yonglixiang/AntiSkillBench

Dataset Card

Overview

This dataset is proposed in our work, "When Agents Learn to Be You: Benchmarking Privacy Leakage, Impersonation Risk, and Defenses in Persona Skills". It includes simulated persona grounded user–assistant dialogues as personal traces for skill distillation, consists of:

1. sync_characters.jsonl contains 50 character profiles.
2. sync_questions.jsonl contains 2,500 character-grounded questions. Each character has 50 questions: 30 general, 10 math, and 10 tool-use questions.
3. sync_dialogues.jsonl contains multi-turn dialogues built from those questions. GPT-5.4 simulates the user, while GPT-5.4-mini simulates the assistant with more concise responses.

Detailed File Descriptions

sync_characters.jsonl

Each record describes one character, including a unique ID, a persona summary, personality and language-style descriptions, a longer natural-language character description, and structured profile attributes such as age, gender, occupation, and education.

sync_questions.jsonl

Each record contains character_id, question_id, question_type, and question. Question types are labeled general_question, math_question, and tool_question.

sync_dialogues.jsonl

Each record contains character_id, the source question, and a dialogue list. Every dialogue turn has a role and content field.

Citation

If you find this dataset useful for your research, please consider citing our paper:

@misc{xiang2026antiskillbench,
  title={When Agents Learn to Be You: Benchmarking Privacy Leakage, Impersonation Risk, and Defenses in Persona Skills},
  author={Yongli Xiang and Zhifang Zhang and Bojun Yang and Ziming Hong and Lei Feng and Miao Xu and Tongliang Liu},
  year={2026},
  eprint={2608.03700},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2608.03700}, 
}