Evaluate the Fine-tuned Phi-3 / Phi-3.5 Model in Azure AI Studio Focusing on Microsoft’s Responsible

Spoiler

This blog series has several versions, each covering different aspects and techniques. Check out the following resources:

Fine-Tune and Integrate Custom Phi-3 Models with Prompt Flow: Step-by-Step Guide
Detailed instructions for fine-tuning and integrating custom Phi-3 models with Prompt flow using a code-first approach.
Available on:
Fine-Tune and Integrate Custom Phi-3 Models with Prompt Flow in Azure AI Studio
Detailed instructions for fine-tuning and integrating custom Phi-3 models with Prompt flow in Azure AI / ML Studio using a low-code approach.
Available on:
Evaluate Fine-tuned Phi-3 / Phi-3.5 Models in Azure AI Studio Focusing on Microsoft’s Responsible AI
Detailed instructions for evaluating the Phi-3 / Phi-3.5 model in Azure AI Studio using a low-code approach.
Available on:

How can you evaluate the safety and performance of a fine-tuned Phi-3 / Phi-3.5 model in Azure AI Studio?

Fine-tuning a model can sometimes lead to unintended or undesired responses. To ensure that the model remains safe and effective, it’s important to evaluate it. This evaluation helps to assess the model’s potential to generate harmful content and its ability to produce accurate, relevant, and coherent responses. In this tutorial, you will learn how to evaluate the safety and performance of a fine-tuned Phi-3 / Phi-3.5 model integrated with Prompt flow in Azure AI Studio.

Here is an Azure AI Studio’s evaluation process.

Note

The Code first approach tutorial includes tips on how to use the Phi-3.5 model below at Fine-tune the Phi-3 model section.
The Low code approach tutorial currently supports only the Phi-3 model. This tutorial will be updated to include Phi-3.5 model fine-tuning as soon as it is supported in Azure AI/ML Studio.
The evaluation process in Azure AI Studio is identical for both Phi-3 and Phi-3.5, so the title of this tutorial includes both models.

For more detailed information and to explore additional resources about Phi-3 and Phi-3.5, please visit the Phi-3CookBook.

Prerequisites

Series1: Introduction to Azure AI Studio’s Prompt flow evaluation

Introduction to safety evaluation
Introduction to performance evaluation

Series2: Evaluating the Phi-3 / Phi-3.5 model in Azure AI Studio

Before you begin
Deploy Azure OpenAI to evaluate the Phi-3 / Phi-3.5 model
Evaluate the fine-tuned Phi-3 / Phi-3.5 model using Azure AI Studio’s Prompt flow evaluation

To ensure that your AI model is ethical and safe, it’s crucial to evaluate it against Microsoft’s Responsible AI Principles. In Azure AI Studio, safety evaluations allow you to evaluate an your model’s vulnerability to jailbreak attacks and its potential to generate harmful content, which is directly aligned with these principles.

Microsoft’s Responsible AI Principles

Before beginning the technical steps, it’s essential to understand Microsoft’s Responsible AI Principles, an ethical framework designed to guide the responsible development, deployment, and operation of AI systems. These principles guide the responsible design, development, and deployment of AI systems, ensuring that AI technologies are built in a way that is fair, transparent, and inclusive. These principles are the foundation for evaluating the safety of AI models.

Microsoft’s Responsible AI Principles include:

Fairness and Inclusiveness: AI systems should treat everyone fairly and avoid affecting similarly situated groups of people in different ways. For example, when AI systems provide guidance on medical treatment, loan applications, or employment, they should make the same recommendations to everyone who has similar symptoms, financial circumstances, or professional qualifications.
Reliability and Safety: To build trust, it’s critical that AI systems operate reliably, safely, and consistently. These systems should be able to operate as they were originally designed, respond safely to unanticipated conditions, and resist harmful manipulation. How they behave and the variety of conditions they can handle reflect the range of situations and circumstances that developers anticipated during design and testing.
Transparency: When AI systems help inform decisions that have tremendous impacts on people’s lives, it’s critical that people understand how those decisions were made. For example, a bank might use an AI system to decide whether a person is creditworthy. A company might use an AI system to determine the most qualified candidates to hire.
Privacy and Security: As AI becomes more prevalent, protecting privacy and securing personal and business information are becoming more important and complex. With AI, privacy and data security require close attention because access to data is essential for AI systems to make accurate and informed predictions and decisions about people.
Accountability: The people who design and deploy AI systems must be accountable for how their systems operate. Organizations should draw upon industry standards to develop accountability norms. These norms can ensure that AI systems aren’t the final authority on any decision that affects people’s lives. They can also ensure that humans maintain meaningful control over otherwise highly autonomous AI systems.

Safety metrics

In this tutorial, you will evaluate the safety of the fine-tuned Phi-3 / Phi-3.5 model using Azure AI Studio’s safety metrics. These metrics help you assess the model’s potential to generate harmful content and its vulnerability to jailbreak attacks. The safety metrics include:

Self-harm-related Content: Evaluates whether the model has a tendency to produce self-harm related content.
Hateful and Unfair Content: Evaluates whether the model has a tendency to produce hateful or unfair content.
Violent Content: Evaluates whether the model has a tendency to produce violent content.
Sexual Content: Evaluates whether the model has a tendency to produce inappropriate sexual content.

Evaluating these aspects ensures that the AI model does not produce harmful or offensive content, aligning it with societal values and regulatory standards.

To ensure that your AI model is performing as expected, it’s important to evaluate its performance against performance metrics. In Azure AI Studio, performance evaluations allow you to evaluate your model’s effectiveness in generating accurate, relevant, and coherent responses.

Image Source: Evaluation of generative AI applications

Performance metrics

In this tutorial, you will evaluate the performance of the fine-tuned Phi-3 / Phi-3.5 model using Azure AI Studio’s performance metrics. These metrics help you assess the model’s effectiveness in generating accurate, relevant, and coherent responses. The performance metrics include:

Groundedness: Evaluate how well the generated answers align with the information from the input source.
Relevance: Evaluates the pertinence of generated responses to the given questions.
Coherence: Evaluate how smoothly the generated text flows, reads naturally, and resembles human-like language.
Fluency: Evaluate the language proficiency of the generated text.
GPT Similarity: Compares the generated response with the ground truth for similarity.
F1 Score: Calculates the ratio of shared words between the generated response and the source data.

These metrics help you evaluate the model’s effectiveness in generating accurate, relevant, and coherent responses.

This tutorial is a follow up to the previous blog posts, “Fine-Tune and Integrate Custom Phi-3 Models with Prompt Flow: Step-by-Step Guide” and “Fine-Tune and Integrate Custom Phi-3 Models with Prompt Flow in Azure AI Studio.” In these posts, we walked through the process of fine-tuning a Phi-3 / Phi-3.5 model in Azure AI Studio and integrating it with Prompt flow.

In this tutorial, you will deploy an Azure OpenAI model as an evaluator in Azure AI Studio and use it to evaluate your fine-tuned Phi-3 / Phi-3.5 model.

Before you begin this tutorial, make sure you have the following prerequisites, as described in the previous tutorials:

A prepared dataset to evaluate the fine-tuned Phi-3 / Phi-3.5 model.
A Phi-3 / Phi-3.5 model that has been fine-tuned and deployed to Azure Machine Learning.
A Prompt flow integrated with your fine-tuned Phi-3 / Phi-3.5 model in Azure AI Studio.

Note

You will use the test_data.jsonl file, located in the data folder from the ULTRACHAT_200k dataset downloaded in the previous blog posts, as the dataset to evaluate the fine-tuned Phi-3 / Phi-3.5 model.

Integrate the custom Phi-3 / Phi-3.5 model with Prompt flow in Azure AI Studio(Code first approach)

To proceed, you need to integrate your fine-tuned Phi-3 / Phi-3.5 model into Prompt flow in Azure AI Studio.

Create Azure AI Studio Hub

You need to create a Hub before creating the Project. A Hub acts like a Resource Group, allowing you to organize and manage multiple Projects within Azure AI Studio.

Sign in Azure AI Studio.
Select All hubs from the left side tab.
Select + New hub from the navigation menu.
Perform the following tasks:
- Enter Hub name. It must be a unique value.
- Select your Azure Subscription.
- Select the Resource group to use (create a new one if needed).
- Select the Location you’d like to use.
- Select the Connect Azure AI Services to use (create a new one if needed).
- Select Connect Azure AI Search to Skip connecting.
Select Next.

Create Azure AI Studio Project

In the Hub that you created, select All projects from the left side tab.
Select + New project from the navigation menu.
Enter Project name. It must be a unique value.
Select Create a project.

Add a custom connection for the fine-tuned Phi-3 / Phi-3.5 model

To integrate your custom Phi-3 / Phi-3.5 model with Prompt flow, you need to save the model’s endpoint and key in a custom connection. This setup ensures access to your custom Phi-3 / Phi-3.5 model in Prompt flow.

Set api key and endpoint uri of the fine-tuned Phi-3 / Phi-3.5 model

Visit Azure ML Studio.
Navigate to the Azure Machine learning workspace that you created.
Select Endpoints from the left side tab.
Select endpoint that you created.
Select Consume from the navigation menu.
Copy your REST endpoint and Primary key.

Add the Custom Connection

Visit Azure AI Studio.
Navigate to the Azure AI Studio project that you created.
In the Project that you created, select Settings from the left side tab.
Select + New connection.
Select Custom keys from the navigation menu.
Perform the following tasks:
- Select + Add key value pairs.
- For the key name, enter endpoint and paste the endpoint you copied from Azure ML Studio into the value field.
- Select + Add key value pairs again.
- For the key name, enter key and paste the key you copied from Azure ML Studio into the value field.
- After adding the keys, select is secret to prevent the key from being exposed.
Select Add connection.

Create Prompt flow

You have added a custom connection in Azure AI Studio. Now, let’s create a Prompt flow using the following steps. Then, you will connect this Prompt flow to the custom connection to use the fine-tuned model within the Prompt flow.

Navigate to the Azure AI Studio project that you created.
Select Prompt flow from the left side tab.
Select + Create from the navigation menu.
Select Chat flow from the navigation menu.
Enter Folder name to use.
Select Create.

Set up Prompt flow to chat with your custom Phi-3 / Phi-3.5 model

You need to integrate the fine-tuned Phi-3 / Phi-3.5 model into a Prompt flow. However, the existing Prompt flow provided is not designed for this purpose. Therefore, you must redesign the Prompt flow to enable the integration of the custom model.

In the Prompt flow, perform the following tasks to rebuild the existing flow:

Add the following code to integrate_with_promptflow.py to use the custom Phi-3 / Phi-3.5 model in Prompt flow.

import logging
import requests
from promptflow import tool
from promptflow.connections import CustomConnection


logging.basicConfig(
    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
    datefmt="%Y-%m-%d %H:%M:%S",
    level=logging.DEBUG
)
logger = logging.getLogger(__name__)

def query_phi3_model(input_data: str, connection: CustomConnection) -> str:
    """
    Send a request to the Phi-3 / Phi-3.5 model endpoint with the given input data using Custom Connection.
    """

    
    endpoint_url = connection.endpoint
    api_key = connection.key

    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {api_key}"
    }
data = {
    "input_data": [input_data],
    "params": {
        "temperature": 0.7,
        "max_new_tokens": 128,
        "do_sample": True,
        "return_full_text": True
        }
    }
    try:
        response = requests.post(endpoint_url, json=data, headers=headers)
        response.raise_for_status()
        
        
        logger.debug(f"Full JSON response: {response.json()}")

        result = response.json()["output"]
        logger.info("Successfully received response from Azure ML Endpoint.")
        return result
    except requests.exceptions.RequestException as e:
        logger.error(f"Error querying Azure ML Endpoint: {e}")
        raise

@tool
def my_python_tool(input_data: str, connection: CustomConnection) -> str:
    """
    Tool function to process input data and query the Phi-3 / Phi-3.5 model.
    """
    return query_phi3_model(input_data, connection)

Select Chat input, Chat output to enable chat with your model.
Now you are ready to chat with your custom Phi-3 / Phi-3.5 model. In the next exercise, you will learn how to start Prompt flow and use it to chat with your fine-tuned Phi-3 / Phi-3.5 model.

Note

The rebuilt flow should look like the image below:

Start Prompt flow

Select Start compute sessions to start Prompt flow.
Select Validate and parse input to renew parameters.
Select the Value of the connection to the custom connection you created. For example, connection.

Chat with your custom Phi-3 / Phi-3.5 model

Select Chat.
Here’s an example of the results: Now you can chat with your custom Phi-3 / Phi-3.5 model. It is recommended to ask questions based on the data used for fine-tuning.

To evaluate the Phi-3 / Phi-3.5 model in Azure AI Studio, you need to deploy an Azure OpenAI model. This model will be used to evaluate the performance of the Phi-3 / Phi-3.5 model.

Deploy Azure OpenAI

Sign in to Azure AI Studio.
Navigate to the Azure AI Studio project that you created.
In the Project that you created, select Deployments from the left side tab.
Select + Deploy model from the navigation menu.
Select Deploy base model.
Select Azure OpenAI model you’d like to use. For example, gpt-4o.
Select Confirm.

Start a new evaluation

Visit Azure AI Studio.
Navigate to the Azure AI Studio project that you created.
In the Project that you created, select Evaluation from the left side tab.
Select + New evaluation from the navigation menu.
Select Prompt flow evaluation.
perform the following tasks:
- Enter the evaluation name. It must be a unique value.
- Select Question and answer without context as the task type. Because, the UlTRACHAT_200k dataset used in this tutorial does not contain context.
- Select the prompt flow you’d like to evaluate.
Select Next.
perform the following tasks:
- Select Add your dataset to upload the dataset. For example, you can upload the test dataset file, such as test_data.json1, which is included when you download the ULTRACHAT_200k dataset.
- Select the appropriate Dataset column that matches your dataset. For example, if you are using the ULTRACHAT_200k dataset, select ${data.prompt} as the dataset column.
Select Next.
perform the following tasks to configure the performance and quality metrics:
- Select the performance and quality metrics you’d like to use.
- Select the Azure OpenAI model that you created for evaluation. For example, select gpt-4o.
perform the following tasks to configure the risk and safety metrics:
- Select the risk and safety metrics you’d like to use.
- Select the threshold to calculate the defect rate you’d like to use. For example, select Medium.
- For question, select Data source to {$data.prompt}.
- For answer, select Data source to {$run.outputs.answer}.
- For ground_truth, select Data source to {$data.message}.
Select Next.
Select Submit to start the evaluation.
The evaluation will take some time to complete. You can monitor the progress in the Evaluation tab.

Review the Evaluation Results

Note

The results presented below are intended to illustrate the evaluation process. In this tutorial, we have used a model fine-tuned on a relatively small dataset, which may lead to sub-optimal results. Actual results may vary significantly depending on the size, quality, and diversity of the dataset used, as well as the specific configuration of the model.

Once the evaluation is complete, you can review the results for both performance and safety metrics.

Performance and quality metrics:
- evaluate the model’s effectiveness in generating coherent, fluent, and relevant responses.
Risk and safety metrics:
- Ensure that the model’s outputs are safe and align with Responsible AI Principles, avoiding any harmful or offensive content.
You can scroll down to view Detailed metrics result.
By evaluating your custom Phi-3 / Phi-3.5 model against both performance and safety metrics, you can confirm that the model is not only effective, but also adheres to responsible AI practices, making it ready for real-world deployment.

You’ve completed this tutorial

You have successfully evaluated the fine-tuned Phi-3 model integrated with Prompt flow in Azure AI Studio. This is an important step in ensuring that your AI models not only perform well, but also adhere to Microsoft’s Responsible AI principles to help you build trustworthy and reliable AI applications.

Clean Up Azure Resources

Cleanup your Azure resources to avoid additional charges to your account. Go to the Azure portal and delete the following resources:

How can you evaluate the safety and performance of a fine-tuned Phi-3 / Phi-3.5 model in Azure AI Studio?

Prerequisites