol=Self-Debias Eval protocol, Backbone=LLaMA-3-8B2026.03 | 12.4 | 33.13 | |
| SELF-DEBIAS (Re.)Eval Protocol=Main Eval protocol, Backbone=LLaMA-3-8B2026.03 | 15.47 | 29.19 | |
| SELF-DEBIAS (Re.)Eval Protocol=Self-Debias Eval protocol, Backbone=LLaMA-3-8B2026.03 | 15.47 | 29.19 |