How to construct validation_data. Which is correct ?

[ken-take-it-so-so]

Hello.
This question is different from #2219. In #2219, I understand validation_data should contain full training_data.
By the way, I want to use exchange_rate_nips dataset provided in GluonTS. This data is used in some papers. Dataset for train has 8 time series and dataset for test has 40(8 times 5) time series because of rolling evaluation.

from gluonts.dataset.repository.datasets import get_dataset

dataset = get_dataset("exchange_rate_nips")
print(f"len(dataset.train): {len(dataset.train)}")
print(f"len(dataset.test): {len(dataset.test)}")

Output

len(dataset.train): 8
len(dataset.test): 40

I think below figures explain the dataset.

I will use DeepVAREstimator after processing dataset properly. After that, I will use .train(training_data, validation_data). In this case, which is correct ?

1. varidation_data has 8 time series. (like dataset.train shape)
2. varidation_data has 40 time series. (like dataset.test shape)

I would appreciate your cooperation.