Skip to main content
Last updated on

Align judges with humans

Judge alignment teaches LLM judges to match human evaluation standards through systematic feedback. This process transforms generic evaluators into domain-specific experts that understand your unique quality criteria, improving agreement with human assessments by 30 to 50 percent compared to baseline judges.

The same alignment workflow applies to both built-in judges (such as RelevanceToQuery, Safety, or Correctness) and custom judges created with make_judge(). Use alignment with built-in judges to adapt their generic criteria to your domain, or with custom judges to refine specialized evaluation logic.

Judge alignment follows a three-step workflow:

  1. Generate initial assessments: Use a built-in or custom judge to evaluate traces and establish a baseline.
  2. Collect human feedback: Domain experts review and correct judge assessments.
  3. Align and deploy: Invoke the judge's align() method to create a new judge that is more aligned with human feedback.

The system supports the optimizers that are available in the package mlflow.genai.judges.optimizers.

Requirements​

  • MLflow 3.4.0 or above to use judge alignment features

    Python
    %pip install --upgrade "mlflow[databricks]>=3.4.0" databricks_openai dspy
    dbutils.library.restartPython()

  • A judge to align. This can be a built-in judge (for example, RelevanceToQuery or Correctness) or a custom judge created with make_judge().

  • The human feedback assessment name must exactly match the judge's name attribute. For built-in judges, this is the default snake_case name (for example, relevance_to_query for RelevanceToQuery) unless you override it by passing name= when instantiating the class. For custom judges, it's the name you passed to make_judge() (for example, product_quality).

  • Alignment is not supported for session-level (multi-turn) judges such as ConversationCompleteness.

Step 1: Set up the judge and generate traces​

Set up your initial judge and generate traces with assessments. You can achieve reasonable alignment with at least 10 traces, but 50-100 traces yield better results.

Instantiate a built-in judge directly. Built-in judges expose a name attribute (the default is a snake_case string such as relevance_to_query) that you'll use when logging human feedback in Step 2.

Python
from mlflow.genai.scorers import RelevanceToQuery
import mlflow

# Create or set an MLflow experiment for alignment.
# Use a workspace path such as /Shared/<name> or /Users/<your-email>/<name>.
experiment = mlflow.set_experiment("/Shared/relevance-alignment")
experiment_id = experiment.experiment_id

# Use a built-in judge
initial_judge = RelevanceToQuery()

Define your application logic. The following example uses a Databricks-hosted foundation model to generate a product description from a query. Replace this with your own application code:

Python
import mlflow
from databricks_openai import DatabricksOpenAI

# Enable automatic tracing of OpenAI calls
mlflow.openai.autolog()

# Create an OpenAI client connected to Databricks-hosted LLMs
client = DatabricksOpenAI()
model_name = "databricks-claude-sonnet-4"


def generate_product_description(query: str) -> str:
    response = client.chat.completions.create(
        model=model_name,
        messages=[
            {
                "role": "system",
                "content": "You write concise, accurate product descriptions.",
            },
            {"role": "user", "content": query},
        ],
    )
    return response.choices[0].message.content

Generate traces and run the judge. Use the judge's name attribute (for example, relevance_to_query for the built-in judge above, or product_quality for the custom judge above) as the feedback name:

Python
# Generate traces for alignment (minimum 10, recommended 50+)
for i in range(50):
    query = f"Tell me about product {i}"
    description = generate_product_description(query)

    # Retrieve the ID of the most recent finished trace
    trace_id = mlflow.get_last_active_trace_id()
    trace = mlflow.get_trace(trace_id)

    # Generate judge assessment
    judge_result = initial_judge(trace=trace)

    # Log judge feedback to the trace using the judge's name
    mlflow.log_feedback(
        trace_id=trace_id,
        name=initial_judge.name,
        value=judge_result.value,
        rationale=judge_result.rationale,
    )

Step 2: Collect human feedback​

Collect human feedback to teach the judge your quality standards. Choose from the following approaches:

Collect human feedback when:

  • You need domain experts to review outputs.
  • You want to iteratively refine feedback criteria.
  • You're working with a smaller dataset (< 100 examples).

Use the MLflow UI to manually review and provide feedback:

  1. Navigate to your MLflow experiment in the Databricks workspace.
  2. Click on the Traces tab to see traces.
  3. Review each trace and its judge assessment.
  4. Add human feedback using the UI's feedback interface.
  5. Ensure the feedback name matches your judge's name attribute exactly (for example, relevance_to_query for a built-in RelevanceToQuery instance or product_quality for the custom judge above).

Best practices for feedback collection​

  • Diverse reviewers: Include multiple domain experts to capture varied perspectives
  • Balanced examples: Include at least 30% negative examples (poor/fair ratings)
  • Clear rationales: Provide detailed explanations for ratings
  • Representative samples: Cover edge cases and common scenarios

Step 3: Align and register the judge​

Once you have sufficient human feedback, align the judge. The same align() method is used for both built-in and custom judges.

When you call align() without specifying an optimizer, the MemAlign optimizer is used automatically:

Python
# Retrieve traces with both judge and human assessments
traces_for_alignment = mlflow.search_traces(
    experiment_ids=[experiment_id],
    max_results=100,
    return_type="list"
)

if len(traces_for_alignment) >= 10:
    # Align the judge based on human feedback using the default optimizer
    aligned_judge = initial_judge.align(traces_for_alignment)

    # Register the aligned judge for production use.
    # Use a new name to distinguish it from the original judge.
    aligned_judge.register(
        experiment_id=experiment_id,
        name=f"{initial_judge.name}_aligned",
        tags={"alignment_date": "2025-10-23", "num_traces": str(len(traces_for_alignment))}
    )

    print(f"Successfully aligned judge using {len(traces_for_alignment)} traces")
else:
    print(f"Insufficient traces for alignment. Found {len(traces_for_alignment)}, need at least 10")

Enable detailed logging​

To monitor the alignment process, enable debug logging for the optimizer:

Python
import logging

# Enable detailed logging
logging.getLogger("mlflow.genai.judges.optimizers.memalign").setLevel(logging.DEBUG)

# Run alignment with verbose output
aligned_judge = initial_judge.align(traces_for_alignment)

Validate alignment​

Validate that alignment improved the judge:

Python

def test_alignment_improvement(
    original_judge, aligned_judge, test_traces: list
) -> dict:
    """Compare judge performance before and after alignment."""

    original_correct = 0
    aligned_correct = 0

    for trace in test_traces:
        # Get human ground truth from trace assessments
        feedbacks = trace.search_assessments(type="feedback")
        human_feedback = next(
            (f for f in feedbacks if f.source.source_type == "HUMAN"), None
        )

        if not human_feedback:
            continue

        # Get judge evaluations
        # Judges can evaluate entire traces instead of individual inputs/outputs
        original_eval = original_judge(trace=trace)
        aligned_eval = aligned_judge(trace=trace)

        # Check agreement with human
        if original_eval.value == human_feedback.value:
            original_correct += 1
        if aligned_eval.value == human_feedback.value:
            aligned_correct += 1

    total = len(test_traces)
    return {
        "original_accuracy": original_correct / total,
        "aligned_accuracy": aligned_correct / total,
        "improvement": (aligned_correct - original_correct) / total,
    }


Create custom alignment optimizers​

For specialized alignment strategies, extend the AlignmentOptimizer base class:

Python
from mlflow.genai.judges.base import AlignmentOptimizer, Judge
from mlflow.entities.trace import Trace

class MyCustomOptimizer(AlignmentOptimizer):
    """Custom optimizer implementation for judge alignment."""

    def __init__(self, model: str = None, **kwargs):
        """Initialize your optimizer with custom parameters."""
        self.model = model
        # Add any custom initialization logic

    def align(self, judge: Judge, traces: list[Trace]) -> Judge:
        """
        Implement your alignment algorithm.

        Args:
            judge: The judge to be optimized
            traces: List of traces containing human feedback

        Returns:
            A new Judge instance with improved alignment
        """
        # Your custom alignment logic here
        # 1. Extract feedback from traces
        # 2. Analyze disagreements between judge and human
        # 3. Generate improved instructions
        # 4. Return new judge with better alignment

        # Example: Return judge with modified instructions
        from mlflow.genai.judges import make_judge

        improved_instructions = self._optimize_instructions(judge.instructions, traces)

        return make_judge(
            name=judge.name,
            instructions=improved_instructions,
            model=judge.model,
        )

    def _optimize_instructions(self, instructions: str, traces: list[Trace]) -> str:
        """Your custom optimization logic."""
        # Implement your optimization strategy
        pass

# Create your custom optimizer
custom_optimizer = MyCustomOptimizer(model="your-model")

# Use it for alignment
aligned_judge = initial_judge.align(traces_with_feedback, custom_optimizer)

Limitations​

  • Judge alignment does not support agent-based or expectation-based evaluation.

Next steps​

On this page