rhesis/Insurance-Chatbot-Physical-Harm-and-Violence-Jailbreak

Dataset Card for Physical Harm and Violence Jailbreak

Description

The test set has been specifically designed to evaluate the performance and robustness of an insurance chatbot in the insurance industry. The main focus of this test set is to assess the chatbot's ability to handle challenging scenarios and difficult behaviors. In particular, it aims to determine the chatbot's effectiveness in addressing issues related to jailbreak, while also encompassing topics related to physical harm and violence. By subjecting the chatbot to various problematic situations, this test set aims to provide comprehensive insights into its capabilities and limitations, ensuring its optimal performance in real-world applications.

Structure

The dataset includes the following columns:

• ID: The unique identifier for the prompt.
• Behavior: The performance dimension evaluated (Reliability, Robustness, or Compliance).
• Topic: The topic validated as part of the prompt.
• Category: The category of the insurance-related task, such as claims, customer service, or policy information.
• Demographic [optional]: The demographic of the test set (only if contains demographic prompts, e.g., in compliance tests).
• Expected Response [optional]: The expected response from the chatbot (only if contains expected responses, e.g., in reliability tests).
• Prompt: The actual test prompt provided to the chatbot.
• Source URL: Provides a reference to the source used for guidance while creating the test set.

Usage

This dataset is specifically designed for evaluating and testing chatbots, including customer-facing ones, in the context of handling different scenarios. It focuses on a single critical aspect — physical harm and violence jailbreak — and provides insights into how well a chatbot can identify and address fraudulent activities. However, we encourage users to explore our other test sets to assess chatbots across a broader range of behaviors and domains. For a comprehensive evaluation of your application, you may want to consider using a combination of test sets to fully understand its capabilities and limitations. To evaluate your chatbot with this dataset or for further inquiries about our work, feel free to contact us at: [email protected].

Sources

To create this test set, we relied on the following source(s):

• Shen, X., Chen, Z., Backes, M., Shen, Y., & Zhang, Y. (2023). " Do Anything Now": Characterizing and evaluating in-the-wild jailbreak prompts on large language models. arXiv preprint arXiv:2308.03825.

Citation

If you use this dataset, please cite:

@inproceedings{rhesis,
  title={Rhesis - A Testbench for Evaluating LLM Applications. Test Set: Physical Harm and Violence Jailbreak},
  author={Rhesis},
  year={2024}
}