A fitted model: a reference into the session's ML cache, and the context needed to use it.

Not a value. The model itself lives on the server, in a cache that offloads it to disk when memory is tight and evicts it when the session ends — so a %Latu.ML.Model{} can name something that is no longer there, and every verb that touches one can fail for that reason. Latu.ML.with_model/3 is the bracket; Latu.ML.delete/1 is the explicit release. There is no finalizer, because BEAM terms have none.

dataset is the plan the model was fitted on, kept so a summary the server has evicted can be rebuilt from it. It is nil for a model from Latu.ML.load/3, which was not fitted here — and which has no summary on the server either, because a model's save format does not carry one. uid is the client's, copied from the estimator: a Fit answers with an obj_ref and nothing else. A loaded model's is the one the saved metadata held, which is the only command that gives a uid back.

class is nil for a model whose estimator can fit more than one — LDA returns a LocalLDAModel or a DistributedLDAModel depending on its optimizer, and a Fit does not say which. candidates is then both of them, and is what an attribute is looked up in; nothing claims a class it cannot know.