A Judge is an automated evaluator, often powered by a Large Language Model (LLM), that assesses the quality of responses or content within the platform. It serves as an intelligent system that can analyze inputs based on specific criteria and provide feedback or scores to indicate the response’s quality or correctness. Judges are used in various contexts such as competitions, assessments, and automated content moderation.

​

Key Functions of a Judge

1. Safety Judge

• Evaluates responses to ensure they adhere to safety and ethical guidelines.
• Detects harmful, inappropriate, or unsafe content such as offensive language, hate speech, or personal information breaches.
• Often used in platforms where user interactions or submissions need to be moderated for safe consumption.

2. Reliability Judge

• Assesses the factual correctness of a response.
• Detects whether the response contains invented or fabricated information, a common issue in LLM-generated content.
• Typically used in environments where factual accuracy is critical, such as in educational settings, research, or content generation.

3. Custom Judges

• Tailored to specific evaluation criteria or use cases.
• Can be designed to assess responses based on unique requirements, industry standards, or user-defined metrics.

​

Steps to Create a Judge

​

1. Select A Space

​

2. Go to the Judge Section

​

3. Judges Dashboard

This page displays all the judges created within the space. You can view, edit,test or delete existing judges and create new judges from this dashboard.

​

4. Create New Judge

Click on the “Create New Judge” button to start the process of creating a new judge.

​

5. Configure a Judge

The first step in configuring a judge is to determine the type of judge that best suits your evaluation needs. Our platform supports three types of judges:

• Safety Judges
• Hallucination Judges
• Custom Judges

Understanding each type will help you select the most appropriate judge for your application.

​

Safety Judges

Safety Judges are designed to detect and mitigate harmful content in generated responses. They play a crucial role in ensuring that the outputs from language models adhere to safety guidelines and do not contain inappropriate or dangerous material.

​

Use Cases for Safety Judges

• Content Moderation: Filtering out offensive language, hate speech, or harassment to maintain a respectful environment.
• Regulatory Compliance: Ensuring that generated content complies with legal standards and industry regulations.
• User Protection: Preventing the dissemination of violent, or harmful information that could distress users.

Click here for more information on adding a Safety Judge.

​

Reliability Judges

Reliability Judges focus on detecting factual inaccuracies in generated responses. Language models sometimes produce information that appears plausible but is incorrect or fabricated—this phenomenon is known as “hallucination.”

​

Use Cases for Reliability Judges

• Information Services: Providing accurate data in applications like virtual assistants or information kiosks.
• Educational Content: Ensuring correctness in learning materials or explanations.
• News and Reports: Delivering reliable information in journalism or analytical reports.

Click here for more information on adding a Reliability Judge.

​

Custom Judges

Custom Judges offer the flexibility to define evaluation criteria tailored to your specific needs. They are ideal when standard safety or factual correctness checks are insufficient for your application.

​

Use Cases for Custom Judges

• Industry-Specific Standards: Assessing compliance with protocols in fields like healthcare, finance, or legal services.
• Brand Voice and Tone: Ensuring that responses align with your organization’s branding guidelines.
• Creative and Stylistic Requirements: Evaluating originality, style, or adherence to a particular writing convention.

Click here for more information on adding a Custom Judge.