The feature training dataset consists of 500 rows, each corresponding to a different observation, and 5000 features, and a response dataset with 500 observations.

The feature test dataset, consists on 1000 observations of the 5000 features. You will be evaluated on the response corresponding to this dataset.

The kind of transformations on your data (original) or reduced, should be limited to simple functions like:

$$log(x)$$
$$|x|$$
$$x^2$$
$$sqrt(x)$$

You are welcome to try more complex relationships, but the underlying model is not more complicated than this. You can also use advance techniques like Neural Networks.