[nxZhai]

What does this PR do?

This PR adds the implementation of the ACL 2026 paper PALU (2601.03190, nxZhai/PALU) to Open-Unlearning.

• Add the PALU trainer with prefix-aware localized logit flattening and configurable forward, reverse, or equally weighted mixed KL constraints.
• Add annotated TOFU data, a dedicated dataset loader, trainer and experiment configurations, and registry entries.
• Add example and parameter-grid scripts, method documentation, and regression tests.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Have you gone through the contributions guide?
• Are your changes documented? Read documentation guidelines here.

[Copilot]

🟡 Changes recommended

Both PALU training scripts default to an unpopulated reference-log directory, causing evaluation to fail before training begins under the documented setup.

1 open issue

• Default reference log path to the documented saves/eval directory · New

What changed in this PR

Adds PALU to Open-Unlearning’s trainer and dataset pipelines, with TOFU configurations, usage examples, and regression tests.

Changes:

• Implements prefix-aware localized logit flattening with forward, reverse, or equally weighted mixed KL preservation.
• Adds an annotation-aware dataset loader, registry entries, and TOFU experiment settings.
• Provides example and parameter-grid scripts, method documentation, and regression coverage.

File	Description
tests/​test_palu.py	Adds objective, gradient, masking, and dataset regression checks.
src/​trainer/​unlearn/​palu.py	Implements the PALU training objective.
src/​trainer/​__init__.py	Registers the PALU trainer.
src/​data/​palu.py	Maps sensitive character spans to token labels.
src/​data/​__init__.py	Registers the PALU dataset loader.
README.md	Lists PALU among supported methods.
docs/​links.md	Adds PALU research and source links.
configs/​trainer/​PALU.yaml	Defines PALU trainer defaults.
configs/​experiment/​unlearn/​tofu/​palu.yaml	Configures PALU experiments on TOFU.
configs/​data/​datasets/​TOFU_QA_forget_palu.yaml	Configures annotated forget-data loading.
community/​methods/​PALU/​scripts/​tofu/​train_tofu_palu_llama3.sh	Adds a Llama-3.1 parameter sweep.
community/​methods/​PALU/​scripts/​tofu/​train_tofu_palu_llama2.sh	Adds a Llama-2 parameter sweep.
community/​methods/​PALU/​run.sh	Provides a single-run example.
community/​methods/​PALU/​README.md	Documents the method, annotations, and usage.
community/​methods/​PALU/​LICENSE	Preserves the upstream MIT notice.
community/​methods/​PALU/​data/​forget10_with_common_words_gpt/​state.json	Defines saved-dataset state for forget10.
community/​methods/​PALU/​data/​forget10_with_common_words_gpt/​dataset_info.json	Defines the forget10 annotation schema.
community/​methods/​PALU/​data/​forget05_with_common_words_gpt/​state.json	Defines saved-dataset state for forget05.
community/​methods/​PALU/​data/​forget05_with_common_words_gpt/​dataset_info.json	Defines the forget05 annotation schema.
community/​methods/​PALU/​data/​forget01_with_common_words_gpt/​state.json	Defines saved-dataset state for forget01.
community/​methods/​PALU/​data/​forget01_with_common_words_gpt/​dataset_info.json	Defines the forget01 annotation schema.
community/​methods/​PALU/​.gitignore	Allows tracking bundled annotation data.

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

A tokenizer-inserted terminal EOS remains masked in answer-containing examples, omitting its preservation loss instead of receiving the expected -200 label.

No open issues 🎉

1 previously missed issue

In code that hasn't changed since last review

Unmask existing terminal EOS when answer tokens are present

src/​data/​palu.py:117

When a fast tokenizer inserts a terminal EOS with (0, 0) offsets, the label loop assigns it -100, and this branch leaves it masked because EOS is already present. Its prediction therefore receives no preservation loss, unlike an explicitly appended EOS. Set an existing terminal EOS label to -200 when answer tokens are present, including at max_length, while leaving prompt-only truncated examples masked.

1 resolved since last review

• Default reference log path to the documented saves/eval directory

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

No active review issues block the PALU trainer and annotated-data integration.

No open issues 🎉

_{Review effort: Balanced}

[molereddy]

couple small nitpicks to cleanup otherwise looks good! thanks!!

[nxZhai]

couple small nitpicks to cleanup otherwise looks good! thanks!!

Thanks for pointing those out! I've fixed them. Could you please take another look?

[molereddy]

thank you for contributing!