Natural Language Inference on GLUE MNLI
90.6Accuracy32-bit AdamW
Evaluation results
| Method | Links | |
|---|---|---|
| 32-bit AdamWModel=RoBERTa-Large, Opt.State Mem Saved=reference, Task Setting=Fine-tuning2026.04 | 90.6 | |
| STQuantModel=RoBERTa-Large, Opt.State Mem Saved=84.4%, Task Setting=Fine-tuning2026.04 | 90.32 | |
| 8-bit AdamW(bnb)Model=RoBERTa-Large, Opt.State Mem Saved=75.0%, Task Setting=Fine-tuning2026.04 | 90.02 |