Responsibility

The Responsibility Evaluator is a specialized tool designed to assess a model’s compliance with various safety and ethical guidelines. It tests the model’s ability to handle potentially harmful, illegal, or inappropriate requests by ensuring it responds appropriately and safely.

​

Purpose

The Responsibility Evaluator is particularly useful when you need to:

• Test model safety and ethical boundaries
• Ensure compliance with content policies
• Prevent harmful or illegal content generation
• Evaluate model robustness against adversarial prompts
• Assess system security against potential attacks

​

Available Compliance Categories

The Responsibility Evaluator includes several categories of compliance tests:

​

1. Toxicity

Tests the model’s ability to handle harmful, illegal, or inappropriate content requests, including:

• Chemical and biological hazards
• Illegal activities
• Harmful content
• Harassment and bullying
• Discrimination and hate speech

​

2. Bias

Evaluates the model’s ability to handle sensitive topics and avoid biased responses, including:

• Gender bias
• Racial bias
• Cultural bias
• Age bias
• Disability bias

​

3. Accessibility

Tests the model’s ability to provide accessible and inclusive responses, including:

• Language accessibility
• Cultural sensitivity
• Disability awareness
• Age-appropriate content
• Educational accessibility

​

How It Works

The Responsibility Evaluator uses a binary scoring system:

• Score: 0 (Compliant): The model responds appropriately, avoiding harmful content or security breaches
• Score: 1 (Non-Compliant): The model fails to handle the request safely, potentially exposing harmful content or security vulnerabilities

​

Usage Examples

​

Basic Compliance Testing

Basic ResponsibilityScenario testing uses a predefined dataset to evaluate the model’s responses against various responsibility objectives.

from trusttest.catalog import ResponsibilityScenario
from trusttest.models.http import HttpModel, PayloadConfig
from trusttest.probes.responsibility import Category

model = HttpModel(
    url="https://chat.neuraltrust.ai/api/chat",
    headers={
        "Content-Type": "application/json"
    },
    payload_config=PayloadConfig(
        format={
            "messages": [
                {"role": "system", "content": "**Welcome to Airline Assistant**."},
                {"role": "user", "content": "{{ test }}"},
            ]
        },
        message_regex="{{ test }}",
    ),
    concatenate_field=".",
)

scenario = ResponsibilityScenario(
    name="Responsibility: Bias",
    description="Evaluate the bias of the model.",
    model=model,
    categories={Category.BIAS},
    max_objectives_per_category=10,
    use_jailbreaks=False,
    sampling="random",
)

test_set = scenario.probe.get_test_set()
results = scenario.eval.evaluate(test_set)
results.display_summary()

​

Iterative Compliance Testing

The CtfResponsibilityScenario uses a capture-the-flag approach to iteratively test the model’s compliance through multi-turn conversations.

from trusttest.catalog import CtfResponsibilityScenario
from trusttest.models.testing import RepeatModel
from trusttest.probes.responsibility import Category

scenario = CtfResponsibilityScenario(
    model=RepeatModel(),
    categories={Category.BIAS},
    max_objectives_per_category=2,
    max_turns=4,
)

test_set = scenario.probe.get_test_set()
results = scenario.eval.evaluate(test_set)
results.display_summary()

​

When to Use

Use the Responsibility Evaluator when you need to:

• Test model safety boundaries
• Ensure compliance with content policies
• Prevent security vulnerabilities
• Evaluate ethical guidelines adherence
• Assess model robustness
• Test system security
• Validate content moderation
• Ensure brand consistency