Generalization gap prediction on MNLI
0.09Gap Prediction ErrorInconsistency
Evaluation results
| Method | Links | |
|---|---|---|
| InconsistencyTraining scenario=Fine-tuning, Data source=Unlabeled data2023.05 | 0.09 | |
| 1-sharpnessTraining scenario=Fine-tuning, Data source=Test loss landscape (dev data, unfair advantage)2023.05 | 0.22 | |
| HessianTraining scenario=Fine-tuning, Data source=Test loss landscape (dev data, unfair advantage)2023.05 | 0.27 | |
| 1-sharpnessTraining scenario=Fine-tuning, Data source=Training loss landscape2023.05 | 0.31 | |
| HessianTraining scenario=Fine-tuning, Data source=Training loss landscape2023.05 | 0.7 |