TRF031

Output-like dataclasses in modeling files must inherit ModelOutput.

Default	Enabled
Scope	Models added on or after 2026-06-20
Source	mlinter/trf031.py
Show in terminal	mlinter --rule TRF031

What it does

In modeling_*.py and modular_*.py, flags a top-level @dataclass whose bases include nothing ending in Output, unless it has two or more mandatory fields – those are internal argument bundles, which ModelOutput rejects at runtime.

Why is this bad?

A plain output dataclass does not index like a tuple, does not survive return_dict=False, and is invisible to @auto_docstring, so its fields never reach the generated API docs. ModelOutput gets all three for free.

Example

 @auto_docstring
 @dataclass
-class AcmeStructureOutput:
+class AcmeStructureOutput(ModelOutput):
     positions: torch.Tensor
     confidence: Optional[torch.Tensor] = None

Suppressing this rule

Add a # trf-ignore: TRF031 comment on the flagged line or the line directly above it. See Suppressing rules for whole-file directives and when a suppression is the wrong answer.