Natural Language Understanding on GLUE (val) - Subset (SST-2, MRPC, RTE)
92.43SST-2 AccuracyUncased baseline
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Uncased baselineBackbone=BERT, Non-linearity Type=Exact (Softmax, LayerNorm, GELU), Execution Domain=Plaintext, Model Casing=Uncased2026.04 | 92.43 | 87.75 | 91.17 | 72.56 | |
| Uncased baselineNon-linearity Type=Exact (softmax, LayerNorm, GELU), Model Casing=Uncased, Backbone=BERT-base2026.04 | 92.43 | 87.75 | 91.17 | 72.56 | |
| Fine-tuned teacherBackbone=BERT, Non-linearity Type=Surrogate (MBMax, MBLN, approx GELU), Execution Domain=Plaintext, Model Casing=Uncased2026.04 | 92.13 | 87.01 | 90.82 | 69.43 | |
| Fine-tuned teacherNon-linearity Type=Surrogate (MBMax, MBLN, approximate GELU), Backbone=BERT-base2026.04 | 92.13 | 87.01 | 90.82 | 69.43 | |
| EncFormerBackbone=BERT, Non-linearity Type=Surrogate (MBMax, MBLN, approx GELU), Execution Domain=Encrypted (FHE+MPC), Model Casing=Uncased2026.04 | 91.78 | 86.76 | 90.34 | 70.03 | |
| EncFormerExecution Domain=Encrypted (FHE+MPC), Backbone=BERT-base, Non-linearity Type=Surrogate (MBMax, MBLN, approximate GELU)2026.04 | 91.78 | 86.76 | 90.34 | 70.03 |