Is it possible to to only use only a training data sample for creating a LIME model explainer?

I have been looking into outputting a model explainer artefact at time of training my Keras+Tensorflow Neural network.

Lime seems like a great choice however my data is very big and I am reading from disk one batch at a time as it is impractical and inefficient to store in memory. Lime appears to require the whole training dataset to be inputted for it to be able to create a surrogate model.

Is it appropriate to use only a sample of training data? Does anyone have experience successfully using Lime like this? What would best practice be in this case?