Document how custom dataset should be formatted

[pfliu-nlp]

Currently, we don't describe how custom datasets (for example: this) should be organized in the doc that describes each task's analysis setup.