Is it possible to train a decoder with embeddings from a multi-session model?

[ELGarulli]

Dear CEBRA community,

I am taking my first steps in the ML/AI world so please forgive any misunderstanding on the functioning and underlying principles. I would love to have any input from you on this.

I have a dataset containing neural and kinematic features from multiple experimental runs (55). I don't want to concatenate all the runs, as mentioned in #49 this would artificially create weird junctions. So I am trying the multi-session training strategy.
At first, just to get the pipeline up and running I used the whole dataset, of course, I plan to do the proper test and training split.
Briefly, the X is a list of 55 2D matrixes containing my features and the y is a list of 1D arrays containing labels that classify different "states" of my subjects (possible values: 0, 1, 2).
Q1: does the pipeline until now make sense?
EDIT: I realized I am misusing the labels as continuous, while discrete values are not supported by multisessions models yet. I'll be waiting for #135

My next step is to train the model only on a subset of sessions and see if I can decode the states of unseen sessions.
However, since I am using the multi-session model I have multiple embeddings. I am unsure about how (if at all) to combine the embeddings to train the decoder.
I know this might be more of a basic question, not strictly related to CEBRA. But I would really appreciate any input and feedback both on my reasoning and practical steps to achieve what I have in mind.
Q2: is it possible to train a decoder with the embeddings from a multi-session model?
EDIT: I realized the Demo_Allen is doing something similar enough, will try to use this.

Thanks a lot for your help! And thanks a lot for this tool, I'm having a lot of fun using it!