README.md

---
license: mit
language:
  - en
tags:
  - unlearning
  - privacy
  - benchmark
  - multimodal
pretty_name: PPU-Bench
task_categories:
  - question-answering
  - text-generation
size_categories:
  - 100K<n<1M
configs:
  - config_name: train_pair_qwen3
    data_files: train_pair_qwen3/train-*.parquet
  - config_name: train_pair_gemma3
    data_files: train_pair_gemma3/train-*.parquet
  - config_name: train_pair_qwen3_qa
    data_files: train_pair_qwen3_qa/train-*.parquet
  - config_name: train_pair_gemma3_qa
    data_files: train_pair_gemma3_qa/train-*.parquet
  - config_name: unlearning_target
    data_files: unlearning_target/train-*.parquet
  - config_name: ppu_eval_cloze
    data_files:
      - split: test
        path: ppu_eval_cloze/test-*.parquet
  - config_name: ppu_eval_classification
    data_files:
      - split: test
        path: ppu_eval_classification/test-*.parquet
  - config_name: ppu_eval_generation
    data_files:
      - split: test
        path: ppu_eval_generation/test-*.parquet
  - config_name: cross_image_complete_r5
    data_files:
      - split: test
        path: cross_image_complete_r5/test-*.parquet
  - config_name: cross_image_complete_r15
    data_files:
      - split: test
        path: cross_image_complete_r15/test-*.parquet
  - config_name: cross_image_complete_r30
    data_files:
      - split: test
        path: cross_image_complete_r30/test-*.parquet
  - config_name: cross_image_selective
    data_files:
      - split: test
        path: cross_image_selective/test-*.parquet
  - config_name: cross_image_persona
    data_files:
      - split: test
        path: cross_image_persona/test-*.parquet
  - config_name: random_prefix_complete30
    data_files:
      - split: test
        path: random_prefix_complete30/test-*.parquet
  - config_name: random_prefix_selective
    data_files:
      - split: test
        path: random_prefix_selective/test-*.parquet
  - config_name: random_prefix_persona
    data_files:
      - split: test
        path: random_prefix_persona/test-*.parquet
  - config_name: jailbreak_style_prompt_complete30
    data_files:
      - split: test
        path: jailbreak_style_prompt_complete30/test-*.parquet
  - config_name: jailbreak_style_prompt_selective
    data_files:
      - split: test
        path: jailbreak_style_prompt_selective/test-*.parquet
  - config_name: jailbreak_style_prompt_persona
    data_files:
      - split: test
        path: jailbreak_style_prompt_persona/test-*.parquet
  - config_name: paraphrase_complete30
    data_files:
      - split: test
        path: paraphrase_complete30/test-*.parquet
  - config_name: paraphrase_selective
    data_files:
      - split: test
        path: paraphrase_selective/test-*.parquet
  - config_name: paraphrase_persona
    data_files:
      - split: test
        path: paraphrase_persona/test-*.parquet
---

# PPU-Bench: Real-World Multimodal Benchmark for Personalized Partial Unlearning

PPU-Bench is a real-world multimodal benchmark designed to evaluate personalized partial unlearning in vision-language models. It supports multiple unlearning settings and provides training/evaluation data for different VLM backbones.