Natural Language Understanding on GLUE large
83.5MNLI AccuracyARMADA
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ARMADAModels=BERT-base, Distillation type=ARMADA - Leuclid2026.03 | 83.5 | 91 | 88.8 | 92 | |
| ARMADAModels=BERT-base, Distillation type=ARMADA - Lcosine2026.03 | 83.4 | 91 | 89.4 | 93.1 | |
| ARMADAModels=BERT-base, Distillation type=ARMADA - Lelementwise2026.03 | 83.4 | 91.2 | 88.9 | 92.8 | |
| ARMADAModels=BERT-6L, Distillation type=ARMADA - Leuclid2026.03 | 80.3 | 87.2 | 88.9 | 90.6 | |
| ARMADAModels=BERT-6L, Distillation type=ARMADA - Lcosine2026.03 | 80.2 | 87.3 | 88.6 | 90.7 | |
| ARMADAModels=BERT-6L, Distillation type=ARMADA - Lelementwise2026.03 | 80.1 | 87.3 | 89.1 | 90.7 | |
| BERT-baseModels=BERT-base, Distillation type=undistilled2026.03 | 79.3 | 87.8 | 83.1 | 89.2 | |
| BERT-6LModels=BERT-6L, Distillation type=undistilled2026.03 | 77.7 | 84.9 | 89 | 87.6 |