To help reproducibility, and to track the critical information of clusters/embeddings and feature extractors used, it makes sense to store this metadata about models at the time the model is learned.
Metadata is hard to embed in binary versions of trained model, and we want to keep model files transferable, which rules out some pickled something or other. A decent first solution is therefore probably an extra file generated after/before the model is trained, named something like model_filename +'.metadata' .
To help reproducibility, and to track the critical information of clusters/embeddings and feature extractors used, it makes sense to store this metadata about models at the time the model is learned.
Metadata is hard to embed in binary versions of trained model, and we want to keep model files transferable, which rules out some pickled something or other. A decent first solution is therefore probably an extra file generated after/before the model is trained, named something like model_filename +'.metadata' .