TRF053
No manual label shifting in modeling code; self.loss_function owns it.
| Default | Enabled |
| Scope | All models |
| Source | mlinter/trf053.py |
| Show in terminal | mlinter --rule TRF053 |
What it does
Checks modeling_*.py and modular_*.py for assignments that build shift_logits/shift_labels (and shifted_ variants) by slicing, as in labels[…, 1:]. Receiving already-shifted labels (shift_labels = kwargs.pop(“shift_labels”, labels)) is the correct idiom and is not flagged.
Why is this bad?
self.loss_function shifts labels itself, so modeling code that pre-shifts trains on doubly-shifted targets or forces a bespoke loss path. Decoder-only models pass the raw labels and let the loss shift them. Encoder-decoder models are the mirror case: their labels are already shifted because the decoder input gets the decoder start token prepended, so they must pass shift_labels=labels to stop the loss from shifting again. Double-shift is the recurring training-loss bug (Git/Florence2/Moonshine family).
Example
if labels is not None:
- shift_logits = logits[..., :-1, :].contiguous()
- shift_labels = labels[..., 1:].contiguous()
- loss = nn.functional.cross_entropy(shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1))
+ # decoder-only: labels are unshifted, the loss shifts them
+ loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size)
+ # encoder-decoder: labels are already shifted, hand them over as shift_labels
+ loss = self.loss_function(logits=logits, labels=labels, shift_labels=labels, vocab_size=self.config.vocab_size)
Suppressing this rule
Add a # trf-ignore: TRF053 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.
Allowlisted models
15 models are exempt from TRF053 in mlinter/rules.toml, because they predate the convention and cannot be changed without breaking backward compatibility.
Show the 15 allowlisted models
blip_2clvpcsmgemma3gitgpt2granite_speechimagegptllama4mambamodernbert_decodermoshiopenaiqwen2_audioxlstm