My model has 0.7142 Accuracy in 99% of the time while my loss value jumps around

I am using a lstm as a classifier for timeseries data for 3 classes. I print out my predictions and labels at the end with my loss and accuracy. No matter if my Loss is 0.1 or 1.0 my accuracy is almost always 0.7142.

I am wondering if I calculated loss and accuracy correctly in my model or if this is just a problem about hyperparamter tuning. I am under a lot of time pressure so if someone could help me I would be very thankful.

My validation loop looks as follows.
EDIT: I am using CrossEntropyLoss() as criterion

for i, (inputs, labels) in enumerate(valloader):

            y_pred = model(inputs)                                                 
            _,prediction2 = torch.max(y_pred, 1)
            labels = labels.type_as(y_pred)
            labels = labels.long()
            val_loss = criterion(y_pred, labels)
            running_val_loss += val_loss.item()

            for x in range(len(labels)):
                y_true_list.append(labels[x].item())
            for x in range(len(prediction2)):
                y_pred_list.append(prediction2[x].item())
            
            total += labels.size(0)
            correct += (prediction2 == labels).sum().item()
            
            #getting output for metric calc correct
            softmax = nn.Softmax(dim=1)
            probability = softmax(y_pred)
           
            for x in range(len(probability)):
                probability_each = probability[x].detach().numpy()
                prob_softmax.append(probability_each)
           val_loss_mean = running_val_loss/len(valloader)
        validation_losses.append(val_loss_mean)

        accuracy = (correct/total)
        validation_accuracys.append(accuracy)
 
        print("Y True and Y Pred :",y_true_list, y_pred_list)
        print(f'Fold {fold}: For Epoche {epoch} the Loss is {val_loss_mean} and the Acc is {accuracy}')

I guess 0.7142 would correlate to the potential class imbalance in your Dataset?
Could you check, if your model is predicting the same class for all samples?

That is what I thought as well but it does not. I am printing the true label list and the predicted label list at the end and my model does not always predict the same class.

Maybe I found one problem in my model which causes this. Maybe my model does not know when a sequence inside a batch ends and the next one starts and it is treating it like a whole sequence? I am having trouble understanding how to make my model effectively use batch training. Different sources say different things about implementing hidden state reset and none work for me.