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)

NLP Metrics (No model required)

Safety Metrics (Require Azure AI project)

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