Perhaps your features (`output` in this case) have excessive nesting (inputs type `list` where type `int` is expected)

This smells like an unresolved issue…
OK, let’s work around it. If we process the dataset in advance, SFTTrainer won’t complain.