Align KTO with DPO: Use _metrics attribute

[albertvillanova]

Align KTO with DPO: Use _metrics attribute.

Part of:

• KTO refactoring #4786

This PR refactors how training and evaluation metrics are stored and managed in the KTOTrainer class, replacing the previous _stored_metrics structure with a new _metrics dictionary. This change centralizes metric tracking and simplifies metric handling throughout the class.

Changes

Metrics storage and management:

• Replaced the _stored_metrics attribute (a nested defaultdict) with a new _metrics dictionary, initialized to separately track "train" and "eval" metrics.
• Updated all methods (store_metrics, log) to use the new _metrics structure instead of _stored_metrics, ensuring consistent metric accumulation and cleanup.

These changes improve code clarity and reduce the risk of errors when handling training and evaluation metrics.

---

Note

Low Risk
Low risk refactor limited to how KTOTrainer accumulates and clears train/eval metrics, but it could affect what gets logged if the new containers are mishandled.

Overview
Updates KTOTrainer to use a unified _metrics container (split into train and eval) instead of _stored_metrics, aligning metric bookkeeping with DPO.

Metric accumulation (store_metrics) and log-time aggregation/cleanup (log) are refactored to write into and clear from the new structure, including the chosen/rejected count-based averaging logic.

^{Reviewed by Cursor Bugbot for commit 881aec3. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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 c2ec97c. 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.