Well, it depends on the purpose of the machine learning mechanisms implemented in the papers you read. In general, a training dataset may contain any type of information, following the addressed problem domain. In the mentioned case (NOMA system), I suppose the training dataset contains info about channel quality, power levels and so on.

Also, you should consider what type of machine learning approach was considered: supervised, unsupervised or by reinforcement?

Can you mention the papers that you read and are basing on?

With respect to the answer given by Einar,I would like to clarify few doubts.

The paper which I follow is "unsupervised machine learning based user clustering in mm wave -NOMA systems".In that paper it was stated raw data set of users channel state information was been used.According to that paper I would like to check how raw data set is been generated.whether it needs to be generated by simulation or is there any other way of generating it.In addition that paper follows clustering as a result unsupervised machine learning is been used.

Following

I have read some parts of the paper you told us. As I understand, the authors consider two scenarios for clustering with K-means: one using the channel direction as a feature, where this information is extracted from the cosine of the angle between the channel vectors; another using only the normalized channel coefficients. The first scenario is employed in the Single-path mm-Wave Channel Model, and the second one is employed in the General Model.

Given the authors are evaluating the solution by simulation, the channel coefficients values are random and must follow a specific distribution model. I believe this information is in the paper or can be obtained from the authors themselves.