• model: klue/roberta-large
• learning rate: 1e-4
• lr scheduler type: linear
• weight decay: 0.01
• epochs: 5
• checkpoint: 2700