Datasets:

ArXiv:
License:

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

ReA-Bench

ReA-Bench is a diagnostic benchmark for auditing responsible agent behavior in execution trajectories.

Dataset Access

The ReA-Bench dataset is available upon request. Many of its trajectories are adapted from in-house production and red-team data. Although we have applied extensive de-identification to ensure data safety, we share the dataset through a request-based review rather than a fully public release.

To apply, please email yezipeng.yzp@alibaba-inc.com with the following information:

  • Your HuggingFace or GitHub username (for access approval)
  • Your name and affiliation (an institutional email is preferred)
  • A brief description of your intended use

Approved applicants will be asked to confirm the data use terms (e.g., research-only use and no redistribution) before access is granted.

We are accepting requests. We will contact approved applicants when dataset distribution begins.

About Us

ReA-Bench is developed by Alibaba AAIG, Alibaba's Artificial Intelligence Governance and Sustainable Development Laboratory. We focus on AI safety and governance: using AI to address real-world risks, while making AI itself more safe, reliable, and trustworthy.

Downloads last month
26

Paper for Alibaba-AAIG/ReA-Bench