Fix dataset fingerprinting in experimental KTO and TPO trainers

[DaoyuanLi2816]

Follow-up to #6206, which fixed dataset fingerprinting in the core DPOTrainer / SFTTrainer / RewardTrainer. The two experimental trainers KTOTrainer and TPOTrainer still have the pattern #6206 removed: their tokenization map is a closure that calls self._tokenize(...), so it captures self. A map function that closes over self (which holds the model and other unpicklable state) can't be hashed stably, so datasets falls back to a random fingerprint and silently disables caching of the tokenized dataset — re-tokenizing on every run.

This applies the same fix as #6206: make _tokenize a @staticmethod and bind it to a local (tokenize = self._tokenize) before defining the map function, so the closure no longer captures self.

• KTO: _tokenize uses the VLM flag, so is_vlm becomes an explicit parameter passed through fn_kwargs (exactly as in Fix dataset fingerprinting in DPO/SFT tokenization #6206's DPO).
• TPO: _tokenize doesn't reference self at all, so it simply becomes a @staticmethod.

No behavior change: the tokenized output is identical; only the map function's hashability (and therefore cache reuse across runs) changes.

CPOTrainer and ORPOTrainer share the same pattern (a self-bound tokenize_row) but are intentionally left out here to avoid conflicts: CPO's data pipeline is being refactored in #5837, and ORPO's _prepare_dataset is touched by #6230. They can be aligned once those land.

Verification

• tests/experimental/test_kto_trainer.py and tests/experimental/test_tpo_trainer.py pass in full (tokenized output unchanged); ruff check / ruff format --check clean. Validated on a single GPU (RTX 4080).
• Like Fix dataset fingerprinting in DPO/SFT tokenization #6206, this doesn't add a fingerprint-specific test: the fallback to a random fingerprint only triggers for processing classes / models that datasets can't hash, which the tiny CI fixtures don't exercise. The existing suites cover tokenization correctness.

---

Note

Low Risk
Refactor limited to dataset preprocessing closures with no training-loss or tokenization logic changes; behavior is intended to be identical aside from caching.

Overview
Aligns experimental KTOTrainer and TPOTrainer with the dataset-fingerprint fix from #6206 so dataset.map during preprocessing can be cached across runs instead of re-tokenizing every time.

In both trainers, _tokenize is now a @staticmethod, and tokenize_fn calls a local tokenize = self._tokenize so the map callback no longer closes over self (which held the model and broke stable hashing). KTO also threads is_vlm as an explicit argument via fn_kwargs, matching DPO’s VLM handling.

Tokenized outputs are unchanged; only cache fingerprint stability improves.

^{Reviewed by Cursor Bugbot for commit bf7b8e4. Bugbot is set up for automated code reviews on this repo. Configure here.}

[albertvillanova]

Duplicate of:

• Align KTO with DPO: Align _tokenize #6212
• Align experimental TPO _tokenize with core trainers #6213