Skip to main content

BeaverTails Dataset for LLM Safety Testing

Overview​

The BeaverTails plugin uses the BeaverTails dataset, a dataset of 330,000 prompt injections published by PKU-Alignment to test LLM systems against a comprehensive collection of potentially harmful prompts.

The dataset covers a wide range of harmful content categories including:

  • Animal and human abuse
  • Criminal activities and unethical behavior
  • Discrimination and hate speech
  • Explicit adult content
  • Misinformation and dangerous advice
  • Privacy violations
  • Violence and terrorism

How it works​

The plugin:

  • Fetches test cases from the BeaverTails dataset
  • Filters out safe/benign test cases to focus on potentially harmful prompts
  • Randomly selects prompts from the filtered set
  • Tests if the model appropriately refuses or handles the harmful prompts

Quick Start​

  1. Set your Hugging Face API token as an environment variable:
export HF_TOKEN=your_huggingface_token  # or HF_API_TOKEN
  1. Enable the plugin in your configuration:
promptfooconfig.yaml
redteam:
  plugins:
    - beavertails

Configuration​

You can control the number of test cases using the numTests parameter (defaults to 5):

redteam:
  plugins:
    - id: beavertails
      numTests: 25

Grading​

The plugin uses a specialized grader that evaluates model outputs against a safety rubric. The grader checks if the output:

  • Refuses to engage with harmful content
  • Avoids promoting or assisting with harmful activities
  • Maintains appropriate boundaries
  • Provides safe and ethical responses

Outputs are graded as either SAFE (pass) or COMPLICIT (fail) based on these criteria.