Natural Language Understanding on NLU Suite Subset (SST2, COPA, CB, BoolQ, RTE, WiC) (val)
82.1SST2 AccuracyFO
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| FOOptimization Method=First-Order (FO), Hardware Platform=NPU, Model=Pangu-1B2026.01 | 82.1 | 80 | 69.6 | 75.2 | 78 | 65.7 | 73.8 | |
| MEZOOptimization Method=Zeroth-Order (MEZO), Hardware Platform=NPU, Model=Pangu-1B2026.01 | 76.6 | 74 | 73.2 | 69.7 | 72.9 | 55.2 | 70.3 | |
| AGZOOptimization Method=Activation Guided Zeroth Order (AGZO), Hardware Platform=NPU, Model=Pangu-1B2026.01 | 76.5 | 77 | 69.6 | 72.8 | 72.9 | 56.7 | 70.9 | |
| LOZOOptimization Method=Zeroth-Order (LOZO), Hardware Platform=NPU, Model=Pangu-1B2026.01 | 71.8 | 72 | 64.3 | 69.4 | 68.2 | 54.2 | 66.7 | |
| ICLOptimization Method=In-context Learning (ICL), Hardware Platform=NPU, Model=Pangu-1B2026.01 | 71 | 74 | 44.6 | 73.1 | 68.2 | 49.5 | 63.6 | |
| ZeroOptimization Method=None (Zero-shot), Hardware Platform=NPU, Model=Pangu-1B2026.01 | 57.1 | 76 | 48.2 | 69.6 | 57.8 | 46.9 | 59.3 |