Describe LLM-as-a-judge for evaluation

  • 7 minutes

When human evaluation is too time-consuming or expensive, you can use another LLM to evaluate your model's outputs. This approach is called LLM-as-a-judge, where one large language model evaluates the quality and performance of another LLM or AI system.

LLM-as-a-judge provides consistent and scalable evaluations. Unlike human evaluators who may have different opinions, an LLM judge applies the same criteria consistently across all evaluations.

Create effective prompts for LLM-as-a-judge

LLM-as-a-judge works by giving the evaluator LLM specific instructions through a prompt template. Here's an example template that evaluates how well a system answers a user's question:

You will be given a `user_question` and `system_answer` pair. Your task is to provide a 'total rating' scoring how well the `system_answer` answers the user concerns expressed in the `user_question`.

Give your answer as a float on a scale of 0 to 10, where 0 means that the `system_answer` is not helpful at all, and 10 means that the answer completely and helpfully addresses the question. 

Provide your feedback as follows: 

*Feedback*
Total rating: (your rating, as a float between 0 and 10)

Now here are the question and answer to evaluate. 

Question: {question} 
Answer: {answer} 

*Feedback* 
Total rating:

This template includes several key components:

  • Clear task definition: The LLM judge knows it needs to rate how well the system answer addresses the user's question
  • Specific scoring scale: A 0-10 float scale with clear endpoints (0 = not helpful, 10 = completely helpful)
  • Structured output format: The "Feedback" and "Total rating:" format ensures consistent responses
  • Input placeholders: {question} and {answer} variables for the actual content to evaluate

The structured approach ensures consistent evaluation and makes it easy to aggregate results across multiple samples.

To improve your LLM-as-a-judge evaluations, include few-shot examples with human-provided scores for guidance. Few-shot examples are sample question-answer pairs with their ideal scores that show the LLM judge what good evaluation looks like. Also define what "good" means for your specific metric and provide a detailed rubric or evaluation scale.

Use MLflow for LLM-as-a-judge evaluation

MLflow provides built-in support for LLM-as-a-judge evaluation in Azure Databricks. You can use MLflow's evaluate module to implement this approach:

  1. Create evaluation records: Define example inputs and expected outputs for your evaluation criteria
  2. Define a metric object: Set up the evaluation framework with your examples, scoring criteria, the LLM judge, and aggregation methods
  3. Run the evaluation: Apply your metric to evaluate models against your reference datasets

MLflow allows you to create custom metrics and run evaluations on your datasets, providing detailed performance assessments.

Tip

Learn more about creating custom LLM-evaluation metrics with MLflow.

Example LLM-as-a-judge metrics

Here are common metrics you can evaluate using LLM-as-a-judge. Each example shows how to structure the evaluation prompt with clear criteria and scoring scales:

Metric Definition
Relevance Evaluate the relevance of the following response to the given query: [Query] - [Response]. Provide a score between 1 and 5, with 5 being highly relevant and 1 being not relevant at all.
Coherence Assess the coherence of the following paragraph. Does it logically flow from one sentence to the next? Provide a score between 1 and 5, with 5 being highly coherent and 1 being not coherent at all.
Accuracy Judge the accuracy of the following statement based on the provided context: [Context] - [Statement]. Provide a score between 1 and 5, with 5 being highly accurate and 1 being not accurate at all.
Fluency Evaluate the fluency of the following text. Does it read naturally and smoothly? Provide a score between 1 and 5, with 5 being highly fluent and 1 being not fluent at all.