Fine-Tuning Large Language Models using Azure Databricks and Azure OpenAI

With Azure Databricks, users can now leverage the power of LLMs for specialized tasks by fine-tuning them with their own data, enhancing domain-specific performance. To fine-tune a language model using Azure Databricks, you can utilize the Mosaic AI Model Training interface which simplifies the process of full model fine-tuning. This feature allows you to fine-tune a model with your custom data, with checkpoints saved to MLflow, ensuring you retain complete control over the fine-tuned model.

This lab will take approximately 60 minutes to complete.

Before you start

You’ll need an Azure subscription in which you have administrative-level access.

Provision an Azure OpenAI resource

If you don’t already have one, provision an Azure OpenAI resource in your Azure subscription.

Sign into the Azure portal at https://portal.azure.com.
Create an Azure OpenAI resource with the following settings:
- Subscription: Select an Azure subscription that has been approved for access to the Azure OpenAI service
- Resource group: Choose or create a resource group
- Region: Make a random choice from any of the following regions*
  - East US 2
  - North Central US
  - Sweden Central
  - Switzerland West
- Name: A unique name of your choice
- Pricing tier: Standard S0

* Azure OpenAI resources are constrained by regional quotas. The listed regions include default quota for the model type(s) used in this exercise. Randomly choosing a region reduces the risk of a single region reaching its quota limit in scenarios where you are sharing a subscription with other users. In the event of a quota limit being reached later in the exercise, there’s a possibility you may need to create another resource in a different region.

Wait for deployment to complete. Then go to the deployed Azure OpenAI resource in the Azure portal.
In the left pane, under Resource Management, select Keys and Endpoint.
Copy the endpoint and one of the available keys as you will use it later in this exercise.
Launch Cloud Shell and run az account get-access-token to get a temporary authorization token for API testing. Keep it together with the endpoint and key copied previously.

Deploy the required model

Azure provides a web-based portal named Azure AI Studio, that you can use to deploy, manage, and explore models. You’ll start your exploration of Azure OpenAI by using Azure AI Studio to deploy a model.

Note: As you use Azure AI Studio, message boxes suggesting tasks for you to perform may be displayed. You can close these and follow the steps in this exercise.

In the Azure portal, on the Overview page for your Azure OpenAI resource, scroll down to the Get Started section and select the button to go to Azure AI Studio.
In Azure AI Studio, in the pane on the left, select the Deployments page and view your existing model deployments. If you don’t already have one, create a new deployment of the gpt-35-turbo model with the following settings:
- Deployment name: gpt-35-turbo-0613
- Model: gpt-35-turbo
- Model version: 0613
- Deployment type: Standard
- Tokens per minute rate limit: 5K*
- Content filter: Default
- Enable dynamic quota: Disabled

* A rate limit of 5,000 tokens per minute is more than adequate to complete this exercise while leaving capacity for other people using the same subscription.

Provision an Azure Databricks workspace

Tip: If you already have an Azure Databricks workspace, you can skip this procedure and use your existing workspace.

Sign into the Azure portal at https://portal.azure.com.
Create an Azure Databricks resource with the following settings:
- Subscription: Select the same Azure subscription that you used to create your Azure OpenAI resource
- Resource group: The same resource group where you created your Azure OpenAI resource
- Region: The same region where you created your Azure OpenAI resource
- Name: A unique name of your choice
- Pricing tier: Premium or Trial
Select Review + create and wait for deployment to complete. Then go to the resource and launch the workspace.

Create a cluster

Azure Databricks is a distributed processing platform that uses Apache Spark clusters to process data in parallel on multiple nodes. Each cluster consists of a driver node to coordinate the work, and worker nodes to perform processing tasks. In this exercise, you’ll create a single-node cluster to minimize the compute resources used in the lab environment (in which resources may be constrained). In a production environment, you’d typically create a cluster with multiple worker nodes.

Tip: If you already have a cluster with a 13.3 LTS ML or higher runtime version in your Azure Databricks workspace, you can use it to complete this exercise and skip this procedure.

In the Azure portal, browse to the resource group where the Azure Databricks workspace was created.
Select your Azure Databricks Service resource.
In the Overview page for your workspace, use the Launch Workspace button to open your Azure Databricks workspace in a new browser tab; signing in if prompted.

Tip: As you use the Databricks Workspace portal, various tips and notifications may be displayed. Dismiss these and follow the instructions provided to complete the tasks in this exercise.

In the sidebar on the left, select the (+) New task, and then select Cluster.
In the New Cluster page, create a new cluster with the following settings:
- Cluster name: User Name’s cluster (the default cluster name)
- Policy: Unrestricted
- Cluster mode: Single Node
- Access mode: Single user (with your user account selected)
- Databricks runtime version: Select the ML edition of the latest non-beta version of the runtime (Not a Standard runtime version) that:
  - Does not use a GPU
  - Includes Scala > 2.11
  - Includes Spark > 3.4
- Use Photon Acceleration: Unselected
- Node type: Standard_DS3_v2
- Terminate after 20 minutes of inactivity
Wait for the cluster to be created. It may take a minute or two.

Note: If your cluster fails to start, your subscription may have insufficient quota in the region where your Azure Databricks workspace is provisioned. See CPU core limit prevents cluster creation for details. If this happens, you can try deleting your workspace and creating a new one in a different region.

Install required libraries

In your cluster’s page, select the Libraries tab.
Select Install New.
Select PyPI as the library source and install the following Python packages:
- numpy==2.1.0
- requests==2.32.3
- openai==1.42.0
- tiktoken==0.7.0

Create a new notebook and ingest data

In the sidebar, use the (+) New link to create a Notebook.
Name your notebook and in the Connect drop-down list, select your cluster if it is not already selected. If the cluster is not running, it may take a minute or so to start.

In the first cell of the notebook, enter the following code, which uses shell commands to download data files from GitHub into the file system used by your cluster.

 %sh
 rm -r /dbfs/fine_tuning
 mkdir /dbfs/fine_tuning
 wget -O /dbfs/fine_tuning/training_set.jsonl https://github.com/MicrosoftLearning/mslearn-databricks/raw/main/data/training_set.jsonl
 wget -O /dbfs/fine_tuning/validation_set.jsonl https://github.com/MicrosoftLearning/mslearn-databricks/raw/main/data/validation_set.jsonl

In a new cell, run the following code with the access information you copied at the beginning of this exercise to assign persistent environment variables for authentication when using Azure OpenAI resources:
```
 import os

 os.environ["AZURE_OPENAI_API_KEY"] = "your_openai_api_key"
 os.environ["AZURE_OPENAI_ENDPOINT"] = "your_openai_endpoint"
 os.environ["TEMP_AUTH_TOKEN"] = "your_access_token"
```

Validade token counts

Both training_set.jsonl and validation_set.jsonl are made of different conversation examples between user and assistant that will serve as data points for training and validating the fine-tuned model. Individual examples need to remain under the gpt-35-turbo model’s input token limit of 4096 tokens.

In a new cell, run the following code to validate the token counts for each file:

 import json
 import tiktoken
 import numpy as np
 from collections import defaultdict

 encoding = tiktoken.get_encoding("cl100k_base")

 def num_tokens_from_messages(messages, tokens_per_message=3, tokens_per_name=1):
     num_tokens = 0
     for message in messages:
         num_tokens += tokens_per_message
         for key, value in message.items():
             num_tokens += len(encoding.encode(value))
             if key == "name":
                 num_tokens += tokens_per_name
     num_tokens += 3
     return num_tokens

 def num_assistant_tokens_from_messages(messages):
     num_tokens = 0
     for message in messages:
         if message["role"] == "assistant":
             num_tokens += len(encoding.encode(message["content"]))
     return num_tokens

 def print_distribution(values, name):
     print(f"\n##### Distribution of {name}:")
     print(f"min / max: {min(values)}, {max(values)}")
     print(f"mean / median: {np.mean(values)}, {np.median(values)}")

 files = ['/dbfs/fine_tuning/training_set.jsonl', '/dbfs/fine_tuning/validation_set.jsonl']

 for file in files:
     print(f"File: {file}")
     with open(file, 'r', encoding='utf-8') as f:
         dataset = [json.loads(line) for line in f]

     total_tokens = []
     assistant_tokens = []

     for ex in dataset:
         messages = ex.get("messages", {})
         total_tokens.append(num_tokens_from_messages(messages))
         assistant_tokens.append(num_assistant_tokens_from_messages(messages))

     print_distribution(total_tokens, "total tokens")
     print_distribution(assistant_tokens, "assistant tokens")
     print('*' * 75)

Upload fine-tuning files to Azure OpenAI

Before you start to fine-tune the model, you need to initialize an OpenAI client and add the fine-tuning files to its environment, generating file IDs that will be used to initialize the job.

In a new cell, run the following code:

 import os
 from openai import AzureOpenAI

 client = AzureOpenAI(
   azure_endpoint = os.getenv("AZURE_OPENAI_ENDPOINT"),
   api_key = os.getenv("AZURE_OPENAI_API_KEY"),
   api_version = "2024-05-01-preview"  # This API version or later is required to access seed/events/checkpoint features
 )

 training_file_name = '/dbfs/fine_tuning/training_set.jsonl'
 validation_file_name = '/dbfs/fine_tuning/validation_set.jsonl'

 training_response = client.files.create(
     file = open(training_file_name, "rb"), purpose="fine-tune"
 )
 training_file_id = training_response.id

 validation_response = client.files.create(
     file = open(validation_file_name, "rb"), purpose="fine-tune"
 )
 validation_file_id = validation_response.id

 print("Training file ID:", training_file_id)
 print("Validation file ID:", validation_file_id)

Submit fine-tuning job

Now that the fine-tuning files have been successfully uploaded you can submit your fine-tuning training job. It isn’t unusual for training to take more than an hour to complete. Once training is completed, you can see the results in Azure AI Studio by selecting the Fine-tuning option in the left pane.

In a new cell, run the following code to start the fine-tuning training job:

 response = client.fine_tuning.jobs.create(
     training_file = training_file_id,
     validation_file = validation_file_id,
     model = "gpt-35-turbo-0613",
     seed = 105 # seed parameter controls reproducibility of the fine-tuning job. If no seed is specified one will be generated automatically.
 )

 job_id = response.id

The seed parameter controls reproducibility of the fine-tuning job. Passing in the same seed and job parameters should produce the same results, but can differ in rare cases. If no seed is specified one will be generated automatically.

In a new cell, you can run the following code to monitor the status of the fine-tuning job:
```
 print("Job ID:", response.id)
 print("Status:", response.status)
```

Once the job status changes to succeeded, run the following code to get the final results:

 response = client.fine_tuning.jobs.retrieve(job_id)

 print(response.model_dump_json(indent=2))
 fine_tuned_model = response.fine_tuned_model

Deploy fine-tuned model

Now that you have a fine-tuned model, you can deploy it as a customized model and use it like any other deployed model in either the Chat Playground of Azure AI Studio, or via the chat completion API.

In a new cell, run the following code to deploy your fine-tuned model:

 import json
 import requests

 token = os.getenv("TEMP_AUTH_TOKEN")
 subscription = "<YOUR_SUBSCRIPTION_ID>"
 resource_group = "<YOUR_RESOURCE_GROUP_NAME>"
 resource_name = "<YOUR_AZURE_OPENAI_RESOURCE_NAME>"
 model_deployment_name = "gpt-35-turbo-ft"

 deploy_params = {'api-version': "2023-05-01"}
 deploy_headers = {'Authorization': 'Bearer {}'.format(token), 'Content-Type': 'application/json'}

 deploy_data = {
     "sku": {"name": "standard", "capacity": 1},
     "properties": {
         "model": {
             "format": "OpenAI",
             "name": "<YOUR_FINE_TUNED_MODEL>",
             "version": "1"
         }
     }
 }
 deploy_data = json.dumps(deploy_data)

 request_url = f'https://management.azure.com/subscriptions/{subscription}/resourceGroups/{resource_group}/providers/Microsoft.CognitiveServices/accounts/{resource_name}/deployments/{model_deployment_name}'

 print('Creating a new deployment...')

 r = requests.put(request_url, params=deploy_params, headers=deploy_headers, data=deploy_data)

 print(r)
 print(r.reason)
 print(r.json())

In a new cell, run the following code to use your customized model in a chat completion call:

 import os
 from openai import AzureOpenAI

 client = AzureOpenAI(
   azure_endpoint = os.getenv("AZURE_OPENAI_ENDPOINT"),
   api_key = os.getenv("AZURE_OPENAI_API_KEY"),
   api_version = "2024-02-01"
 )

 response = client.chat.completions.create(
     model = "gpt-35-turbo-ft", # model = "Custom deployment name you chose for your fine-tuning model"
     messages = [
         {"role": "system", "content": "You are a helpful assistant."},
         {"role": "user", "content": "Does Azure OpenAI support customer managed keys?"},
         {"role": "assistant", "content": "Yes, customer managed keys are supported by Azure OpenAI."},
         {"role": "user", "content": "Do other Azure AI services support this too?"}
     ]
 )

 print(response.choices[0].message.content)

Clean up

When you’re done with your Azure OpenAI resource, remember to delete the deployment or the entire resource in the Azure portal at https://portal.azure.com.

In Azure Databricks portal, on the Compute page, select your cluster and select ■ Terminate to shut it down.

If you’ve finished exploring Azure Databricks, you can delete the resources you’ve created to avoid unnecessary Azure costs and free up capacity in your subscription.