Relearning-Controlled-300

A controlled multimodal person-fact dataset for studying cross-identity recovery after machine unlearning and partial relearning attacks.

The dataset contains 300 core synthetic identities:

• 100 target identities whose synthetic birth-city bindings are removed;
• 200 retain identities;
• five independent S=20, H=80 relearning partitions;
• fixed-H attacks with S=10/20/30;
• reference-model, forget, retain, attack-control and novel-fact datasets;
• 2,000 four-cell binding contrasts for representation analysis;
• 900 held-out transformed-pose images;
• a 500-identity visual QA collection with 1,500 AI-annotated questions.

Names, city labels, occupations and code words are synthetic. Images originate from MLLMU-Bench. The visual-attribute answers are preliminary AI annotations, not official MLLMU-Bench labels, and should be manually reviewed before being treated as publication-quality ground truth.

Documentation

See 数据集说明与实验流程.md for the complete Chinese dataset specification, file counts, leakage constraints, metrics and the recommended experiment workflow.

Key directories:

Directory	Purpose
00_metadata/	Identity mapping, vocabularies and source partitions
01_core/	Full/reference injection, forget and retain training data
02_evaluation/	Text, image and transformed-image evaluation
03_attacks/	Random S/H splits and fixed-H attack-size experiments
04_controls/	Template, identity-access and continuation controls
05_novel_fact_learning/	Allowed new-fact learning evaluation
06_idea3_binding/	A/B × city/profession contrast quadruples
Full_Set_Visual_QA/	Complete 500-identity visual QA annotations

Data format

Most files use JSON Lines records with prompt, answer, image, task, modality, and Qwen-style multimodal messages. Image paths are relative to the repository root.

Validation

The current package contains 54 JSONL files and 23,260 JSONL records. Automated checks cover identity separation, S/H isolation, target-city leakage, prompt separation, image integrity and transformed-image byte equality. See validation_report.json.

Source and usage

This is an unofficial derived research dataset. Users must review and comply with the terms of the upstream MLLMU-Bench dataset. No license is asserted here beyond rights granted by the upstream source and the authors of the newly generated annotations and metadata.

Do not interpret synthetic names or attributes as information about real people. Do not count the core-300 visual_attributes.jsonl and its source records in Full_Set_Visual_QA as independent observations.

Citation

Please cite MLLMU-Bench when using the source images:

@article{liu2024protecting,
  title={Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench},
  author={Liu, Zheyuan and Dou, Guangyao and Jia, Mengzhao and Tan, Zhaoxuan and Zeng, Qingkai and Yuan, Yongle and Jiang, Meng},
  journal={arXiv preprint arXiv:2410.22108},
  year={2024}
}