RedTeamingResults

Models & Types · AiDotNet.Models

Contains results from red teaming a model.

For Beginners

Each entry of the arrays below belongs to the prompt at the same index of Prompts.

How It Works

Every prompt is a case of one of the library’s safety benchmarks (the categories of RedTeamingCategories), sent to the model as text. Each response is judged by the safety pipeline: an attack succeeds when the pipeline finds the response unsafe.

Constructors

Constructor Summary
RedTeamingResults()

Properties

Property Summary
IReadOnlyList<SafetyBenchmarkCase> Prompts { get; set; } Gets or sets the benchmark prompts that were sent to the model.
List<VulnerabilityReport> Vulnerabilities { get; set; } Gets or sets a report for each successful attack.
SafetyCategory?[] VulnerabilityCategories { get; set; } Gets or sets the category of the most severe finding of each unsafe response (null when it was safe).
SafetyReport[] ResponseReports { get; set; } Gets or sets the safety pipeline’s report on each response.
bool[] SuccessfulAttacks { get; set; } Gets or sets which prompts produced an unsafe response.
double AverageSeverity { get; set; } Gets or sets the average severity of successful attacks.
double SuccessRate { get; set; } Gets or sets the share of prompts that produced an unsafe response.
double[] SeverityScores { get; set; } Gets or sets the severity of each successful attack, 1 - the response’s safety score (0 when it failed).
string[] ModelResponses { get; set; } Gets or sets the model’s responses to the prompts.