shichenghu/personal-info-unlearning

Synthetic Personal Information Unlearning Dataset

Dataset Description

This dataset is designed for research on large language model (LLM) unlearning in controlled synthetic personal-information settings.

It contains synthetic profiles and question-answer data for four personal attributes:

• Year of birth
• Blood type
• Postcode
• Social insurance number

The benchmark provides three forget-set sizes: N = 5, 20, 40.

All personal-profile data are synthetically generated for studying memorization, unlearning, information recovery, and privacy auditing.

Dataset Configurations

The dataset contains 12 configurations:

Configuration	N	Target Attribute
n5_year_of_birth	5	Year of Birth
n5_blood_type	5	Blood Type
n5_postcode	5	Postcode
n5_social_insurance_number	5	Social Insurance Number
n20_year_of_birth	20	Year of Birth
n20_blood_type	20	Blood Type
n20_postcode	20	Postcode
n20_social_insurance_number	20	Social Insurance Number
n40_year_of_birth	40	Year of Birth
n40_blood_type	40	Blood Type
n40_postcode	40	Postcode
n40_social_insurance_number	40	Social Insurance Number

Example:

from datasets import load_dataset

dataset = load_dataset(
    "<USERNAME>/<DATASET_NAME>",
    "n20_year_of_birth"
)

Splits

Each configuration contains:

Split	Description
forget	Data targeted for unlearning.
retain	Data retained during unlearning.
remain	Remaining data outside the forget and retain sets.
retain_same_attr	Attribute-matched retain subset.
remain_same_attr	Attribute-matched remaining subset.
retain_same_fn_attr	Specialized matched retain subset.
remain_same_fn_attr	Specialized matched remaining subset.
test_forget	Test data corresponding to the forget set.
test_retain	Test data corresponding to the retain set.
test_remain	Test data corresponding to the remaining set.
test_retain_same_attr	Attribute-matched retain test subset.
test_remain_same_attr	Attribute-matched remaining test subset.
test_retain_same_fn_attr	Specialized matched retain test subset.
test_remain_same_fn_attr	Specialized matched remaining test subset.

The construction of matched subsets is provided in the accompanying codebase.

Base Data

Shared files are stored under data/base/.

File	Description
profiles.json	Synthetic profiles.
training_dataset.json	Main training dataset.
training_testset.json	Training test set.
validation_dataset.json	Validation dataset.
validation_testset.json	Validation test set.
common_knowledge_questions.json	Common-knowledge evaluation data.
real_world_dataset.json	Auxiliary real-world evaluation data.
idk.jsonl	Auxiliary "I don't know" responses.

Repository Structure

data/
├── base/
├── n5/
│   ├── year_of_birth/
│   ├── blood_type/
│   ├── postcode/
│   └── social_insurance_number/
├── n20/
│   ├── year_of_birth/
│   ├── blood_type/
│   ├── postcode/
│   └── social_insurance_number/
└── n40/
    ├── year_of_birth/
    ├── blood_type/
    ├── postcode/
    └── social_insurance_number/

Usage

from datasets import load_dataset

dataset = load_dataset(
    "<USERNAME>/<DATASET_NAME>",
    "n20_postcode"
)

print(dataset)
print(dataset["forget"][0])

Configuration names follow:

n{5,20,40}_{year_of_birth,blood_type,postcode,social_insurance_number}

Intended Use

This dataset is intended for research on:

• LLM unlearning
• memorization and knowledge retention
• privacy auditing
• recovery attacks
• forget quality and model utility evaluation

Synthetic Data Notice

The personal-profile data are synthetically generated and do not represent real individuals.

Auxiliary real-world and common-knowledge evaluation files may contain publicly available factual information.

Limitations

This benchmark evaluates controlled factual memorization and unlearning. Results should not be interpreted as formal privacy guarantees or proof of effective deletion in arbitrary real-world settings.