Great catch—you’re absolutely right to notice this mismatch between the tutorial’s terminology and standard machine learning workflow norms. Let’s break this down clearly:

The Core Issue: Mislabeled Dataset Role #

Yes, the "test dataset" referenced in the tutorial is functionally a validation dataset, not a true test dataset. Here’s why:

• In standard ML practice, a test dataset should only be used once, after full training is complete, to evaluate the final model’s generalization to unseen data.
• But in the tutorial, this dataset is used every training epoch to track the model’s performance during training—this is exactly the role of a validation dataset, which helps monitor overfitting, tune hyperparameters, and gauge how well the model is generalizing during the training process.

Why the Tutorial Does This #

The PyTorch team likely chose this simplified approach to keep the example focused on the core topic: optimizing model parameters. Introducing a three-way split (train/validation/test) would add extra code and complexity that distracts from the main lesson (loss calculation, optimizer steps, etc.). For a beginner-focused tutorial, reducing moving parts helps keep the learning curve gentle.

What’s Missing: A True Test Dataset #

You’re also correct that the tutorial doesn’t include a proper, held-out test dataset. In a real-world project, you’d typically split your data into three distinct subsets:

• Training set: Used to update the model’s weights directly
• Validation set: Used during training to adjust hyperparameters and monitor performance
• Test set: Held completely separate until the end, to get an unbiased measure of the final model’s real-world performance

If you wanted to adapt the tutorial to follow standard practices, you’d modify the dataset split to create a separate validation set (using the tutorial’s current "test" data as validation) and set aside a small, untouched portion of data as the true test set.

内容的提问来源于stack exchange，提问作者Horizon