Natural Language Understanding on SuperGLUE (BoolQ, COPA, WSC, RTE, MultiRC, WiC)
88.5BoolQ AccuracyChunkFT
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| ChunkFTModel=RoBERTa-large, Hardware=RTX 4090-24GB2026.05 | 88.5 | 73.6 | 66.4 | 85.8 | 77.5 | 73.5 | 77.6 | |
| HiFTModel=RoBERTa-large, Hardware=H800-80GB2026.05 | 87.3 | 71.5 | 67.7 | 83.5 | 78.6 | 66.5 | 75.9 | |
| AdamModel=RoBERTa-large, Hardware=H800-80GB2026.05 | 86.5 | 59.5 | 68.7 | 87.6 | 76.6 | 70.5 | 74.9 | |
| BAdamModel=RoBERTa-large, Hardware=H800-80GB2026.05 | 85.4 | 69.4 | 65.5 | 76.3 | 77.3 | 64.6 | 73.1 | |
| GaLoreModel=RoBERTa-large, Hardware=H800-80GB2026.05 | 85.2 | 68.3 | 63.5 | 74.4 | 77.2 | 71.5 | 73.4 | |
| LoRAModel=RoBERTa-large, Hardware=H800-80GB2026.05 | 81.5 | 56.7 | 62.6 | 79.8 | 69.5 | 59.5 | 68.3 |