Text Classification on AG News short protocol (600 steps)
82.8Accuracyvanilla Adam
Evaluation Results
| Method | Links | |
|---|---|---|
| vanilla AdamLearning Rate=3×10⁻⁴, Warmup=None, Architecture=Transformer (12-layer), Tuning=best tuned2026.07 | 82.8 | |
| Adam-InitOnlyLearning Rate=10⁻³, Initialization Cap (kappa)=0.005, Warmup=None, Architecture=Transformer (12-layer)2026.07 | 81 | |
| vanilla AdamLearning Rate=10⁻³, Warmup=None, Architecture=Transformer (12-layer)2026.07 | 25 |