Azure ml model evaluation

Azure ML Model Evaluation

Evaluate generative AI applications using Azure AI Evaluation SDK with built-in quality and safety metrics. Local or cloud-based evaluation integrated with CI/CD pipelines.

Three evaluation approaches:

1. Local Evaluation — Run evaluations on your machine with fast iteration
2. Cloud Evaluation — Scale to large datasets on Azure compute
3. Continuous Monitoring — Post-deployment evaluation for production applications

Use this skill when:

• Evaluating LLM responses for quality (relevance, coherence, groundedness)
• Assessing safety risks (hate, violence, sexual, self-harm)
• Comparing models or prompts
• Tracking evaluation metrics over time
• Building CI/CD gates with quality thresholds

Prerequisites

Local Evaluation

• Packages: azure-ai-evaluation, azure-identity
• Azure OpenAI deployment for AI-assisted evaluators
• Test dataset in JSONL format

Cloud Evaluation

• Azure AI Foundry hub-based project
• Azure OpenAI deployment with chat completion
• Connected storage account for datasets
• CLI: az login

Safety Evaluators

• Project in East US 2, France Central, UK South, or Sweden Central

Template Files

These are templates in examples/ directory. Copy and adapt them for your project:

examples/
  ├── local_evaluation.py                    # Template: Evaluate with built-in metrics
  ├── cloud_evaluation.py                    # Template: Cloud-scale evaluation job
  └── utils.py                               # Template: Helper functions

Do NOT reference these files directly. Copy and adapt them for your project structure.

Quick Start

Local Evaluation

1. Copy examples/local_evaluation.py and examples/utils.py to your project
2. Run: python local_evaluation.py
3. Runs 5 evaluators locally (relevance, groundedness, coherence, fluency, F1 score) on test_data.jsonl
4. Output: evaluation_results.json

Cloud Evaluation

1. Copy examples/cloud_evaluation.py and examples/utils.py to your project
2. Run: python cloud_evaluation.py
3. Submits evaluation job to Azure; monitor via Azure AI Foundry → Evaluation runs
4. Includes relevance, groundedness, violence, coherence evaluators

Data Formats

Single-Turn Q&A (JSONL)

{"query": "What is Azure ML?", "response": "Azure ML is...", "context": "...", "ground_truth": "..."}

Multi-Turn Conversation (JSONL)

{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

Multi-Modal (Image + Text)

{
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "What is in this image?"},
        {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,..."}}
      ]
    }
  ]
}

Built-In Evaluators

Quality Metrics (Require Azure OpenAI GPT-4)

Evaluator	Inputs	Desc
GroundednessEvaluator	query, response, context	Response supported by context
RelevanceEvaluator	query, response	Response addresses query
CoherenceEvaluator	query, response	Logical flow and clarity
FluencyEvaluator	query, response	Language quality
RetrievalEvaluator	query, context	Context relevance to query
IntentResolutionEvaluator	conversation	User intent resolved
TaskAdherenceEvaluator	conversation	Adherence to instructions

NLP Metrics (No model required)

Evaluator	Inputs	Desc
F1ScoreEvaluator	response, ground_truth	Token overlap F1
SimilarityEvaluator	response, ground_truth	Cosine embedding similarity
BleuScoreEvaluator	response, ground_truth	Translation quality
RougeScoreEvaluator	response, ground_truth	Summarization quality
MeteorScoreEvaluator	response, ground_truth	Semantic similarity

Safety Metrics (Require Azure AI project)

Evaluator	Desc	Severity
ViolenceEvaluator	Violent content	Very Low / Low / Med / High
SexualEvaluator	Sexual content	Very Low / Low / Med / High
SelfHarmEvaluator	Self-harm content	Very Low / Low / Med / High
HateUnfairnessEvaluator	Hate/discrimination	Very Low / Low / Med / High
IndirectAttackEvaluator	XPIA jailbreak attempts	True / False
ProtectedMaterialEvaluator	Copyrighted content	True / False
ContentSafetyEvaluator	Composite safety eval	Combined metrics

Local Evaluation Example

See examples/local_evaluation.py for complete implementation with quality, NLP, and similarity evaluators.

Safety Evaluation Example

See examples/local_evaluation.py for safety evaluator setup and execution.

Cloud Evaluation Example

See examples/cloud_evaluation.py for Azure cloud-based evaluation with dataset upload and evaluator configuration.

Custom Evaluators

Code-Based Custom Evaluator

Define evaluators as Python functions with @tool decorator. Functions receive inputs and return dict with score/metric.

Prompt-Based Custom Evaluator (Prompty)

Create .prompty YAML files with model config and evaluation prompt. Load via Prompty.load() and pass to evaluate.

See examples/local_evaluation.py for custom evaluator patterns and integration.

Composite Evaluators

• QAEvaluator: Combines Groundedness + Relevance + Coherence + Fluency + Similarity + F1
• ContentSafetyEvaluator: Combines Violence + Sexual + SelfHarm + HateUnfairness

See examples/local_evaluation.py for composite evaluator usage.

Evaluation with Target Application

Pass a callable target function to evaluate() to automatically generate responses. Function receives query and returns response dict. See examples/local_evaluation.py for implementation.

Integration with Training

Pre-Training Baseline

Evaluate baseline model to establish metric baseline for comparison.

Post-Training Comparison

Evaluate fine-tuned model on same data and compare metrics to baseline.

CI/CD Gate

Add quality gates by checking evaluation metrics against thresholds before deployment.

See examples/local_evaluation.py for evaluation and metric comparison patterns.

Results Structure

{
    "metrics": {
        "relevance": 4.5,
        "groundedness": 4.8,
        "coherence": 4.6,
        "fluency": 4.9,
        "f1_score": 0.92,
    },
    "rows": [
        {
            "inputs.query": "...",
            "inputs.response": "...",
            "outputs.relevance.relevance": 5.0,
            "outputs.groundedness.groundedness": 4.5,
        },
        ...
    ],
    "studio_url": "https://ai.azure.com/...",
}

View results in Azure AI Foundry: Navigate to Evaluation → Evaluation runs and click your run ID.

Notes

• Model selection: Use GPT-5-mini for cost-effective evaluation
• Batch size: Evaluate 100-500 samples for statistical significance
• Data quality: Validate JSONL format and column mappings
• Token usage: Monitor Azure OpenAI quota—evaluators consume 800-3000 tokens each
• Regional limits: Safety evaluators only in East US 2, France Central, UK South, Sweden Central
• Privacy: Evaluation data may be logged—sanitize PII before submission

Common Issues

Cloud Evaluation Stuck in "Running" → Azure OpenAI model lacks capacity; cancel job, increase capacity, retry

"Model not found" Error → Verify deployment exists: az cognitiveservices account deployment list

Safety Evaluator "Region not supported" → Create project in East US 2, France Central, UK South, or Sweden Central

"Storage account not connected" → Follow storage account setup

References

• Azure AI Evaluation SDK
• Quality Evaluators
• Safety Evaluators
• Custom Evaluators