Natural Language Understanding on GLUE (test) (MNLI, QQP, SST-2, SST-5)
82.2MNLI ScoreDP-SelFT+LoRA
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| DP-SelFT+LoRAPrivacy Budget (epsilon)=10, Backbone=RoBERTa-Large2026.05 | 82.2 | 79.6 | 93.81 | 51.2 | 76.7 | |
| LoRAPrivacy Budget (epsilon)=10, Backbone=RoBERTa-Large2026.05 | 81.2 | 78.7 | 93.12 | 50.9 | 75.98 | |
| DP-SelFT+LoRAPrivacy Budget (epsilon)=5, Backbone=RoBERTa-Large2026.05 | 80.5 | 79.5 | 93.46 | 51.2 | 76.17 | |
| DP-SelFTPrivacy Budget (epsilon)=10, Backbone=RoBERTa-Large2026.05 | 80.3 | 79.4 | 94.38 | 51.6 | 76.42 | |
| LoRAPrivacy Budget (epsilon)=5, Backbone=RoBERTa-Large2026.05 | 79.8 | 77.7 | 93 | 49 | 74.88 | |
| DP-SelFTPrivacy Budget (epsilon)=5, Backbone=RoBERTa-Large2026.05 | 79.7 | 79.1 | 94.15 | 51.4 | 76.09 | |
| Full ParameterPrivacy Budget (epsilon)=10, Backbone=RoBERTa-Large2026.05 | 78.2 | 77.3 | 94.15 | 50.6 | 75.06 | |
| Heuristic-basedPrivacy Budget (epsilon)=10, Backbone=RoBERTa-Large2026.05 | 77.9 | 78.8 | 93.69 | 47.9 | 74.57 | |
| Full ParameterPrivacy Budget (epsilon)=5, Backbone=RoBERTa-Large2026.05 | 77.7 | 77.2 | 94.04 | 50.4 | 74.84 | |
| Heuristic-basedPrivacy Budget (epsilon)=5, Backbone=RoBERTa-Large2026.05 | 77.6 | 78.5 | 93.92 | 49.3 | 74.83 | |
| DP-SelFTPrivacy Budget (epsilon)=1, Backbone=RoBERTa-Large2026.05 | 77 | 78.6 | 94.15 | 48.3 | 74.51 | |
| Full ParameterPrivacy Budget (epsilon)=1, Backbone=RoBERTa-Large2026.05 | 76.3 | 76.3 | 93.01 | 46.3 | 72.98 | |
| Heuristic-basedPrivacy Budget (epsilon)=1, Backbone=RoBERTa-Large2026.05 | 76.1 | 78.2 | 93.46 | 46.3 | 73.52 | |
| LoRAPrivacy Budget (epsilon)=1, Backbone=RoBERTa-Large2026.05 | 74.6 | 76.3 | 92.78 | 41.3 | 71.25 | |
| DP-SelFT+LoRAPrivacy Budget (epsilon)=1, Backbone=RoBERTa-Large2026.05 | 74.5 | 76.8 | 93 | 41.4 | 71.43 | |
| DP-SelFTPrivacy Budget (epsilon)=10, Backbone=OPT-350M2026.05 | 59 | 76.3 | 90.48 | 50.8 | 69.15 | |
| Full ParameterPrivacy Budget (epsilon)=10, Backbone=OPT-350M2026.05 | 55.5 | 74.8 | 89.91 | 48.8 | 67.25 | |
| DP-SelFTPrivacy Budget (epsilon)=5, Backbone=OPT-350M2026.05 | 54.4 | 74.6 | 90.37 | 50 | 67.34 | |
| Full ParameterPrivacy Budget (epsilon)=5, Backbone=OPT-350M2026.05 | 47.5 | 74.4 | 89.45 | 48.6 | 64.99 | |
| Heuristic-basedPrivacy Budget (epsilon)=10, Backbone=OPT-350M2026.05 | 46.5 | 74.5 | 90.37 | 48.6 | 64.99 | |
| Heuristic-basedPrivacy Budget (epsilon)=5, Backbone=OPT-350M2026.05 | 42.7 | 71.6 | 89.79 | 49 | 63.27 | |
| DP-SelFT+LoRAPrivacy Budget (epsilon)=10, Backbone=OPT-350M2026.05 | 41.4 | 66.8 | 89.11 | 48.2 | 61.38 | |
| Full ParameterPrivacy Budget (epsilon)=1, Backbone=OPT-350M2026.05 | 40.8 | 63.8 | 87.73 | 47.3 | 59.91 | |
| DP-SelFTPrivacy Budget (epsilon)=1, Backbone=OPT-350M2026.05 | 40.2 | 64.5 | 88.76 | 47.8 | 60.32 | |
| DP-SelFT+LoRAPrivacy Budget (epsilon)=5, Backbone=OPT-350M2026.05 | 38.8 | 64.3 | 88.53 | 46.4 | 59.51 | |
| LoRAPrivacy Budget (epsilon)=10, Backbone=OPT-350M2026.05 | 38.7 | 64.8 | 88.88 | 47.4 | 59.95 | |
| Heuristic-basedPrivacy Budget (epsilon)=1, Backbone=OPT-350M2026.05 | 36.8 | 61.2 | 88.18 | 47.3 | 58.37 | |
| LoRAPrivacy Budget (epsilon)=5, Backbone=OPT-350M2026.05 | 36.1 | 63.1 | 88.07 | 46.2 | 58.37 | |
| DP-SelFT+LoRAPrivacy Budget (epsilon)=1, Backbone=OPT-350M2026.05 | 35.6 | 61.7 | 85.32 | 43.1 | 56.43 | |
| LoRAPrivacy Budget (epsilon)=1, Backbone=OPT-350M2026.05 | 34.5 | 61.6 | 84.86 | 43.8 | 56.19 |