🔐 MAMA-SPIRIT

Synthetic PII Role-based Interaction Tasks for Measuring Memory Leakage in Multi-Agent LLMs

SPIRIT (Synthetic PII Role-based Interaction Tasks) is the synthetic PII evaluation dataset released with MAMA (Multi-Agent Memory Attack), a framework for studying how communication topology affects private-memory leakage in multi-agent LLM systems. The dataset is hosted on Hugging Face in this MAMA-SPIRIT repository.

The dataset accompanies our Findings of ACL 2026 paper:

Topology Matters: Measuring Memory Leakage in Multi-Agent LLMs

---

📊 Dataset at a Glance

Instances	484
PII annotations	2,074
PII types	41
Language	English
Data type	Synthetic
Format	CSV
Task	Multi-agent privacy leakage evaluation
License	Apache-2.0

All PII contained in the dataset is synthetic and is not intended to correspond to real individuals or valid credentials.

---

🚀 Quick Start

Install the Hugging Face datasets library:

pip install datasets

Load MAMA-SPIRIT directly from the Hub:

from datasets import load_dataset

dataset = load_dataset(
    "liu1111121/MAMA-SPIRIT",
    split="test"
)

print(dataset)
print(dataset[0])

No manual download or extraction is required.

---

🧩 What Is SPIRIT?

SPIRIT provides synthetic private documents containing annotated Personally Identifiable Information (PII), together with generated public tasks designed for controlled multi-agent interaction.

A typical evaluation setting separates information into two components:

Component	Role
🔒 Private memory	A synthetic document containing annotated PII, visible only to the target agent
🌐 Public task	A generated background and question used by agents during interaction

This structure enables controlled measurement of whether private information stored in one agent's memory propagates through a multi-agent communication network.

MAMA uses these instances to investigate how properties such as network topology, communication distance, and agent centrality affect memory leakage.

---

🗂️ Dataset Structure

Each row contains five fields:

Field	Description
text	Synthetic PII-bearing document used as the target agent's private memory
pii	Ground-truth PII annotations, stored as a serialized list of entity/type dictionaries
generated_texts	Complete generated public task description
task_backgrounds	Generated public background context
questions	Generated task question

A SPIRIT instance consists of two components:

Component	Dataset fields	Role
🔒 Private Memory	text, pii	Synthetic PII-bearing memory available to the target agent
🌐 Public Task	task_backgrounds, questions, generated_texts	Public interaction context provided to the agents

---

🏗️ Dataset Construction

SPIRIT is derived from the Gretel Synthetic Domain-Specific Documents Dataset (English):

🤗 gretelai/gretel-pii-masking-en-v1

The source dataset provides synthetic documents containing labeled PII entities.

For SPIRIT, these synthetic private documents are paired with LLM-generated task backgrounds and questions to construct role-based interaction scenarios for multi-agent evaluation.

The resulting benchmark therefore combines:

Synthetic PII-bearing document
            +
Generated public task context
            +
Generated question
            ↓
        SPIRIT instance

The original Gretel dataset is distributed under the Apache License 2.0.

---

🔬 Intended Use

SPIRIT is designed for research on:

• multi-agent LLM systems
• privacy and information leakage
• memory propagation
• communication topology
• agent security
• privacy-preserving agent architectures
• controlled evaluation of private-memory exposure

The dataset is intended for authorized research and evaluation.

---

📄 Paper

Topology Matters: Measuring Memory Leakage in Multi-Agent LLMs

Jinbo Liu, Defu Cao, Yifei Wei, Tianyao Su, Yuan Liang, Yushun Dong, Yan Liu, Yue Zhao, and Xiyang Hu

Findings of the Association for Computational Linguistics: ACL 2026, pp. 39728–39746.

Resource	Link
📘 ACL Anthology	2026.findings-acl.1980
📄 PDF	ACL Anthology PDF
📝 arXiv	arXiv:2512.04668
💻 Code	github.com/llll121/mama-eval
🤗 Dataset	liu1111121/MAMA-SPIRIT

---

💻 Code

The reference implementation of MAMA is available on GitHub:

➡️ MAMA GitHub Repository

The repository contains the evaluation pipeline and code for reproducing MAMA experiments.

---

🛡️ Ethical Considerations

SPIRIT is specifically designed for controlled privacy-leakage research.

All documents and annotated PII entities in the dataset are synthetic.

Some synthetic entities imitate the structure of sensitive information, including identifiers, addresses, account information, and credentials. These entities are included because realistic PII formats are necessary for evaluating leakage behavior.

They are not intended to:

• represent real individuals,
• provide valid credentials,
• provide access to real accounts or services.

The dataset should not be used to target real individuals, systems, or services without authorization.

---

📚 Citation

If you use SPIRIT or the MAMA evaluation framework in your research, please cite:

@inproceedings{liu-etal-2026-topology,
    title = "Topology Matters: Measuring Memory Leakage in Multi-Agent {LLM}s",
    author = "Liu, Jinbo and
              Cao, Defu and
              Wei, Yifei and
              Su, Tianyao and
              Liang, Yuan and
              Dong, Yushun and
              Liu, Yan and
              Zhao, Yue and
              Hu, Xiyang",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.1980/",
    doi = "10.18653/v1/2026.findings-acl.1980",
    pages = "39728--39746"
}

---

📜 License

SPIRIT is distributed under the Apache License 2.0.

The source MAMA implementation is maintained separately in the GitHub repository under its repository license.