huseyinatahaninan/ContextualIntegritySyntheticDataset

Contextual Integrity Synthetic Dataset

This repository contains the synthetic dataset introduced in the paper "Contextual Integrity in LLMs via Reasoning and Reinforcement Learning".

Paper | Code | Blog

Dataset Summary

The Contextual Integrity (CI) synthetic dataset consists of 729 examples featuring diverse contexts and information disclosure norms. It is designed to instill reasoning capabilities in LLMs regarding what information is appropriate to share while carrying out specific tasks.

As autonomous agents increasingly make decisions on behalf of users, ensuring they respect privacy boundaries (Contextual Integrity) becomes critical. This dataset provides vignettes where an agent must reason about the context to decide whether specific information should be disclosed. The authors demonstrate that training on this data using reinforcement learning (RL) substantially reduces inappropriate information disclosure while maintaining task performance.

Dataset Structure

The dataset contains the following features:

• seed: A unique identifier/string related to the scenario generation.
• dataset_item: The synthetic content including the prompt, context, and expected behavior regarding information sharing.

Citation

If you use this dataset in your research, please cite the following paper:

@inproceedings{lan2025contextual,
  title={Contextual Integrity in LLMs via Reasoning and Reinforcement Learning},
  author={Guangchen Lan and Huseyin A. Inan and Sahar Abdelnabi and Janardhan Kulkarni and Lukas Wutschitz and Reza Shokri and Christopher G. Brinton and Robert Sim},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025}
}