Reduce inconsistency across trainer test files

[qgallouedec]

What changed

• Renamed test_training_* → test_train_* (stragglers in DPO, GRPO, RLOO, and 10 experimental files)
• Aligned wording: parameters → params, Check the → Check that the
• Removed redundant section comments that describe WHAT instead of WHY (# Get the dataset, # Initialize the trainer, # Train the model, etc. — per CLAUDE.md guidance)
• Fixed within-file outliers (period style, torch.equal vs torch.allclose, missing inline comments on shared config args)

---

Note

Low Risk
Test-only refactors (renames, comment cleanup, and assertion consistency) with minor dataset-loading adjustments; low risk outside of potentially changing which splits are exercised in a few tests.

Overview
Improves consistency across trainer test suites by renaming remaining test_training* methods to test_train*, tightening/standardizing assertion messaging, and removing redundant narrative comments.

Also aligns dataset usage in several tests (e.g., consistently using load_dataset(..., split="train") where a single-split dataset is expected) and standardizes parameter-change checks to use torch.equal (and params wording) for clearer, uniform test behavior.

^{Reviewed by Cursor Bugbot for commit c06fb06. 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.

There are 2 total unresolved issues (including 1 from previous review).

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

^{Reviewed by Cursor Bugbot for commit c5a2467. Configure here.}