As mentioned in the portal, but worth highlighting, the test set is released at the start of the 1 week long submission period. At this point, you should have two other sets, i.e., train and val. For training a model to generate submission for the test, train and val can be used however best fits the method. In other words, it can all be considered a train set, then even split in 5-fold; could be used as-is. Point is, all labeled data currently available can be used for supervised training. In any case, include all details in README, and then the paper later on. Note that the testing is blind. Hence, it is expected that the only time it will be handled is to generate a submission provided a system (input pair list and output submission), and we hope any attempt to understand or analyze the blind set will be avoided (this is what the validation was/is for). Thanks and best luck! We look forward to your submissions! Any doubts or questions, just ask :) In the end, just make sure to document and make clear wrt submission.
Posted by: jvision @ Jan. 13, 2020, 10:25 p.m.