- Task: Multi-dialect Japanese dialect identification
- The weight decay was drawn from the range 1e-6 to 1e-1.
- Fifteen trials were conducted using Optuna as the search framework.
- Input truncation was enabled.
The model released here corresponds to the best-performing trial and checkpoint selected based on macro F1.
Training configuration for this model:
- Optimizer: AdamW
- Learning rate: 5e - 5
- Learning rate scheduler: Linear
- Batch size: 16
- Epochs: 3
- Warmup ratio: 0
- Warmup steps: 0
- Random seed: 42
- Weight decay: 0.07422069766185352