Remove redundant alignment of pad_token_id

[albertvillanova]

Remove redundant alignment of pad_token_id.

Note that this alignment is already done by transformers.Trainer.train() → self.align_special_tokens() / self._align_special_tokens()

This PR removes redundant alignment of pad_token_id between the tokenizer and the mdoel config, relying instead on the transformers.Trainer.align_special_tokens(), called during training.

Changes

Padding token management cleanup:

• Removed explicit assignment of model.config.pad_token_id = tokenizer.pad_token_id in examples/scripts/prm.py.
• Removed explicit assignment of model.config.pad_token_id = processing_class.pad_token_id in the RewardTrainer initialization.

---

Note

Low Risk
Low risk cleanup that removes duplicate pad_token_id assignments and relies on Transformers' built-in special-token alignment during training; behavior could change only if code paths depend on the earlier manual override before Trainer.train() runs.

Overview
Removes manual alignment of model.config.pad_token_id with the tokenizer/processing class in both examples/scripts/prm.py and RewardTrainer, relying on Transformers' Trainer special-token alignment instead.

This reduces redundant configuration mutation and keeps padding token handling centralized in the upstream training flow.

^{Reviewed by Cursor Bugbot for commit bcd4e6c. Bugbot is set up for automated code reviews on this repo. Configure here.}

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit bcd4e6c. Configure here.}

[cursor]

Removal relies on feature absent in minimum transformers version

Medium Severity

The removed model.config.pad_token_id = processing_class.pad_token_id assignment relies on Trainer._align_special_tokens() to handle alignment automatically, but this mechanism only exists in transformers v5.5.0+ (main branch). The project's minimum supported version is transformers>=4.56.2, where no such alignment occurs. For models with config.pad_token_id = None (e.g., Llama, Mistral), sequence classification models use config.pad_token_id during forward-pass pooling — without alignment, this can cause ValueError for batch sizes > 1 or incorrect pooling on older supported versions.

Additional Locations (1)

• examples/scripts/prm.py#L90-L91

 

^{Reviewed by Cursor Bugbot for commit bcd4e6c. Configure here.}

[albertvillanova]

I have checked transformers code:

• _align_special_tokens was implemented in v4.56.0 by:
  • [trainer] ensure special tokens in model configs are aligned with tokenizer at train time transformers#38441
• it was renamed to align_special_tokens in v5.2.0 by:
  • Minor changes trainer transformers#43744

[qgallouedec]

can you quickly check when _align_special_tokens was introduced in trainer? it's not that very old iirc

[albertvillanova]

Done @qgallouedec. See: #5487 (comment)

[qgallouedec]

lgtm