Model Validation Loss Zero

I am working on multiclass classification using a custom dataset with four classes. The dataset consists of 4080 images with dimensions of 2040x2040, and the classes are balanced. I am using the Convnext_tiny model with ADAM optimization (weight_decay=0.01, lr=0.0001) and crossentropy loss. My training losses are as follows: Val_loss ≈ 0.0, F1 ≈ 1. The model performs well on the validation dataset but poorly on the test dataset. How can I address this issue?

Convnext fully connected layer -
model=models.convnext_tiny(pretrained=True)
num_ftrs = model.classifier[2].in_features # Son fully connected katmanın giriş özellik sayısı

        model.classifier[2] = nn.Sequential(
            nn.Linear(num_ftrs, 1000),
            nn.ReLU(),
            nn.LayerNorm(1000),
            nn.Linear(1000,4),
            
        )

My train dimension 1728*1728