In the case of the training set there are certain ranges of dates with no images (I'm guessing due to clouds). In such dataset this tends to happen for rows with pretty small date ranges (there are registers with "unrealistically" short differences between Harvest and Sowing, even less than 10 days). Such thing also happens on the test set.

Also, while in the test dataset the rows with Harvest date before Sowing date where eliminated, there ares still rows that fullfill that condition on the training set. Thus, setting an invalid date range for the tiff files download.