Customize a model with fine-tuning
Azure OpenAI Service lets you tailor our models to your personal datasets by using a process known as fine-tuning. This customization step lets you get more out of the service by providing:
- Higher quality results than what you can get just from prompt engineering
- The ability to train on more examples than can fit into a model's max request context limit.
- Token savings due to shorter prompts
- Lower-latency requests, particularly when using smaller models.
In contrast to few-shot learning, fine tuning improves the model by training on many more examples than can fit in a prompt, letting you achieve better results on a wide number of tasks. Because fine tuning adjusts the base model’s weights to improve performance on the specific task, you won’t have to include as many examples or instructions in your prompt. This means less text sent and fewer tokens processed on every API call, potentially saving cost, and improving request latency.
We use LoRA, or low rank approximation, to fine-tune models in a way that reduces their complexity without significantly affecting their performance. This method works by approximating the original high-rank matrix with a lower rank one, thus only fine-tuning a smaller subset of "important" parameters during the supervised training phase, making the model more manageable and efficient. For users, this makes training faster and more affordable than other techniques.
Note
Azure OpenAI currently only supports text-to-text fine-tuning for all supported models including GPT-4o mini.
Prerequisites
- Read the When to use Azure OpenAI fine-tuning guide.
- An Azure subscription. Create one for free.
- An Azure OpenAI resource that's located in a region that supports fine-tuning of the Azure OpenAI model. Check the Model summary table and region availability for the list of available models by region and supported functionality. For more information, see Create a resource and deploy a model with Azure OpenAI.
- Fine-tuning access requires Cognitive Services OpenAI Contributor.
- If you do not already have access to view quota, and deploy models in Azure OpenAI Studio you will require additional permissions.
Models
The following models support fine-tuning:
babbage-002
davinci-002
gpt-35-turbo
(0613)gpt-35-turbo
(1106)gpt-35-turbo
(0125)gpt-4
(0613)*gpt-4o-mini
(2024-07-18)*
* Fine-tuning for this model is currently in public preview.
Consult the models page to check which regions currently support fine-tuning.
If you plan to use gpt-4
for fine-tuning, please refer to the GPT-4 public preview safety evaluation guidance
Review the workflow for Azure OpenAI Studio
Take a moment to review the fine-tuning workflow for using Azure OpenAI Studio:
- Prepare your training and validation data.
- Use the Create custom model wizard in Azure OpenAI Studio to train your custom model.
- Select a base model.
- Choose your training data.
- Optionally, choose your validation data.
- Optionally, configure advanced options for your fine-tuning job.
- Review your choices and train your new custom model.
- Check the status of your custom fine-tuned model.
- Deploy your custom model for use.
- Use your custom model.
- Optionally, analyze your custom model for performance and fit.
Prepare your training and validation data
Your training data and validation data sets consist of input and output examples for how you would like the model to perform.
Different model types require a different format of training data.
The training and validation data you use must be formatted as a JSON Lines (JSONL) document. For gpt-35-turbo-0613
the fine-tuning dataset must be formatted in the conversational format that is used by the Chat completions API.
If you would like a step-by-step walk-through of fine-tuning a gpt-35-turbo-0613
model please refer to the Azure OpenAI fine-tuning tutorial.
Example file format
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "Is Xbox better than PlayStation?"}, {"role": "assistant", "content": "I apologize, but I cannot provide personal opinions. My primary job is to assist you with any issues related to your Xbox device. Do you have any Xbox-related issues that need addressing?"}]}
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "Can I play PC games on my Xbox?"}, {"role": "assistant", "content": "While Xbox consoles don't natively support PC games, they can play an array of Xbox games, many of which are also available on PC. Can I assist you with finding games compatible with your Xbox device?"}]}
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "I'm having trouble connecting my Xbox to the Wi-Fi."}, {"role": "assistant", "content": "No worries, let's go through the network settings on your Xbox. Can you please tell me what happens when you try to connect it to the Wi-Fi?"}]}
Multi-turn chat file format
Multiple turns of a conversation in a single line of your jsonl training file is also supported. To skip fine-tuning on specific assistant messages add the optional weight
key value pair. Currently weight
can be set to 0 or 1.
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "What's the capital of France?"}, {"role": "assistant", "content": "Paris", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Paris, as if everyone doesn't know that already.", "weight": 1}]}
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "Who wrote 'Romeo and Juliet'?"}, {"role": "assistant", "content": "William Shakespeare", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Oh, just some guy named William Shakespeare. Ever heard of him?", "weight": 1}]}
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "How far is the Moon from Earth?"}, {"role": "assistant", "content": "384,400 kilometers", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Around 384,400 kilometers. Give or take a few, like that really matters.", "weight": 1}]}
In addition to the JSONL format, training and validation data files must be encoded in UTF-8 and include a byte-order mark (BOM). The file must be less than 512 MB in size.
Create your training and validation datasets
The more training examples you have, the better. Fine tuning jobs will not proceed without at least 10 training examples, but such a small number are not enough to noticeably influence model responses. It is best practice to provide hundreds, if not thousands, of training examples to be successful.
In general, doubling the dataset size can lead to a linear increase in model quality. But keep in mind, low quality examples can negatively impact performance. If you train the model on a large amount of internal data, without first pruning the dataset for only the highest quality examples you could end up with a model that performs much worse than expected.
Use the Create custom model wizard
Azure OpenAI Studio provides the Create custom model wizard, so you can interactively create and train a fine-tuned model for your Azure resource.
Open Azure OpenAI Studio at https://oai.azure.com/ and sign in with credentials that have access to your Azure OpenAI resource. During the sign-in workflow, select the appropriate directory, Azure subscription, and Azure OpenAI resource.
In Azure OpenAI Studio, browse to the Management > Models pane, and select Create a custom model.
The Create custom model wizard opens.
Select the base model
The first step in creating a custom model is to choose a base model. The Base model pane lets you choose a base model to use for your custom model. Your choice influences both the performance and the cost of your model.
Select the base model from the Base model type dropdown, and then select Next to continue.
You can create a custom model from one of the following available base models:
babbage-002
davinci-002
gpt-35-turbo
(0613)gpt-35-turbo
(1106)gpt-35-turbo
(0125)gpt-4
(0613)Or you can fine tune a previously fine-tuned model, formatted as base-model.ft-{jobid}.
For more information about our base models that can be fine-tuned, see Models.
Choose your training data
The next step is to either choose existing prepared training data or upload new prepared training data to use when customizing your model. The Training data pane displays any existing, previously uploaded datasets and also provides options to upload new training data.
If your training data is already uploaded to the service, select Choose dataset.
- Select the file from the list shown in the Training data pane.
To upload new training data, use one of the following options:
Select Local file to upload training data from a local file.
Select Azure blob or other shared web locations to import training data from Azure Blob or another shared web location.
For large data files, we recommend that you import from an Azure Blob store. Large files can become unstable when uploaded through multipart forms because the requests are atomic and can't be retried or resumed. For more information about Azure Blob Storage, see What is Azure Blob Storage?
Note
Training data files must be formatted as JSONL files, encoded in UTF-8 with a byte-order mark (BOM). The file must be less than 512 MB in size.
Upload training data from local file
You can upload a new training dataset to the service from a local file by using one of the following methods:
Drag and drop the file into the client area of the Training data pane, and then select Upload file.
Select Browse for a file from the client area of the Training data pane, choose the file to upload from the Open dialog, and then select Upload file.
After you select and upload the training dataset, select Next to continue.
Import training data from Azure Blob store
You can import a training dataset from Azure Blob or another shared web location by providing the name and location of the file.
Enter the File name for the file.
For the File location, provide the Azure Blob URL, the Azure Storage shared access signature (SAS), or other link to an accessible shared web location.
Select Upload file to import the training dataset to the service.
After you select and upload the training dataset, select Next to continue.
Choose your validation data
The next step provides options to configure the model to use validation data in the training process. If you don't want to use validation data, you can choose Next to continue to the advanced options for the model. Otherwise, if you have a validation dataset, you can either choose existing prepared validation data or upload new prepared validation data to use when customizing your model.
The Validation data pane displays any existing, previously uploaded training and validation datasets and provides options by which you can upload new validation data.
If your validation data is already uploaded to the service, select Choose dataset.
- Select the file from the list shown in the Validation data pane.
To upload new validation data, use one of the following options:
Select Local file to upload validation data from a local file.
Select Azure blob or other shared web locations to import validation data from Azure Blob or another shared web location.
For large data files, we recommend that you import from an Azure Blob store. Large files can become unstable when uploaded through multipart forms because the requests are atomic and can't be retried or resumed.
Note
Similar to training data files, validation data files must be formatted as JSONL files, encoded in UTF-8 with a byte-order mark (BOM). The file must be less than 512 MB in size.
Upload validation data from local file
You can upload a new validation dataset to the service from a local file by using one of the following methods:
Drag and drop the file into the client area of the Validation data pane, and then select Upload file.
Select Browse for a file from the client area of the Validation data pane, choose the file to upload from the Open dialog, and then select Upload file.
After you select and upload the validation dataset, select Next to continue.
Import validation data from Azure Blob store
You can import a validation dataset from Azure Blob or another shared web location by providing the name and location of the file.
Enter the File name for the file.
For the File location, provide the Azure Blob URL, the Azure Storage shared access signature (SAS), or other link to an accessible shared web location.
Select Upload file to import the training dataset to the service.
After you select and upload the validation dataset, select Next to continue.
Configure advanced options
The Create custom model wizard shows the parameters for training your fine-tuned model on the Advanced options pane. The following parameters are available:
Name | Type | Description |
---|---|---|
batch_size |
integer | The batch size to use for training. The batch size is the number of training examples used to train a single forward and backward pass. In general, we've found that larger batch sizes tend to work better for larger datasets. The default value as well as the maximum value for this property are specific to a base model. A larger batch size means that model parameters are updated less frequently, but with lower variance. |
learning_rate_multiplier |
number | The learning rate multiplier to use for training. The fine-tuning learning rate is the original learning rate used for pre-training multiplied by this value. Larger learning rates tend to perform better with larger batch sizes. We recommend experimenting with values in the range 0.02 to 0.2 to see what produces the best results. A smaller learning rate may be useful to avoid overfitting. |
n_epochs |
integer | The number of epochs to train the model for. An epoch refers to one full cycle through the training dataset. |
seed |
integer | The seed controls the reproducibility of the job. Passing in the same seed and job parameters should produce the same results, but may differ in rare cases. If a seed isn't specified, one will be generated for you |
Select Default to use the default values for the fine-tuning job, or select Advanced to display and edit the hyperparameter values. When defaults are selected, we determine the correct value algorithmically based on your training data.
After you configure the advanced options, select Next to review your choices and train your fine-tuned model.
Review your choices and train your model
The Review pane of the wizard displays information about your configuration choices.
If you're ready to train your model, select Start Training job to start the fine-tuning job and return to the Models pane.
Check the status of your custom model
The Models pane displays information about your custom model in the Customized models tab. The tab includes information about the status and job ID of the fine-tune job for your custom model. When the job completes, the tab displays the file ID of the result file. You might need to select Refresh in order to see an updated status for the model training job.
After you start a fine-tuning job, it can take some time to complete. Your job might be queued behind other jobs on the system. Training your model can take minutes or hours depending on the model and dataset size.
Here are some of the tasks you can do on the Models pane:
Check the status of the fine-tuning job for your custom model in the Status column of the Customized models tab.
In the Model name column, select the model name to view more information about the custom model. You can see the status of the fine-tuning job, training results, training events, and hyperparameters used in the job.
Select Download training file to download the training data you used for the model.
Select Download results to download the result file attached to the fine-tuning job for your model and analyze your custom model for training and validation performance.
Select Refresh to update the information on the page.
Checkpoints
When each training epoch completes a checkpoint is generated. A checkpoint is a fully functional version of a model which can both be deployed and used as the target model for subsequent fine-tuning jobs. Checkpoints can be particularly useful, as they can provide a snapshot of your model prior to overfitting having occurred. When a fine-tuning job completes you will have the three most recent versions of the model available to deploy.
Safety evaluation GPT-4 fine-tuning - public preview
GPT-4 and GPT-4o-mini are our most advanced models that can be fine-tuned to your needs. As with Azure OpenAI models generally, the advanced capabilities of fine-tuned models come with increased responsible AI challenges related to harmful content, manipulation, human-like behavior, privacy issues, and more. Learn more about risks, capabilities, and limitations in the Overview of Responsible AI practices and Transparency Note. To help mitigate the risks associated with GPT-4 and GPT-4o-mini fine-tuned models, we have implemented additional evaluation steps to help detect and prevent harmful content in the training and outputs of fine-tuned models. These steps are grounded in the Microsoft Responsible AI Standard and Azure OpenAI Service content filtering.
- Evaluations are conducted in dedicated, customer specific, private workspaces;
- Evaluation endpoints are in the same geography as the Azure OpenAI resource;
- Training data is not stored in connection with performing evaluations; only the final model assessment (deployable or not deployable) is persisted; and
GPT-4 and GPT-4o-mini fine-tuned model evaluation filters are set to predefined thresholds and cannot be modified by customers; they aren't tied to any custom content filtering configuration you may have created.
Data evaluation
Before training starts, your data is evaluated for potentially harmful content (violence, sexual, hate, and fairness, self-harm – see category definitions here). If harmful content is detected above the specified severity level, your training job will fail, and you'll receive a message informing you of the categories of failure.
Sample message:
The provided training data failed RAI checks for harm types: [hate_fairness, self_harm, violence]. Please fix the data and try again.
Your training data is evaluated automatically within your data import job as part of providing the fine-tuning capability.
If the fine-tuning job fails due to the detection of harmful content in training data, you won't be charged.
Model evaluation
After training completes but before the fine-tuned model is available for deployment, the resulting model is evaluated for potentially harmful responses using Azure’s built-in risk and safety metrics. Using the same approach to testing that we use for the base large language models, our evaluation capability simulates a conversation with your fine-tuned model to assess the potential to output harmful content, again using specified harmful content categories (violence, sexual, hate, and fairness, self-harm).
If a model is found to generate output containing content detected as harmful at above an acceptable rate, you'll be informed that your model isn't available for deployment, with information about the specific categories of harm detected:
Sample Message:
This model is unable to be deployed. Model evaluation identified that this fine tuned model scores above acceptable thresholds for [Violence, Self Harm]. Please review your training data set and resubmit the job.
As with data evaluation, the model is evaluated automatically within your fine-tuning job as part of providing the fine-tuning capability. Only the resulting assessment (deployable or not deployable) is logged by the service. If deployment of the fine-tuned model fails due to the detection of harmful content in model outputs, you won't be charged for the training run.
Deploy a fine-tuned model
When the fine-tuning job succeeds, you can deploy the custom model from the Models pane. You must deploy your custom model to make it available for use with completion calls.
Important
After you deploy a customized model, if at any time the deployment remains inactive for greater than fifteen (15) days, the deployment is deleted. The deployment of a customized model is inactive if the model was deployed more than fifteen (15) days ago and no completions or chat completions calls were made to it during a continuous 15-day period.
The deletion of an inactive deployment doesn't delete or affect the underlying customized model, and the customized model can be redeployed at any time. As described in Azure OpenAI Service pricing, each customized (fine-tuned) model that's deployed incurs an hourly hosting cost regardless of whether completions or chat completions calls are being made to the model. To learn more about planning and managing costs with Azure OpenAI, refer to the guidance in Plan to manage costs for Azure OpenAI Service.
Note
Only one deployment is permitted for a custom model. An error message is displayed if you select an already-deployed custom model.
To deploy your custom model, select the custom model to deploy, and then select Deploy model.
The Deploy model dialog box opens. In the dialog box, enter your Deployment name and then select Create to start the deployment of your custom model.
You can monitor the progress of your deployment on the Deployments pane in Azure OpenAI Studio.
Cross region deployment
Fine-tuning supports deploying a fine-tuned model to a different region than where the model was originally fine-tuned. You can also deploy to a different subscription/region.
The only limitations are that the new region must also support fine-tuning and when deploying cross subscription the account generating the authorization token for the deployment must have access to both the source and destination subscriptions.
Cross subscription/region deployment can be accomplished via Python or REST.
Use a deployed custom model
After your custom model deploys, you can use it like any other deployed model. You can use the Playgrounds in Azure OpenAI Studio to experiment with your new deployment. You can continue to use the same parameters with your custom model, such as temperature
and max_tokens
, as you can with other deployed models. For fine-tuned babbage-002
and davinci-002
models you will use the Completions playground and the Completions API. For fine-tuned gpt-35-turbo-0613
models you will use the Chat playground and the Chat completion API.
Analyze your custom model
Azure OpenAI attaches a result file named results.csv to each fine-tuning job after it completes. You can use the result file to analyze the training and validation performance of your custom model. The file ID for the result file is listed for each custom model in the Result file Id column on the Models pane for Azure OpenAI Studio. You can use the file ID to identify and download the result file from the Data files pane of Azure OpenAI Studio.
The result file is a CSV file that contains a header row and a row for each training step performed by the fine-tuning job. The result file contains the following columns:
Column name | Description |
---|---|
step |
The number of the training step. A training step represents a single pass, forward and backward, on a batch of training data. |
train_loss |
The loss for the training batch. |
train_mean_token_accuracy |
The percentage of tokens in the training batch correctly predicted by the model. For example, if the batch size is set to 3 and your data contains completions [[1, 2], [0, 5], [4, 2]] , this value is set to 0.83 (5 of 6) if the model predicted [[1, 1], [0, 5], [4, 2]] . |
valid_loss |
The loss for the validation batch. |
validation_mean_token_accuracy |
The percentage of tokens in the validation batch correctly predicted by the model. For example, if the batch size is set to 3 and your data contains completions [[1, 2], [0, 5], [4, 2]] , this value is set to 0.83 (5 of 6) if the model predicted [[1, 1], [0, 5], [4, 2]] . |
full_valid_loss |
The validation loss calculated at the end of each epoch. When training goes well, loss should decrease. |
full_valid_mean_token_accuracy |
The valid mean token accuracy calculated at the end of each epoch. When training is going well, token accuracy should increase. |
You can also view the data in your results.csv file as plots in Azure OpenAI Studio. Select the link for your trained model, and you will see three charts: loss, mean token accuracy, and token accuracy. If you provided validation data, both datasets will appear on the same plot.
Look for your loss to decrease over time, and your accuracy to increase. If you see a divergence between your training and validation data, that may indicate that you are overfitting. Try training with fewer epochs, or a smaller learning rate multiplier.
Clean up your deployments, custom models, and training files
When you're done with your custom model, you can delete the deployment and model. You can also delete the training and validation files you uploaded to the service, if needed.
Delete your model deployment
Important
After you deploy a customized model, if at any time the deployment remains inactive for greater than fifteen (15) days, the deployment is deleted. The deployment of a customized model is inactive if the model was deployed more than fifteen (15) days ago and no completions or chat completions calls were made to it during a continuous 15-day period.
The deletion of an inactive deployment doesn't delete or affect the underlying customized model, and the customized model can be redeployed at any time. As described in Azure OpenAI Service pricing, each customized (fine-tuned) model that's deployed incurs an hourly hosting cost regardless of whether completions or chat completions calls are being made to the model. To learn more about planning and managing costs with Azure OpenAI, refer to the guidance in Plan to manage costs for Azure OpenAI Service.
You can delete the deployment for your custom model on the Deployments pane in Azure OpenAI Studio. Select the deployment to delete, and then select Delete to delete the deployment.
Delete your custom model
You can delete a custom model on the Models pane in Azure OpenAI Studio. Select the custom model to delete from the Customized models tab, and then select Delete to delete the custom model.
Note
You can't delete a custom model if it has an existing deployment. You must first delete your model deployment before you can delete your custom model.
Delete your training files
You can optionally delete training and validation files that you uploaded for training, and result files generated during training, on the Management > Data files pane in Azure OpenAI Studio. Select the file to delete, and then select Delete to delete the file.
Continuous fine-tuning
Once you have created a fine-tuned model you may wish to continue to refine the model over time through further fine-tuning. Continuous fine-tuning is the iterative process of selecting an already fine-tuned model as a base model and fine-tuning it further on new sets of training examples.
To perform fine-tuning on a model that you have previously fine-tuned you would use the same process as described in create a customized model but instead of specifying the name of a generic base model you would specify your already fine-tuned model. A custom fine-tuned model would look like gpt-35-turbo-0613.ft-5fd1918ee65d4cd38a5dcf6835066ed7
We also recommend including the suffix
parameter to make it easier to distinguish between different iterations of your fine-tuned model. suffix
takes a string, and is set to identify the fine-tuned model. With the OpenAI Python API a string of up to 18 characters is supported that will be added to your fine-tuned model name.
Important
Some of the features described in this article might only be available in preview. This preview is provided without a service-level agreement, and we don't recommend it for production workloads. Certain features might not be supported or might have constrained capabilities. For more information, see Supplemental Terms of Use for Microsoft Azure Previews.
Prerequisites
An Azure subscription - Create one for free.
An Azure AI project in Azure AI Studio.
An Azure OpenAI connection to a resource in a region where fine-tuning is supported.
Fine-tuning access requires Cognitive Services OpenAI Contributor role on the Azure OpenAI resource.
If you don't already have access to view quota and deploy models in Azure AI Studio you need more permissions.
Models
The following models support fine-tuning:
babbage-002
davinci-002
gpt-35-turbo
(0613)gpt-35-turbo
(1106)gpt-35-turbo
(0125)gpt-4
(0613)*gpt-4o-mini
(2024-07-18)*
* Fine-tuning for this model is currently in public preview.
Consult the models page to check which regions currently support fine-tuning.
If you plan to use gpt-4
for fine-tuning, please refer to the GPT-4 public preview safety evaluation guidance
Review the workflow for Azure AI Studio
Take a moment to review the fine-tuning workflow for using Azure AI Studio:
- Prepare your training and validation data.
- Use the Fine-tune model wizard in Azure AI Studio to train your custom model.
- Select a model.
- Choose your training data.
- Optionally, choose your validation data.
- Optionally, configure your parameters for your fine-tuning job.
- Review your choices and train your new custom model.
- Check the status of your fine-tuned model.
- Optionally, analyze your fine-tuned model for performance and fit.
- Deploy your fine-tuned model for use.
- Use your fine-tuned model.
Prepare your training and validation data
Your training data and validation data sets consist of input and output examples for how you would like the model to perform.
Different model types require a different format of training data.
The training and validation data you use must be formatted as a JSON Lines (JSONL) document. For gpt-35-turbo-0613
the fine-tuning dataset must be formatted in the conversational format that is used by the Chat completions API.
If you would like a step-by-step walk-through of fine-tuning a gpt-35-turbo-0613
model please refer to the Azure OpenAI fine-tuning tutorial.
Example file format
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "Is Xbox better than PlayStation?"}, {"role": "assistant", "content": "I apologize, but I cannot provide personal opinions. My primary job is to assist you with any issues related to your Xbox device. Do you have any Xbox-related issues that need addressing?"}]}
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "Can I play PC games on my Xbox?"}, {"role": "assistant", "content": "While Xbox consoles don't natively support PC games, they can play an array of Xbox games, many of which are also available on PC. Can I assist you with finding games compatible with your Xbox device?"}]}
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "I'm having trouble connecting my Xbox to the Wi-Fi."}, {"role": "assistant", "content": "No worries, let's go through the network settings on your Xbox. Can you please tell me what happens when you try to connect it to the Wi-Fi?"}]}
Multi-turn chat file format
Multiple turns of a conversation in a single line of your jsonl training file is also supported. To skip fine-tuning on specific assistant messages add the optional weight
key value pair. Currently weight
can be set to 0 or 1.
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "What's the capital of France?"}, {"role": "assistant", "content": "Paris", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Paris, as if everyone doesn't know that already.", "weight": 1}]}
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "Who wrote 'Romeo and Juliet'?"}, {"role": "assistant", "content": "William Shakespeare", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Oh, just some guy named William Shakespeare. Ever heard of him?", "weight": 1}]}
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "How far is the Moon from Earth?"}, {"role": "assistant", "content": "384,400 kilometers", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Around 384,400 kilometers. Give or take a few, like that really matters.", "weight": 1}]}
In addition to the JSONL format, training and validation data files must be encoded in UTF-8 and include a byte-order mark (BOM). The file must be less than 512 MB in size.
Create your training and validation datasets
The more training examples you have, the better. Fine tuning jobs will not proceed without at least 10 training examples, but such a small number are not enough to noticeably influence model responses. It is best practice to provide hundreds, if not thousands, of training examples to be successful.
In general, doubling the dataset size can lead to a linear increase in model quality. But keep in mind, low quality examples can negatively impact performance. If you train the model on a large amount of internal data, without first pruning the dataset for only the highest quality examples you could end up with a model that performs much worse than expected.
Create your fine-tuned model
To fine-tune an Azure OpenAI model in an existing Azure AI Studio project, follow these steps:
Sign in to Azure AI Studio and select your project. If you don't have a project already, first create a project.
From the collapsible left menu, select Fine-tuning > + Fine-tune model.
Select the base model
Select a base model to fine-tune. Your choice influences both the performance and the cost of your model. In this example, we are choosing the
gpt-35-turbo
model. Then select Confirm.For
gpt-35-turbo
we have different versions available for fine-tuning, so please choose which version you'd like to fine-tune. We will choose (0301).We also recommend including the
suffix
parameter to make it easier to distinguish between different iterations of your fine-tuned model.suffix
takes a string, and is set to identify the fine-tuned model. With the OpenAI Python API a string of up to 18 characters is supported that will be added to your fine-tuned model name.
If you have more than one Azure OpenAI connection enabled for fine-tuning, then choose which resource you'd like to use. Please note that all users with access to your Azure OpenAI resource will have access to this fine-tuned model.
Choose your training data
The next step is to either choose existing prepared training data or upload new prepared training data to use when customizing your model. The Training data pane displays any existing, previously uploaded datasets and also provides options to upload new training data.
If your training data is already in your project, select Data in Azure AI Studio.
- Select the file from the list shown in the Training data pane.
If your training data is already uploaded to the Azure OpenAI service, select your Azure OpenAI connection under Azure OpenAI Connection.
To upload training data to fine-tune your model, select Upload data and then select Upload file.
- Make sure all your training examples follow the expected format for inference. To fine-tune models effectively, ensure a balanced and diverse dataset. This involves maintaining data balance, including various scenarios, and periodically refining training data to align with real-world expectations, ultimately leading to more accurate and balanced model responses. See data preparation for more information.
- For large data files, we recommend that you import from an Azure Blob store. Large files can become unstable when uploaded through multipart forms because the requests are atomic and can't be retried or resumed. For more information about Azure Blob Storage, see What is Azure Blob Storage?
Note
Training data files must be formatted as JSONL files, encoded in UTF-8 with a byte-order mark (BOM). The file must be less than 512 MB in size.
After uploading files, you will see a preview of your training data. Select Next to continue.
Choose your validation data
Optionally, you can choose to provide validation data to fine-tune your model. If you don't want to use validation data, you can select None and select Next to continue to the advanced options for the model.
Otherwise, if you have a validation dataset, you can either choose from previously uploaded data or upload newly prepared validation data to use for fine-tuning your model.
Note
Similar to training data files, validation data files must be formatted as JSONL files, encoded in UTF-8 with a byte-order mark (BOM). The file must be less than 512 MB in size.
Configure your parameters
Optionally, configure parameters for your fine-tuning job. The following are available:
Name | Type | Description |
---|---|---|
batch_size |
integer | The batch size to use for training. The batch size is the number of training examples used to train a single forward and backward pass. In general, we've found that larger batch sizes tend to work better for larger datasets. The default value as well as the maximum value for this property are specific to a base model. A larger batch size means that model parameters are updated less frequently, but with lower variance. When set to -1, batch_size is calculated as 0.2% of examples in training set and the max is 256. |
learning_rate_multiplier |
number | The learning rate multiplier to use for training. The fine-tuning learning rate is the original learning rate used for pre-training multiplied by this value. Larger learning rates tend to perform better with larger batch sizes. We recommend experimenting with values in the range 0.02 to 0.2 to see what produces the best results. A smaller learning rate may be useful to avoid overfitting. |
n_epochs |
integer | The number of epochs to train the model for. An epoch refers to one full cycle through the training dataset. If set to -1, the number of epochs is determined dynamically based on the input data. |
seed |
integer | The seed controls the reproducibility of the job. Passing in the same seed and job parameters should produce the same results, but may differ in rare cases. If a seed isn't specified, one will be generated for you. |
You can choose to leave the default configuration or customize the values to your preference. After you finish making your configurations, select Next.
Review your choices and train your model
Review your choices and select Submit to start training your new fine-tuned model.
Check the status of your fine-tuned model
After you submit your fine-tuning job, you see a page with details about your fine-tuned model. You can find the status and more information about your fine-tuned model on the Fine-tuning > Models page in Azure AI Studio.
Your job might be queued behind other jobs on the system. Training your model can take minutes or hours depending on the model and dataset size.
Analyze your fine-tuned model
After fine-tuning is successfully completed, you can download a result file named results.csv from the fine-tuned model page under the Details tab. You can use the result file to analyze the training and validation performance of your custom model.
The result file is a CSV file that contains a header row and a row for each training step performed by the fine-tuning job. The result file contains the following columns:
Column name | Description |
---|---|
step |
The number of the training step. A training step represents a single pass, forward and backward, on a batch of training data. |
train_loss |
The loss for the training batch. |
train_mean_token_accuracy |
The percentage of tokens in the training batch correctly predicted by the model. For example, if the batch size is set to 3 and your data contains completions [[1, 2], [0, 5], [4, 2]] , this value is set to 0.83 (5 of 6) if the model predicted [[1, 1], [0, 5], [4, 2]] . |
valid_loss |
The loss for the validation batch. |
validation_mean_token_accuracy |
The percentage of tokens in the validation batch correctly predicted by the model. For example, if the batch size is set to 3 and your data contains completions [[1, 2], [0, 5], [4, 2]] , this value is set to 0.83 (5 of 6) if the model predicted [[1, 1], [0, 5], [4, 2]] . |
full_valid_loss |
The validation loss calculated at the end of each epoch. When training goes well, loss should decrease. |
full_valid_mean_token_accuracy |
The valid mean token accuracy calculated at the end of each epoch. When training is going well, token accuracy should increase. |
You can also view the data in your results.csv file as plots in Azure AI Studio under the Metrics tab of your fine-tuned model. Select the link for your trained model, and you will see two charts: loss, and token accuracy. If you provided validation data, both datasets will appear on the same plot.
Look for your loss to decrease over time, and your accuracy to increase. If you see a divergence between your training and validation data that may indicate that you are overfitting. Try training with fewer epochs, or a smaller learning rate multiplier.
Checkpoints
When each training epoch completes a checkpoint is generated. A checkpoint is a fully functional version of a model which can both be deployed and used as the target model for subsequent fine-tuning jobs. Checkpoints can be particularly useful, as they can provide a snapshot of your model prior to overfitting having occurred. When a fine-tuning job completes you will have the three most recent versions of the model available to deploy.
Safety evaluation GPT-4 fine-tuning - public preview
GPT-4 and GPT-4o-mini are our most advanced models that can be fine-tuned to your needs. As with Azure OpenAI models generally, the advanced capabilities of fine-tuned models come with increased responsible AI challenges related to harmful content, manipulation, human-like behavior, privacy issues, and more. Learn more about risks, capabilities, and limitations in the Overview of Responsible AI practices and Transparency Note. To help mitigate the risks associated with GPT-4 and GPT-4o-mini fine-tuned models, we have implemented additional evaluation steps to help detect and prevent harmful content in the training and outputs of fine-tuned models. These steps are grounded in the Microsoft Responsible AI Standard and Azure OpenAI Service content filtering.
- Evaluations are conducted in dedicated, customer specific, private workspaces;
- Evaluation endpoints are in the same geography as the Azure OpenAI resource;
- Training data is not stored in connection with performing evaluations; only the final model assessment (deployable or not deployable) is persisted; and
GPT-4 and GPT-4o-mini fine-tuned model evaluation filters are set to predefined thresholds and cannot be modified by customers; they aren't tied to any custom content filtering configuration you may have created.
Data evaluation
Before training starts, your data is evaluated for potentially harmful content (violence, sexual, hate, and fairness, self-harm – see category definitions here). If harmful content is detected above the specified severity level, your training job will fail, and you'll receive a message informing you of the categories of failure.
Sample message:
The provided training data failed RAI checks for harm types: [hate_fairness, self_harm, violence]. Please fix the data and try again.
Your training data is evaluated automatically within your data import job as part of providing the fine-tuning capability.
If the fine-tuning job fails due to the detection of harmful content in training data, you won't be charged.
Model evaluation
After training completes but before the fine-tuned model is available for deployment, the resulting model is evaluated for potentially harmful responses using Azure’s built-in risk and safety metrics. Using the same approach to testing that we use for the base large language models, our evaluation capability simulates a conversation with your fine-tuned model to assess the potential to output harmful content, again using specified harmful content categories (violence, sexual, hate, and fairness, self-harm).
If a model is found to generate output containing content detected as harmful at above an acceptable rate, you'll be informed that your model isn't available for deployment, with information about the specific categories of harm detected:
Sample Message:
This model is unable to be deployed. Model evaluation identified that this fine tuned model scores above acceptable thresholds for [Violence, Self Harm]. Please review your training data set and resubmit the job.
As with data evaluation, the model is evaluated automatically within your fine-tuning job as part of providing the fine-tuning capability. Only the resulting assessment (deployable or not deployable) is logged by the service. If deployment of the fine-tuned model fails due to the detection of harmful content in model outputs, you won't be charged for the training run.
Deploy a fine-tuned model
Once your model is fine-tuned, you can deploy the model and can use it in your own application. You can't deploy a fine-tuned model from the deployments page or the playground page in Azure AI Studio. The only way, currently, to deploy a fine-tuned model is from the model details page for that model.
When you deploy the model, you make the model available for inferencing, and that incurs an hourly hosting charge. Fine-tuned models, however, can be stored in Azure AI Studio at no cost until you're ready to use them.
Important
After you deploy a customized model, if at any time the deployment remains inactive for greater than fifteen (15) days, the deployment is deleted. The deployment of a customized model is inactive if the model was deployed more than fifteen (15) days ago and no completions or chat completions calls were made to it during a continuous 15-day period.
The deletion of an inactive deployment doesn't delete or affect the underlying customized model, and the customized model can be redeployed at any time. As described in Azure OpenAI Service pricing, each customized (fine-tuned) model that's deployed incurs an hourly hosting cost regardless of whether completions or chat completions calls are being made to the model. To learn more about planning and managing costs with Azure OpenAI, refer to the guidance in Plan to manage costs for Azure OpenAI Service.
Note
Only one deployment is permitted for a fine-tuned model. An error message is displayed if you select an already-deployed fine-tuned model.
You can monitor the progress of your deployment on the Deployments page in Azure AI Studio.
Use a deployed fine-tuned model
After your fine-tuned model deploys, you can use it like any other deployed model. You can use the Playground in Azure AI Studio to experiment with your new deployment. You can also use the REST API to call your fine-tuned model from your own application. You can even begin to use this new fine-tuned model in your prompt flow to build your generative AI application.
Note
For chat models, the system message that you use to guide your fine-tuned model (whether it's deployed or available for testing in the playground) must be the same as the system message you used for training. If you use a different system message, the model might not perform as expected.
Clean up your deployments, fine-tuned models, and training files
When you're done with your fine-tuned model, you can delete the deployment and model. You can also delete the training and validation files you uploaded to the service, if needed.
Delete your fine-tuned model deployment
Important
After you deploy a customized model, if at any time the deployment remains inactive for greater than fifteen (15) days, the deployment is deleted. The deployment of a customized model is inactive if the model was deployed more than fifteen (15) days ago and no completions or chat completions calls were made to it during a continuous 15-day period.
The deletion of an inactive deployment doesn't delete or affect the underlying customized model, and the customized model can be redeployed at any time. As described in Azure OpenAI Service pricing, each customized (fine-tuned) model that's deployed incurs an hourly hosting cost regardless of whether completions or chat completions calls are being made to the model. To learn more about planning and managing costs with Azure OpenAI, refer to the guidance in Plan to manage costs for Azure OpenAI Service.
You can delete the deployment for your fine-tuned model on the Deployments page in Azure AI Studio. Select the deployment to delete, and then select Delete to delete the deployment.
Delete your fine-tuned model
You can delete a fine-tuned model on the Fine-tuning page in Azure AI Studio. Select the fine-tuned model to delete and then select Delete to delete the fine-tuned model.
Note
You can't delete a fine-tuned model if it has an existing deployment. You must first delete your model deployment before you can delete your fine-tuned model.
Delete your training files
You can optionally delete training and validation files that you uploaded for training, and result files generated during training. For this you need to go to Azure OpenAI Studio and navigate to the Management > Data files pane. Select the file to delete, and then select Delete to delete the file.
Prerequisites
- Read the When to use Azure OpenAI fine-tuning guide.
- An Azure subscription. Create one for free.
- An Azure OpenAI resource. For more information, see Create a resource and deploy a model with Azure OpenAI.
- The following Python libraries:
os
,json
,requests
,openai
. - The OpenAI Python library should be at least version 0.28.1.
- Fine-tuning access requires Cognitive Services OpenAI Contributor.
- If you do not already have access to view quota, and deploy models in Azure OpenAI Studio you will require additional permissions.
Models
The following models support fine-tuning:
babbage-002
davinci-002
gpt-35-turbo
(0613)gpt-35-turbo
(1106)gpt-35-turbo
(0125)gpt-4
(0613)*gpt-4o-mini
(2024-07-18)*
* Fine-tuning for this model is currently in public preview.
If you plan to use gpt-4
for fine-tuning, please refer to the GPT-4 public preview safety evaluation guidance
Or you can fine tune a previously fine-tuned model, formatted as base-model.ft-{jobid}.
Consult the models page to check which regions currently support fine-tuning.
Review the workflow for the Python SDK
Take a moment to review the fine-tuning workflow for using the Python SDK with Azure OpenAI:
- Prepare your training and validation data.
- Select a base model.
- Upload your training data.
- Train your new customized model.
- Check the status of your customized model.
- Deploy your customized model for use.
- Use your customized model.
- Optionally, analyze your customized model for performance and fit.
Prepare your training and validation data
Your training data and validation data sets consist of input and output examples for how you would like the model to perform.
Different model types require a different format of training data.
The training and validation data you use must be formatted as a JSON Lines (JSONL) document. For gpt-35-turbo-0613
the fine-tuning dataset must be formatted in the conversational format that is used by the Chat completions API.
If you would like a step-by-step walk-through of fine-tuning a gpt-35-turbo-0613
please refer to the Azure OpenAI fine-tuning tutorial
Example file format
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "Is Xbox better than PlayStation?"}, {"role": "assistant", "content": "I apologize, but I cannot provide personal opinions. My primary job is to assist you with any issues related to your Xbox device. Do you have any Xbox-related issues that need addressing?"}]}
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "Can I play PC games on my Xbox?"}, {"role": "assistant", "content": "While Xbox consoles don't natively support PC games, they can play an array of Xbox games, many of which are also available on PC. Can I assist you with finding games compatible with your Xbox device?"}]}
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "I'm having trouble connecting my Xbox to the Wi-Fi."}, {"role": "assistant", "content": "No worries, let's go through the network settings on your Xbox. Can you please tell me what happens when you try to connect it to the Wi-Fi?"}]}
Multi-turn chat file format
Multiple turns of a conversation in a single line of your jsonl training file is also supported. To skip fine-tuning on specific assistant messages add the optional weight
key value pair. Currently weight
can be set to 0 or 1.
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "What's the capital of France?"}, {"role": "assistant", "content": "Paris", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Paris, as if everyone doesn't know that already.", "weight": 1}]}
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "Who wrote 'Romeo and Juliet'?"}, {"role": "assistant", "content": "William Shakespeare", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Oh, just some guy named William Shakespeare. Ever heard of him?", "weight": 1}]}
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "How far is the Moon from Earth?"}, {"role": "assistant", "content": "384,400 kilometers", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Around 384,400 kilometers. Give or take a few, like that really matters.", "weight": 1}]}
In addition to the JSONL format, training and validation data files must be encoded in UTF-8 and include a byte-order mark (BOM). The file must be less than 512 MB in size.
Create your training and validation datasets
The more training examples you have, the better. Fine tuning jobs will not proceed without at least 10 training examples, but such a small number are not enough to noticeably influence model responses. It is best practice to provide hundreds, if not thousands, of training examples to be successful.
In general, doubling the dataset size can lead to a linear increase in model quality. But keep in mind, low quality examples can negatively impact performance. If you train the model on a large amount of internal data, without first pruning the dataset for only the highest quality examples you could end up with a model that performs much worse than expected.
Upload your training data
The next step is to either choose existing prepared training data or upload new prepared training data to use when customizing your model. After you prepare your training data, you can upload your files to the service. There are two ways to upload training data:
For large data files, we recommend that you import from an Azure Blob store. Large files can become unstable when uploaded through multipart forms because the requests are atomic and can't be retried or resumed. For more information about Azure Blob storage, see What is Azure Blob storage?
Note
Training data files must be formatted as JSONL files, encoded in UTF-8 with a byte-order mark (BOM). The file must be less than 512 MB in size.
The following Python example uploads local training and validation files by using the Python SDK, and retrieves the returned file IDs.
# Upload fine-tuning files
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 capabilities
)
training_file_name = 'training_set.jsonl'
validation_file_name = 'validation_set.jsonl'
# Upload the training and validation dataset files to Azure OpenAI with the SDK.
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)
Create a customized model
After you upload your training and validation files, you're ready to start the fine-tuning job.
The following Python code shows an example of how to create a new fine-tune job with the Python SDK:
In this example we are also passing the seed parameter. The seed controls the reproducibility of the job. Passing in the same seed and job parameters should produce the same results, but may differ in rare cases. If a seed isn't specified, one will be generated for you.
response = client.fine_tuning.jobs.create(
training_file=training_file_id,
validation_file=validation_file_id,
model="gpt-35-turbo-0613", # Enter base model name. Note that in Azure OpenAI the model name contains dashes and cannot contain dot/period characters.
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
# You can use the job ID to monitor the status of the fine-tuning job.
# The fine-tuning job will take some time to start and complete.
print("Job ID:", response.id)
print("Status:", response.id)
print(response.model_dump_json(indent=2))
You can also pass additional optional parameters like hyperparameters to take greater control of the fine-tuning process. For initial training we recommend using the automatic defaults that are present without specifying these parameters.
The current supported hyperparameters for fine-tuning are:
Name | Type | Description |
---|---|---|
batch_size |
integer | The batch size to use for training. The batch size is the number of training examples used to train a single forward and backward pass. In general, we've found that larger batch sizes tend to work better for larger datasets. The default value as well as the maximum value for this property are specific to a base model. A larger batch size means that model parameters are updated less frequently, but with lower variance. |
learning_rate_multiplier |
number | The learning rate multiplier to use for training. The fine-tuning learning rate is the original learning rate used for pre-training multiplied by this value. Larger learning rates tend to perform better with larger batch sizes. We recommend experimenting with values in the range 0.02 to 0.2 to see what produces the best results. A smaller learning rate can be useful to avoid overfitting. |
n_epochs |
integer | The number of epochs to train the model for. An epoch refers to one full cycle through the training dataset. |
To set custom hyperparameters with the 1.x version of the OpenAI Python API:
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" # This API version or later is required to access fine-tuning for turbo/babbage-002/davinci-002
)
client.fine_tuning.jobs.create(
training_file="file-abc123",
model="gpt-35-turbo-0613", # Enter base model name. Note that in Azure OpenAI the model name contains dashes and cannot contain dot/period characters.
hyperparameters={
"n_epochs":2
}
)
Check fine-tuning job status
response = client.fine_tuning.jobs.retrieve(job_id)
print("Job ID:", response.id)
print("Status:", response.status)
print(response.model_dump_json(indent=2))
List fine-tuning events
To examine the individual fine-tuning events that were generated during training:
You may need to upgrade your OpenAI client library to the latest version with pip install openai --upgrade
to run this command.
response = client.fine_tuning.jobs.list_events(fine_tuning_job_id=job_id, limit=10)
print(response.model_dump_json(indent=2))
Checkpoints
When each training epoch completes a checkpoint is generated. A checkpoint is a fully functional version of a model which can both be deployed and used as the target model for subsequent fine-tuning jobs. Checkpoints can be particularly useful, as they can provide a snapshot of your model prior to overfitting having occurred. When a fine-tuning job completes you will have the three most recent versions of the model available to deploy. The final epoch will be represented by your fine-tuned model, the previous two epochs will be available as checkpoints.
You can run the list checkpoints command to retrieve the list of checkpoints associated with an individual fine-tuning job:
You may need to upgrade your OpenAI client library to the latest version with pip install openai --upgrade
to run this command.
response = client.fine_tuning.jobs.list_events(fine_tuning_job_id=job_id, limit=10)
print(response.model_dump_json(indent=2))
Safety evaluation GPT-4 fine-tuning - public preview
GPT-4 and GPT-4o-mini are our most advanced models that can be fine-tuned to your needs. As with Azure OpenAI models generally, the advanced capabilities of fine-tuned models come with increased responsible AI challenges related to harmful content, manipulation, human-like behavior, privacy issues, and more. Learn more about risks, capabilities, and limitations in the Overview of Responsible AI practices and Transparency Note. To help mitigate the risks associated with GPT-4 and GPT-4o-mini fine-tuned models, we have implemented additional evaluation steps to help detect and prevent harmful content in the training and outputs of fine-tuned models. These steps are grounded in the Microsoft Responsible AI Standard and Azure OpenAI Service content filtering.
- Evaluations are conducted in dedicated, customer specific, private workspaces;
- Evaluation endpoints are in the same geography as the Azure OpenAI resource;
- Training data is not stored in connection with performing evaluations; only the final model assessment (deployable or not deployable) is persisted; and
GPT-4 and GPT-4o-mini fine-tuned model evaluation filters are set to predefined thresholds and cannot be modified by customers; they aren't tied to any custom content filtering configuration you may have created.
Data evaluation
Before training starts, your data is evaluated for potentially harmful content (violence, sexual, hate, and fairness, self-harm – see category definitions here). If harmful content is detected above the specified severity level, your training job will fail, and you'll receive a message informing you of the categories of failure.
Sample message:
The provided training data failed RAI checks for harm types: [hate_fairness, self_harm, violence]. Please fix the data and try again.
Your training data is evaluated automatically within your data import job as part of providing the fine-tuning capability.
If the fine-tuning job fails due to the detection of harmful content in training data, you won't be charged.
Model evaluation
After training completes but before the fine-tuned model is available for deployment, the resulting model is evaluated for potentially harmful responses using Azure’s built-in risk and safety metrics. Using the same approach to testing that we use for the base large language models, our evaluation capability simulates a conversation with your fine-tuned model to assess the potential to output harmful content, again using specified harmful content categories (violence, sexual, hate, and fairness, self-harm).
If a model is found to generate output containing content detected as harmful at above an acceptable rate, you'll be informed that your model isn't available for deployment, with information about the specific categories of harm detected:
Sample Message:
This model is unable to be deployed. Model evaluation identified that this fine tuned model scores above acceptable thresholds for [Violence, Self Harm]. Please review your training data set and resubmit the job.
As with data evaluation, the model is evaluated automatically within your fine-tuning job as part of providing the fine-tuning capability. Only the resulting assessment (deployable or not deployable) is logged by the service. If deployment of the fine-tuned model fails due to the detection of harmful content in model outputs, you won't be charged for the training run.
Deploy a fine-tuned model
When the fine-tuning job succeeds, the value of the fine_tuned_model
variable in the response body is set to the name of your customized model. Your model is now also available for discovery from the list Models API. However, you can't issue completion calls to your customized model until your customized model is deployed. You must deploy your customized model to make it available for use with completion calls.
Important
After you deploy a customized model, if at any time the deployment remains inactive for greater than fifteen (15) days, the deployment is deleted. The deployment of a customized model is inactive if the model was deployed more than fifteen (15) days ago and no completions or chat completions calls were made to it during a continuous 15-day period.
The deletion of an inactive deployment doesn't delete or affect the underlying customized model, and the customized model can be redeployed at any time. As described in Azure OpenAI Service pricing, each customized (fine-tuned) model that's deployed incurs an hourly hosting cost regardless of whether completions or chat completions calls are being made to the model. To learn more about planning and managing costs with Azure OpenAI, refer to the guidance in Plan to manage costs for Azure OpenAI Service.
You can also use Azure OpenAI Studio or the Azure CLI to deploy your customized model.
Note
Only one deployment is permitted for a customized model. An error occurs if you select an already-deployed customized model.
Unlike the previous SDK commands, deployment must be done using the control plane API which requires separate authorization, a different API path, and a different API version.
variable | Definition |
---|---|
token | There are multiple ways to generate an authorization token. The easiest method for initial testing is to launch the Cloud Shell from the Azure portal. Then run az account get-access-token . You can use this token as your temporary authorization token for API testing. We recommend storing this in a new environment variable. |
subscription | The subscription ID for the associated Azure OpenAI resource. |
resource_group | The resource group name for your Azure OpenAI resource. |
resource_name | The Azure OpenAI resource name. |
model_deployment_name | The custom name for your new fine-tuned model deployment. This is the name that will be referenced in your code when making chat completion calls. |
fine_tuned_model | Retrieve this value from your fine-tuning job results in the previous step. It will look like gpt-35-turbo-0613.ft-b044a9d3cf9c4228b5d393567f693b83 . You will need to add that value to the deploy_data json. Alternatively you can also deploy a checkpoint, by passing the checkpoint ID which will appear in the format ftchkpt-e559c011ecc04fc68eaa339d8227d02d |
import json
import os
import requests
token= os.getenv("<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" # custom deployment name that you will use to reference the model when making inference calls.
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": <"fine_tuned_model">, #retrieve this value from the previous call, it will look like gpt-35-turbo-0613.ft-b044a9d3cf9c4228b5d393567f693b83
"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())
Cross region deployment
Fine-tuning supports deploying a fine-tuned model to a different region than where the model was originally fine-tuned. You can also deploy to a different subscription/region.
The only limitations are that the new region must also support fine-tuning and when deploying cross subscription the account generating the authorization token for the deployment must have access to both the source and destination subscriptions.
Below is an example of deploying a model that was fine-tuned in one subscription/region to another.
import json
import os
import requests
token= os.getenv("<TOKEN>")
subscription = "<DESTINATION_SUBSCRIPTION_ID>"
resource_group = "<DESTINATION_RESOURCE_GROUP_NAME>"
resource_name = "<DESTINATION_AZURE_OPENAI_RESOURCE_NAME>"
source_subscription = "<SOURCE_SUBSCRIPTION_ID>"
source_resource_group = "<SOURCE_RESOURCE_GROUP>"
source_resource = "<SOURCE_RESOURCE>"
source = f'/subscriptions/{source_subscription}/resourceGroups/{source_resource_group}/providers/Microsoft.CognitiveServices/accounts/{source_resource}'
model_deployment_name ="gpt-35-turbo-ft" # custom deployment name that you will use to reference the model when making inference calls.
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": <"FINE_TUNED_MODEL_NAME">, # This value will look like gpt-35-turbo-0613.ft-0ab3f80e4f2242929258fff45b56a9ce
"version": "1",
"source": source
}
}
}
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())
To deploy between the same subscription, but different regions you would just have subscription and resource groups be identical for both source and destination variables and only the source and destination resource names would need to be unique.
Deploy a model with Azure CLI
The following example shows how to use the Azure CLI to deploy your customized model. With the Azure CLI, you must specify a name for the deployment of your customized model. For more information about how to use the Azure CLI to deploy customized models, see az cognitiveservices account deployment
.
To run this Azure CLI command in a console window, you must replace the following <placeholders> with the corresponding values for your customized model:
Placeholder | Value |
---|---|
<YOUR_AZURE_SUBSCRIPTION> | The name or ID of your Azure subscription. |
<YOUR_RESOURCE_GROUP> | The name of your Azure resource group. |
<YOUR_RESOURCE_NAME> | The name of your Azure OpenAI resource. |
<YOUR_DEPLOYMENT_NAME> | The name you want to use for your model deployment. |
<YOUR_FINE_TUNED_MODEL_ID> | The name of your customized model. |
az cognitiveservices account deployment create
--resource-group <YOUR_RESOURCE_GROUP>
--name <YOUR_RESOURCE_NAME>
--deployment-name <YOUR_DEPLOYMENT_NAME>
--model-name <YOUR_FINE_TUNED_MODEL_ID>
--model-version "1"
--model-format OpenAI
--sku-capacity "1"
--sku-name "Standard"
Use a deployed customized model
After your custom model deploys, you can use it like any other deployed model. You can use the Playgrounds in Azure OpenAI Studio to experiment with your new deployment. You can continue to use the same parameters with your custom model, such as temperature
and max_tokens
, as you can with other deployed models. For fine-tuned babbage-002
and davinci-002
models you will use the Completions playground and the Completions API. For fine-tuned gpt-35-turbo-0613
models you will use the Chat playground and the Chat completion API.
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)
Analyze your customized model
Azure OpenAI attaches a result file named results.csv to each fine-tune job after it completes. You can use the result file to analyze the training and validation performance of your customized model. The file ID for the result file is listed for each customized model, and you can use the Python SDK to retrieve the file ID and download the result file for analysis.
The following Python example retrieves the file ID of the first result file attached to the fine-tuning job for your customized model, and then uses the Python SDK to download the file to your working directory for analysis.
# Retrieve the file ID of the first result file from the fine-tuning job
# for the customized model.
response = client.fine_tuning.jobs.retrieve(job_id)
if response.status == 'succeeded':
result_file_id = response.result_files[0]
retrieve = client.files.retrieve(result_file_id)
# Download the result file.
print(f'Downloading result file: {result_file_id}')
with open(retrieve.filename, "wb") as file:
result = client.files.content(result_file_id).read()
file.write(result)
The result file is a CSV file that contains a header row and a row for each training step performed by the fine-tuning job. The result file contains the following columns:
Column name | Description |
---|---|
step |
The number of the training step. A training step represents a single pass, forward and backward, on a batch of training data. |
train_loss |
The loss for the training batch. |
train_mean_token_accuracy |
The percentage of tokens in the training batch correctly predicted by the model. For example, if the batch size is set to 3 and your data contains completions [[1, 2], [0, 5], [4, 2]] , this value is set to 0.83 (5 of 6) if the model predicted [[1, 1], [0, 5], [4, 2]] . |
valid_loss |
The loss for the validation batch. |
validation_mean_token_accuracy |
The percentage of tokens in the validation batch correctly predicted by the model. For example, if the batch size is set to 3 and your data contains completions [[1, 2], [0, 5], [4, 2]] , this value is set to 0.83 (5 of 6) if the model predicted [[1, 1], [0, 5], [4, 2]] . |
full_valid_loss |
The validation loss calculated at the end of each epoch. When training goes well, loss should decrease. |
full_valid_mean_token_accuracy |
The valid mean token accuracy calculated at the end of each epoch. When training is going well, token accuracy should increase. |
You can also view the data in your results.csv file as plots in Azure OpenAI Studio. Select the link for your trained model, and you will see three charts: loss, mean token accuracy, and token accuracy. If you provided validation data, both datasets will appear on the same plot.
Look for your loss to decrease over time, and your accuracy to increase. If you see a divergence between your training and validation data that can indicate that you are overfitting. Try training with fewer epochs, or a smaller learning rate multiplier.
Clean up your deployments, customized models, and training files
When you're done with your customized model, you can delete the deployment and model. You can also delete the training and validation files you uploaded to the service, if needed.
Delete your model deployment
Important
After you deploy a customized model, if at any time the deployment remains inactive for greater than fifteen (15) days, the deployment is deleted. The deployment of a customized model is inactive if the model was deployed more than fifteen (15) days ago and no completions or chat completions calls were made to it during a continuous 15-day period.
The deletion of an inactive deployment doesn't delete or affect the underlying customized model, and the customized model can be redeployed at any time. As described in Azure OpenAI Service pricing, each customized (fine-tuned) model that's deployed incurs an hourly hosting cost regardless of whether completions or chat completions calls are being made to the model. To learn more about planning and managing costs with Azure OpenAI, refer to the guidance in Plan to manage costs for Azure OpenAI Service.
You can use various methods to delete the deployment for your customized model:
Delete your customized model
Similarly, you can use various methods to delete your customized model:
Note
You can't delete a customized model if it has an existing deployment. You must first delete your model deployment before you can delete your customized model.
Delete your training files
You can optionally delete training and validation files that you uploaded for training, and result files generated during training, from your Azure OpenAI subscription. You can use the following methods to delete your training, validation, and result files:
- Azure OpenAI Studio
- The REST APIs
- The Python SDK
The following Python example uses the Python SDK to delete the training, validation, and result files for your customized model:
print('Checking for existing uploaded files.')
results = []
# Get the complete list of uploaded files in our subscription.
files = openai.File.list().data
print(f'Found {len(files)} total uploaded files in the subscription.')
# Enumerate all uploaded files, extracting the file IDs for the
# files with file names that match your training dataset file and
# validation dataset file names.
for item in files:
if item["filename"] in [training_file_name, validation_file_name, result_file_name]:
results.append(item["id"])
print(f'Found {len(results)} already uploaded files that match our files')
# Enumerate the file IDs for our files and delete each file.
print(f'Deleting already uploaded files.')
for id in results:
openai.File.delete(sid = id)
Continuous fine-tuning
Once you have created a fine-tuned model you might want to continue to refine the model over time through further fine-tuning. Continuous fine-tuning is the iterative process of selecting an already fine-tuned model as a base model and fine-tuning it further on new sets of training examples.
To perform fine-tuning on a model that you have previously fine-tuned you would use the same process as described in create a customized model but instead of specifying the name of a generic base model you would specify your already fine-tuned model's ID. The fine-tuned model ID looks like gpt-35-turbo-0613.ft-5fd1918ee65d4cd38a5dcf6835066ed7
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.fine_tuning.jobs.create(
training_file=training_file_id,
validation_file=validation_file_id,
model="gpt-35-turbo-0613.ft-5fd1918ee65d4cd38a5dcf6835066ed7" # Enter base model name. Note that in Azure OpenAI the model name contains dashes and cannot contain dot/period characters.
)
job_id = response.id
# You can use the job ID to monitor the status of the fine-tuning job.
# The fine-tuning job will take some time to start and complete.
print("Job ID:", response.id)
print("Status:", response.id)
print(response.model_dump_json(indent=2))
We also recommend including the suffix
parameter to make it easier to distinguish between different iterations of your fine-tuned model. suffix
takes a string, and is set to identify the fine-tuned model. With the OpenAI Python API a string of up to 18 characters is supported that will be added to your fine-tuned model name.
If you are unsure of the ID of your existing fine-tuned model this information can be found in the Models page of Azure OpenAI Studio, or you can generate a list of models for a given Azure OpenAI resource using the REST API.
Prerequisites
- Read the When to use Azure OpenAI fine-tuning guide.
- An Azure subscription. Create one for free.
- An Azure OpenAI resource. For more information, see Create a resource and deploy a model with Azure OpenAI.
- Fine-tuning access requires Cognitive Services OpenAI Contributor.
- If you don't already have access to view quota, and deploy models in Azure OpenAI Studio you'll require additional permissions.
Models
The following models support fine-tuning:
babbage-002
davinci-002
gpt-35-turbo
(0613)gpt-35-turbo
(1106)gpt-35-turbo
(0125)gpt-4
(0613)*gpt-4o-mini
(2024-07-18)*
* Fine-tuning for this model is currently in public preview.
Consult the models page to check which regions currently support fine-tuning.
If you plan to use gpt-4
for fine-tuning, please refer to the GPT-4 public preview safety evaluation guidance.
Review the workflow for the REST API
Take a moment to review the fine-tuning workflow for using the REST APIS and Python with Azure OpenAI:
- Prepare your training and validation data.
- Select a base model.
- Upload your training data.
- Train your new customized model.
- Check the status of your customized model.
- Deploy your customized model for use.
- Use your customized model.
- Optionally, analyze your customized model for performance and fit.
Prepare your training and validation data
Your training data and validation data sets consist of input and output examples for how you would like the model to perform.
Different model types require a different format of training data.
The training and validation data you use must be formatted as a JSON Lines (JSONL) document. For gpt-35-turbo-0613
and other related models, the fine-tuning dataset must be formatted in the conversational format that is used by the Chat completions API.
If you would like a step-by-step walk-through of fine-tuning a gpt-35-turbo-0613
please refer to the Azure OpenAI fine-tuning tutorial.
Example file format
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "Is Xbox better than PlayStation?"}, {"role": "assistant", "content": "I apologize, but I cannot provide personal opinions. My primary job is to assist you with any issues related to your Xbox device. Do you have any Xbox-related issues that need addressing?"}]}
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "Can I play PC games on my Xbox?"}, {"role": "assistant", "content": "While Xbox consoles don't natively support PC games, they can play an array of Xbox games, many of which are also available on PC. Can I assist you with finding games compatible with your Xbox device?"}]}
{"messages": [{"role": "system", "content": "You are an Xbox customer support agent whose primary goal is to help users with issues they are experiencing with their Xbox devices. You are friendly and concise. You only provide factual answers to queries, and do not provide answers that are not related to Xbox."}, {"role": "user", "content": "I'm having trouble connecting my Xbox to the Wi-Fi."}, {"role": "assistant", "content": "No worries, let's go through the network settings on your Xbox. Can you please tell me what happens when you try to connect it to the Wi-Fi?"}]}
Multi-turn chat file format
Multiple turns of a conversation in a single line of your jsonl training file is also supported. To skip fine-tuning on specific assistant messages add the optional weight
key value pair. Currently weight
can be set to 0 or 1.
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "What's the capital of France?"}, {"role": "assistant", "content": "Paris", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Paris, as if everyone doesn't know that already.", "weight": 1}]}
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "Who wrote 'Romeo and Juliet'?"}, {"role": "assistant", "content": "William Shakespeare", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Oh, just some guy named William Shakespeare. Ever heard of him?", "weight": 1}]}
{"messages": [{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."}, {"role": "user", "content": "How far is the Moon from Earth?"}, {"role": "assistant", "content": "384,400 kilometers", "weight": 0}, {"role": "user", "content": "Can you be more sarcastic?"}, {"role": "assistant", "content": "Around 384,400 kilometers. Give or take a few, like that really matters.", "weight": 1}]}
In addition to the JSONL format, training and validation data files must be encoded in UTF-8 and include a byte-order mark (BOM). The file must be less than 512 MB in size.
Create your training and validation datasets
The more training examples you have, the better. Fine tuning jobs will not proceed without at least 10 training examples, but such a small number are not enough to noticeably influence model responses. It is best practice to provide hundreds, if not thousands, of training examples to be successful.
In general, doubling the dataset size can lead to a linear increase in model quality. But keep in mind, low quality examples can negatively impact performance. If you train the model on a large amount of internal data without first pruning the dataset for only the highest quality examples, you could end up with a model that performs much worse than expected.
Select the base model
The first step in creating a custom model is to choose a base model. The Base model pane lets you choose a base model to use for your custom model. Your choice influences both the performance and the cost of your model.
Select the base model from the Base model type dropdown, and then select Next to continue.
You can create a custom model from one of the following available base models:
babbage-002
davinci-002
gpt-35-turbo
(0613)gpt-35-turbo
(1106)gpt-35-turbo
(0125)gpt-4
(0613)
Or you can fine tune a previously fine-tuned model, formatted as base-model.ft-{jobid}.
For more information about our base models that can be fine-tuned, see Models.
Upload your training data
The next step is to either choose existing prepared training data or upload new prepared training data to use when fine-tuning your model. After you prepare your training data, you can upload your files to the service. There are two ways to upload training data:
For large data files, we recommend that you import from an Azure Blob store. Large files can become unstable when uploaded through multipart forms because the requests are atomic and can't be retried or resumed. For more information about Azure Blob storage, see What is Azure Blob storage?
Note
Training data files must be formatted as JSONL files, encoded in UTF-8 with a byte-order mark (BOM). The file must be less than 512 MB in size.
Upload training data
curl -X POST $AZURE_OPENAI_ENDPOINT/openai/files?api-version=2023-12-01-preview \
-H "Content-Type: multipart/form-data" \
-H "api-key: $AZURE_OPENAI_API_KEY" \
-F "purpose=fine-tune" \
-F "file=@C:\\fine-tuning\\training_set.jsonl;type=application/json"
Upload validation data
curl -X POST $AZURE_OPENAI_ENDPOINT/openai/files?api-version=2023-12-01-preview \
-H "Content-Type: multipart/form-data" \
-H "api-key: $AZURE_OPENAI_API_KEY" \
-F "purpose=fine-tune" \
-F "file=@C:\\fine-tuning\\validation_set.jsonl;type=application/json"
Create a customized model
After you uploaded your training and validation files, you're ready to start the fine-tuning job. The following code shows an example of how to create a new fine-tuning job with the REST API.
In this example we are also passing the seed parameter. The seed controls the reproducibility of the job. Passing in the same seed and job parameters should produce the same results, but can differ in rare cases. If a seed is not specified, one will be generated for you.
curl -X POST $AZURE_OPENAI_ENDPOINT/openai/fine_tuning/jobs?api-version=2024-05-01-preview \
-H "Content-Type: application/json" \
-H "api-key: $AZURE_OPENAI_API_KEY" \
-d '{
"model": "gpt-35-turbo-0613",
"training_file": "<TRAINING_FILE_ID>",
"validation_file": "<VALIDATION_FILE_ID>",
"seed": 105
}'
You can also pass additional optional parameters like hyperparameters to take greater control of the fine-tuning process. For initial training we recommend using the automatic defaults that are present without specifying these parameters.
The current supported hyperparameters for fine-tuning are:
Name | Type | Description |
---|---|---|
batch_size |
integer | The batch size to use for training. The batch size is the number of training examples used to train a single forward and backward pass. In general, we've found that larger batch sizes tend to work better for larger datasets. The default value as well as the maximum value for this property are specific to a base model. A larger batch size means that model parameters are updated less frequently, but with lower variance. |
learning_rate_multiplier |
number | The learning rate multiplier to use for training. The fine-tuning learning rate is the original learning rate used for pre-training multiplied by this value. Larger learning rates tend to perform better with larger batch sizes. We recommend experimenting with values in the range 0.02 to 0.2 to see what produces the best results. A smaller learning rate can be useful to avoid overfitting. |
n_epochs |
integer | The number of epochs to train the model for. An epoch refers to one full cycle through the training dataset. |
Check the status of your customized model
After you start a fine-tune job, it can take some time to complete. Your job might be queued behind other jobs in the system. Training your model can take minutes or hours depending on the model and dataset size. The following example uses the REST API to check the status of your fine-tuning job. The example retrieves information about your job by using the job ID returned from the previous example:
curl -X GET $AZURE_OPENAI_ENDPOINT/openai/fine_tuning/jobs/<YOUR-JOB-ID>?api-version=2024-05-01-preview \
-H "api-key: $AZURE_OPENAI_API_KEY"
List fine-tuning events
To examine the individual fine-tuning events that were generated during training:
curl -X POST $AZURE_OPENAI_ENDPOINT/openai/fine_tuning/jobs/{fine_tuning_job_id}/events?api-version=2024-05-01-preview \
-H "Content-Type: application/json" \
-H "api-key: $AZURE_OPENAI_API_KEY"
Checkpoints
When each training epoch completes a checkpoint is generated. A checkpoint is a fully functional version of a model which can both be deployed and used as the target model for subsequent fine-tuning jobs. Checkpoints can be particularly useful, as they can provide a snapshot of your model prior to overfitting having occurred. When a fine-tuning job completes you will have the three most recent versions of the model available to deploy. The final epoch will be represented by your fine-tuned model, the previous two epochs will be available as checkpoints.
You can run the list checkpoints command to retrieve the list of checkpoints associated with an individual fine-tuning job:
curl -X POST $AZURE_OPENAI_ENDPOINT/openai/fine_tuning/jobs/{fine_tuning_job_id}/checkpoints?api-version=2024-05-01-preview \
-H "Content-Type: application/json" \
-H "api-key: $AZURE_OPENAI_API_KEY"
Safety evaluation GPT-4 fine-tuning - public preview
GPT-4 and GPT-4o-mini are our most advanced models that can be fine-tuned to your needs. As with Azure OpenAI models generally, the advanced capabilities of fine-tuned models come with increased responsible AI challenges related to harmful content, manipulation, human-like behavior, privacy issues, and more. Learn more about risks, capabilities, and limitations in the Overview of Responsible AI practices and Transparency Note. To help mitigate the risks associated with GPT-4 and GPT-4o-mini fine-tuned models, we have implemented additional evaluation steps to help detect and prevent harmful content in the training and outputs of fine-tuned models. These steps are grounded in the Microsoft Responsible AI Standard and Azure OpenAI Service content filtering.
- Evaluations are conducted in dedicated, customer specific, private workspaces;
- Evaluation endpoints are in the same geography as the Azure OpenAI resource;
- Training data is not stored in connection with performing evaluations; only the final model assessment (deployable or not deployable) is persisted; and
GPT-4 and GPT-4o-mini fine-tuned model evaluation filters are set to predefined thresholds and cannot be modified by customers; they aren't tied to any custom content filtering configuration you may have created.
Data evaluation
Before training starts, your data is evaluated for potentially harmful content (violence, sexual, hate, and fairness, self-harm – see category definitions here). If harmful content is detected above the specified severity level, your training job will fail, and you'll receive a message informing you of the categories of failure.
Sample message:
The provided training data failed RAI checks for harm types: [hate_fairness, self_harm, violence]. Please fix the data and try again.
Your training data is evaluated automatically within your data import job as part of providing the fine-tuning capability.
If the fine-tuning job fails due to the detection of harmful content in training data, you won't be charged.
Model evaluation
After training completes but before the fine-tuned model is available for deployment, the resulting model is evaluated for potentially harmful responses using Azure’s built-in risk and safety metrics. Using the same approach to testing that we use for the base large language models, our evaluation capability simulates a conversation with your fine-tuned model to assess the potential to output harmful content, again using specified harmful content categories (violence, sexual, hate, and fairness, self-harm).
If a model is found to generate output containing content detected as harmful at above an acceptable rate, you'll be informed that your model isn't available for deployment, with information about the specific categories of harm detected:
Sample Message:
This model is unable to be deployed. Model evaluation identified that this fine tuned model scores above acceptable thresholds for [Violence, Self Harm]. Please review your training data set and resubmit the job.
As with data evaluation, the model is evaluated automatically within your fine-tuning job as part of providing the fine-tuning capability. Only the resulting assessment (deployable or not deployable) is logged by the service. If deployment of the fine-tuned model fails due to the detection of harmful content in model outputs, you won't be charged for the training run.
Deploy a fine-tuned model
Important
After you deploy a customized model, if at any time the deployment remains inactive for greater than fifteen (15) days, the deployment is deleted. The deployment of a customized model is inactive if the model was deployed more than fifteen (15) days ago and no completions or chat completions calls were made to it during a continuous 15-day period.
The deletion of an inactive deployment doesn't delete or affect the underlying customized model, and the customized model can be redeployed at any time. As described in Azure OpenAI Service pricing, each customized (fine-tuned) model that's deployed incurs an hourly hosting cost regardless of whether completions or chat completions calls are being made to the model. To learn more about planning and managing costs with Azure OpenAI, refer to the guidance in Plan to manage costs for Azure OpenAI Service.
The following Python example shows how to use the REST API to create a model deployment for your customized model. The REST API generates a name for the deployment of your customized model.
variable | Definition |
---|---|
token | There are multiple ways to generate an authorization token. The easiest method for initial testing is to launch the Cloud Shell from the Azure portal. Then run az account get-access-token . You can use this token as your temporary authorization token for API testing. We recommend storing this in a new environment variable. |
subscription | The subscription ID for the associated Azure OpenAI resource. |
resource_group | The resource group name for your Azure OpenAI resource. |
resource_name | The Azure OpenAI resource name. |
model_deployment_name | The custom name for your new fine-tuned model deployment. This is the name that will be referenced in your code when making chat completion calls. |
fine_tuned_model | Retrieve this value from your fine-tuning job results in the previous step. It will look like gpt-35-turbo-0613.ft-b044a9d3cf9c4228b5d393567f693b83 . You'll need to add that value to the deploy_data json. Alternatively you can also deploy a checkpoint, by passing the checkpoint ID which will appear in the format ftchkpt-e559c011ecc04fc68eaa339d8227d02d |
curl -X POST "https://management.azure.com/subscriptions/<SUBSCRIPTION>/resourceGroups/<RESOURCE_GROUP>/providers/Microsoft.CognitiveServices/accounts/<RESOURCE_NAME>/deployments/<MODEL_DEPLOYMENT_NAME>api-version=2023-05-01" \
-H "Authorization: Bearer <TOKEN>" \
-H "Content-Type: application/json" \
-d '{
"sku": {"name": "standard", "capacity": 1},
"properties": {
"model": {
"format": "OpenAI",
"name": "<FINE_TUNED_MODEL>",
"version": "1"
}
}
}'
Cross region deployment
Fine-tuning supports deploying a fine-tuned model to a different region than where the model was originally fine-tuned. You can also deploy to a different subscription/region.
The only limitations are that the new region must also support fine-tuning and when deploying cross subscription the account generating the authorization token for the deployment must have access to both the source and destination subscriptions.
Below is an example of deploying a model that was fine-tuned in one subscription/region to another.
curl -X PUT "https://management.azure.com/subscriptions/<SUBSCRIPTION>/resourceGroups/<RESOURCE_GROUP>/providers/Microsoft.CognitiveServices/accounts/<RESOURCE_NAME>/deployments/<MODEL_DEPLOYMENT_NAME>api-version=2023-05-01" \
-H "Authorization: Bearer <TOKEN>" \
-H "Content-Type: application/json" \
-d '{
"sku": {"name": "standard", "capacity": 1},
"properties": {
"model": {
"format": "OpenAI",
"name": "<FINE_TUNED_MODEL>",
"version": "1",
"source": "/subscriptions/{sourceSubscriptionID}/resourceGroups/{sourceResourceGroupName}/providers/Microsoft.CognitiveServices/accounts/{sourceAccount}"
}
}
}'
To deploy between the same subscription, but different regions, you would just have subscription and resource groups be identical for both source and destination variables and only the source and destination resource names would need to be unique.
Deploy a model with Azure CLI
The following example shows how to use the Azure CLI to deploy your customized model. With the Azure CLI, you must specify a name for the deployment of your customized model. For more information about how to use the Azure CLI to deploy customized models, see az cognitiveservices account deployment
.
To run this Azure CLI command in a console window, you must replace the following <placeholders> with the corresponding values for your customized model:
Placeholder | Value |
---|---|
<YOUR_AZURE_SUBSCRIPTION> | The name or ID of your Azure subscription. |
<YOUR_RESOURCE_GROUP> | The name of your Azure resource group. |
<YOUR_RESOURCE_NAME> | The name of your Azure OpenAI resource. |
<YOUR_DEPLOYMENT_NAME> | The name you want to use for your model deployment. |
<YOUR_FINE_TUNED_MODEL_ID> | The name of your customized model. |
az cognitiveservices account deployment create
--resource-group <YOUR_RESOURCE_GROUP>
--name <YOUR_RESOURCE_NAME>
--deployment-name <YOUR_DEPLOYMENT_NAME>
--model-name <YOUR_FINE_TUNED_MODEL_ID>
--model-version "1"
--model-format OpenAI
--sku-capacity "1"
--sku-name "Standard"
Use a deployed customized model
After your custom model deploys, you can use it like any other deployed model. You can use the Playgrounds in Azure OpenAI Studio to experiment with your new deployment. You can continue to use the same parameters with your custom model, such as temperature
and max_tokens
, as you can with other deployed models. For fine-tuned babbage-002
and davinci-002
models you'll use the Completions playground and the Completions API. For fine-tuned gpt-35-turbo-0613
models you'll use the Chat playground and the Chat completion API.
curl $AZURE_OPENAI_ENDPOINT/openai/deployments/<deployment_name>/chat/completions?api-version=2023-05-15 \
-H "Content-Type: application/json" \
-H "api-key: $AZURE_OPENAI_API_KEY" \
-d '{"messages":[{"role": "system", "content": "You are a helpful assistant."},{"role": "user", "content": "Does Azure OpenAI support customer managed keys?"},{"role": "assistant", "content": "Yes, customer managed keys are supported by Azure OpenAI."},{"role": "user", "content": "Do other Azure AI services support this too?"}]}'
Analyze your customized model
Azure OpenAI attaches a result file named results.csv to each fine-tune job after it completes. You can use the result file to analyze the training and validation performance of your customized model. The file ID for the result file is listed for each customized model, and you can use the REST API to retrieve the file ID and download the result file for analysis.
The following Python example uses the REST API to retrieve the file ID of the first result file attached to the fine-tuning job for your customized model, and then downloads the file to your working directory for analysis.
curl -X GET "$AZURE_OPENAI_ENDPOINT/openai/fine_tuning/jobs/<JOB_ID>?api-version=2023-12-01-preview" \
-H "api-key: $AZURE_OPENAI_API_KEY")
curl -X GET "$AZURE_OPENAI_ENDPOINT/openai/files/<RESULT_FILE_ID>/content?api-version=2023-12-01-preview" \
-H "api-key: $AZURE_OPENAI_API_KEY" > <RESULT_FILENAME>
The result file is a CSV file that contains a header row and a row for each training step performed by the fine-tuning job. The result file contains the following columns:
Column name | Description |
---|---|
step |
The number of the training step. A training step represents a single pass, forward and backward, on a batch of training data. |
train_loss |
The loss for the training batch. |
train_mean_token_accuracy |
The percentage of tokens in the training batch correctly predicted by the model. For example, if the batch size is set to 3 and your data contains completions [[1, 2], [0, 5], [4, 2]] , this value is set to 0.83 (5 of 6) if the model predicted [[1, 1], [0, 5], [4, 2]] . |
valid_loss |
The loss for the validation batch. |
validation_mean_token_accuracy |
The percentage of tokens in the validation batch correctly predicted by the model. For example, if the batch size is set to 3 and your data contains completions [[1, 2], [0, 5], [4, 2]] , this value is set to 0.83 (5 of 6) if the model predicted [[1, 1], [0, 5], [4, 2]] . |
full_valid_loss |
The validation loss calculated at the end of each epoch. When training goes well, loss should decrease. |
full_valid_mean_token_accuracy |
The valid mean token accuracy calculated at the end of each epoch. When training is going well, token accuracy should increase. |
You can also view the data in your results.csv file as plots in Azure OpenAI Studio. Select the link for your trained model, and you will see three charts: loss, mean token accuracy, and token accuracy. If you provided validation data, both datasets will appear on the same plot.
Look for your loss to decrease over time, and your accuracy to increase. If you see a divergence between your training and validation data that may indicate that you are overfitting. Try training with fewer epochs, or a smaller learning rate multiplier.
Clean up your deployments, customized models, and training files
When you're done with your customized model, you can delete the deployment and model. You can also delete the training and validation files you uploaded to the service, if needed.
Delete your model deployment
You can use various methods to delete the deployment for your customized model:
Delete your customized model
Similarly, you can use various methods to delete your customized model:
Note
You can't delete a customized model if it has an existing deployment. You must first delete your model deployment before you can delete your customized model.
Delete your training files
You can optionally delete training and validation files that you uploaded for training, and result files generated during training, from your Azure OpenAI subscription. You can use the following methods to delete your training, validation, and result files:
Continuous fine-tuning
Once you have created a fine-tuned model, you might want to continue to refine the model over time through further fine-tuning. Continuous fine-tuning is the iterative process of selecting an already fine-tuned model as a base model and fine-tuning it further on new sets of training examples.
To perform fine-tuning on a model that you have previously fine-tuned, you would use the same process as described in create a customized model but instead of specifying the name of a generic base model you would specify your already fine-tuned model's ID. The fine-tuned model ID looks like gpt-35-turbo-0613.ft-5fd1918ee65d4cd38a5dcf6835066ed7
curl -X POST $AZURE_OPENAI_ENDPOINT/openai/fine_tuning/jobs?api-version=2023-12-01-preview \
-H "Content-Type: application/json" \
-H "api-key: $AZURE_OPENAI_API_KEY" \
-d '{
"model": "gpt-35-turbo-0613.ft-5fd1918ee65d4cd38a5dcf6835066ed7",
"training_file": "<TRAINING_FILE_ID>",
"validation_file": "<VALIDATION_FILE_ID>",
"suffix": "<additional text used to help identify fine-tuned models>"
}'
We also recommend including the suffix
parameter to make it easier to distinguish between different iterations of your fine-tuned model. suffix
takes a string, and is set to identify the fine-tuned model. The suffix can contain up to 40 characters (a-z, A-Z, 0-9,- and _) that will be added to your fine-tuned model name.
If you're unsure of the ID of your fine-tuned model this information can be found in the Models page of Azure OpenAI Studio, or you can generate a list of models for a given Azure OpenAI resource using the REST API.
Troubleshooting
How do I enable fine-tuning? Create a custom model is greyed out in Azure OpenAI Studio?
In order to successfully access fine-tuning, you need Cognitive Services OpenAI Contributor assigned. Even someone with high-level Service Administrator permissions would still need this account explicitly set in order to access fine-tuning. For more information, please review the role-based access control guidance.
Why did my upload fail?
If your file upload fails in Azure OpenAI Studio, you can view the error message under “data files” in Azure OpenAI Studio. Hover your mouse over where it says “error” (under the status column) and an explanation of the failure will be displayed.
My fine-tuned model does not seem to have improved
Missing system message: You need to provide a system message when you fine tune; you will want to provide that same system message when you use the fine-tuned model. If you provide a different system message, you may see different results than what you fine-tuned for.
Not enough data: while 10 is the minimum for the pipeline to run, you need hundreds to thousands of data points to teach the model a new skill. Too few data points risks overfitting and poor generalization. Your fine-tuned model may perform well on the training data, but poorly on other data because it has memorized the training examples instead of learning patterns. For best results, plan to prepare a data set with hundreds or thousands of data points.
Bad data: A poorly curated or unrepresentative dataset will produce a low-quality model. Your model may learn inaccurate or biased patterns from your dataset. For example, if you are training a chatbot for customer service, but only provide training data for one scenario (e.g. item returns) it will not know how to respond to other scenarios. Or, if your training data is bad (contains incorrect responses), your model will learn to provide incorrect results.
Next steps
- Explore the fine-tuning capabilities in the Azure OpenAI fine-tuning tutorial.
- Review fine-tuning model regional availability
- Learn more about Azure OpenAI quotas