Tune Gemini models by using preference tuning

This document describes how to perform preference tuning on a Gemini model.

Before you begin

Before you begin, you must prepare a dataset by following the instructions in Prepare preference tuning data for Gemini models.

Create a tuning job

Console

To configure preference tuning by using the Google Cloud console, perform the following steps:

  1. In the Vertex AI section of the Google Cloud console, go to the Vertex AI Studio page.

    Go to Vertex AI Studio

  2. Click Create tuned model.

  3. Select Preference tuning.

  4. Under Model details, configure the following:

    1. In the Tuned model name field, enter a name for your new tuned model, up to 128 characters.
    2. In the Base model field, select gemini-2.5-flash.
    3. In the Region drop-down field, Select the region where the pipeline tuning job runs and where the tuned model is deployed.

  5. Under Tuning setting, configure the following:

    1. In the Number of epochs field, enter the number of steps to run for model tuning.
    2. In the Adapter Size field, enter the adapter size to use for model tuning.
    3. Optional: In the Beta field, enter the desired beta value. Beta is the coefficient that controls how closely the tuned model aligns with its baseline.
    4. In the Learning rate multiplier field, enter the step size at each iteration. The default value is 1. .

  6. Optional: To disable intermediate checkpoints and use only the latest checkpoint, click the Export last checkpoint only toggle.

  7. Click Continue.

    The Tuning dataset page opens.

  8. To upload a dataset file, select one of the following:

    1. If you haven't uploaded a dataset yet, select the radio button for Upload file to Cloud Storage.
    2. In the Select JSONL file field, click Browse and select your dataset file.
    3. In the Dataset location field, click Browse and select the Cloud Storage bucket where you want to store your dataset file.
    4. If your dataset file is already in a Cloud Storage bucket, select the radio button for Existing file on Cloud Storage.
    5. In Cloud Storage file path field, click Browse and select the Cloud Storage bucket where your dataset file is located.

  9. (Optional) To get validation metrics during training, click the Enable model validation toggle.

    1. In the Validation dataset file, enter the Cloud Storage path of your validation dataset.

  10. Click Start Tuning.

    Your new model appears under the Managed tuning section on the Tune and Distill page. When the model is finished tuning, the Status says Succeeded.

REST

To create a model tuning job, send a POST request by using the tuningJobs.create method. Some of the parameters are not supported by all of the models. Ensure that you include only the applicable parameters for the model that you're tuning.

Before using any of the request data, make the following replacements:

  • PROJECT_ID: Your project ID.
  • TUNING_JOB_REGION: The region where the tuning job runs. This is also the default region for where the tuned model is uploaded.
  • BASE_MODEL: Name of the foundation model to tune.
  • TRAINING_DATASET_URI: Cloud Storage URI of your training dataset. The dataset must be formatted as a JSONL file. For best results, provide at least 100 to 500 examples. For more information, see About supervised tuning datasets .
  • VALIDATION_DATASET_URIOptional: The Cloud Storage URI of your validation dataset file.
  • EPOCH_COUNTOptional: The number of complete passes the model makes over the entire training dataset during training. Leave it unset to use the pre-populated recommended value.
  • BETAOptional: The coefficient that controls how closely the tuned model aligns with its baseline.
  • ADAPTER_SIZEOptional: The Adapter size to use for the tuning job. The adapter size influences the number of trainable parameters for the tuning job. A larger adapter size implies that the model can learn more complex tasks, but it requires a larger training dataset and longer training times.
  • LEARNING_RATE_MULTIPLIER: Optional: A multiplier to apply to the recommended learning rate. Leave it unset to use the recommended value.
  • EXPORT_LAST_CHECKPOINT_ONLYOptional: Set to true to use only the latest checkpoint.
  • TUNED_MODEL_DISPLAYNAMEOptional: A display name for the tuned model. If not set, a random name is generated.
  • KMS_KEY_NAMEOptional: The Cloud KMS resource identifier of the customer-managed encryption key used to protect a resource. The key has the format: projects/my-project/locations/my-region/keyRings/my-kr/cryptoKeys/my-key. The key needs to be in the same region as where the compute resource is created. For more information, see Customer-managed encryption keys (CMEK).
  • SERVICE_ACCOUNTOptional: The service account that the tuningJob workload runs as. If not specified, the Vertex AI Secure Fine-Tuning Service Agent in the project is used. See Tuning Service Agent. If you plan to use a customer-managed Service Account, you must grant the roles/aiplatform.tuningServiceAgent role to the service account. Also grant the Tuning Service Agent roles/iam.serviceAccountTokenCreator role to the customer-managed Service Account.

HTTP method and URL: 

POST https://TUNING_JOB_REGION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/TUNING_JOB_REGION/tuningJobs

Request JSON body: 

{
  "baseModel": "BASE_MODEL",
  "preferenceOptimizationSpec" : {
      "trainingDatasetUri": "TRAINING_DATASET_URI",
      "validationDatasetUri": "VALIDATION_DATASET_URI",
      "hyperParameters": {
          "epochCount": "EPOCH_COUNT",
          "beta": "BETA",
          "adapterSize": "ADAPTER_SIZE",
          "learningRateMultiplier": "LEARNING_RATE_MULTIPLIER"
      },
      "exportLastCheckpointOnly": EXPORT_LAST_CHECKPOINT_ONLY,
  },
  "tunedModelDisplayName": "TUNED_MODEL_DISPLAYNAME",
  "encryptionSpec": {
    "kmsKeyName": "KMS_KEY_NAME"
  },
  "serviceAccount": "SERVICE_ACCOUNT"
}

To send your request, expand one of these options:

Example curl command

PROJECT_ID=myproject
LOCATION=global
curl \
-X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json; charset=utf-8" \
"https://${LOCATION}-aiplatform.googleapis.com/v1/projects/${PROJECT_ID}/locations/${LOCATION}/tuningJobs" \
-d \
$'{
   "baseModel": "gemini-2.5-flash",
   "preferenceOptimizationSpec" : {
      "trainingDatasetUri": "gs://mybucket/preference_tuning/data/train_data.jsonl",
      "validationDatasetUri": "gs://mybucket/preference_tuning/data/validation_data.jsonl"
   },
   "tunedModelDisplayName": "tuned_gemini"
}'

Tuning hyperparameters

It's recommended to submit your first tuning job without changing the hyperparameters. The default value is the recommended value based on our benchmarking results to yield the best model output quality.

  • Epochs: The number of complete passes the model makes over the entire training dataset during training. Vertex AI automatically adjusts the default value to your training dataset size. This value is based on benchmarking results to optimize model output quality.
  • Beta: The coefficient that controls how closely the tuned model aligns with its baseline. Lower beta means more aggressive updates toward preferred responses. The recommended range is from 0.01 to 0.5, inclusive. A beta value of 0 stops the model from learning.
  • Adapter size: The Adapter size to use for the tuning job. The adapter size influences the number of trainable parameters for the tuning job. A larger adapter size implies that the model can learn more complex tasks, but it requires a larger training dataset and longer training times.
  • Learning Rate Multiplier: A multiplier to apply to the recommended learning rate. You can increase the value to converge faster, or decrease the value to avoid overfitting.

View or cancel a tuning job

To list tuning jobs, view tuning job details, cancel a tuning job, or delete a tuned model, see the following sections in the supervised fine-tuning documentation:

Model tuning metrics

The model tuning job automatically collects the following tuning metrics:

  • /preference_optimization_train_loss: Preference loss for the tuning dataset at a training step.

Model validation metrics

When validation data is provided, the model tuning job automatically collects the following validation metrics:

  • /eval_total_loss: Preference loss for the validation dataset at a validation step.

The metrics visualizations are available after the tuning job starts running. They are updated in real time as tuning progresses. If you don't specify a validation dataset when you create the tuning job, only the visualizations for the tuning metrics are available.

What's next