Classification on SuperGLUE
72.2RTE ScoreCurvZO
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| CurvZOModel=OPT-6.7B, Number of training samples=1,000, Optimization Method=CurvZO, Fine-tuning Protocol=Full fine-tuning2026.03 | 72.2 | 61.7 | 61.5 | |
| MeZOModel=OPT-6.7B, Number of training samples=1,000, Optimization Method=MeZO, Fine-tuning Protocol=Full fine-tuning2026.03 | 70.3 | 62.5 | 60.5 | |
| DiZOModel=OPT-6.7B, Number of training samples=1,000, Optimization Method=DiZO, Fine-tuning Protocol=Full fine-tuning2026.03 | 66.4 | 61.6 | 59.6 | |
| MeZO + LoRAModel=OPT-6.7B, Number of training samples=1,000, Optimization Method=MeZO, Fine-tuning Protocol=LoRA2026.03 | 64.9 | 61.4 | 60.5 | |
| CurvZO + LoRAModel=OPT-6.7B, Number of training samples=1,000, Optimization Method=CurvZO, Fine-tuning Protocol=LoRA2026.03 | 63.5 | 61.5 | 63.4 | |
| DiZO + LoRAModel=OPT-6.7B, Number of training samples=1,000, Optimization Method=DiZO, Fine-tuning Protocol=LoRA2026.03 | 58.8 | 57.2 | 58.6 |