Quick start: Reinforcement learning fine-tuning

This product is a Pre-GA offering, subject to the "Pre-GA Offerings Terms" in the General Service Terms section of the Service Specific Terms, the "Agentic AI Services" terms in the Service Specific Terms, and the Additional Terms for Generative AI Preview Products. Pre-GA products and features may have limited support, and changes to pre-GA products and features may not be compatible with other pre-GA versions. For more information, see the launch stage descriptions.

These Pre-GA products are in various stages of internal testing and review. As such, do not use proprietary, sensitive, or other confidential data with these products. These products are made available to Customers solely for limited testing and evaluation, and may not be used for commercial or production purposes.

This page walks you through the end-to-end workflow for reinforcement learning fine-tuning of Gemini models: creating a tuning job, checking its status, retrieving the tuned-model endpoint, and running inference against it.

Before you begin, see About reinforcement learning fine-tuning for an overview of the feature, supported models, and supported regions.

Create a reinforcement learning fine-tuning job

A reinforcement learning fine-tuning job is created by sending a POST request to the tuningJobs.create endpoint. For the full request schema and all configurable fields, see the Reinforcement learning fine-tuning job page.

The examples on this page use us-central1 as the tuning region. The resulting tuned model is served from the us multi-region endpoint. For the full list of supported tuning and serving regions, see the Supported models and regions section.

curl -X POST \
  -H "Authorization: Bearer $(gcloud auth application-default print-access-token)" \
  -H "Content-Type: application/json" \
  "https://us-central1-aiplatform.googleapis.com/v1beta1/projects/PROJECT_ID/locations/us-central1/tuningJobs" \
  -d \
  $'{
    "tunedModelDisplayName": "dai-image-test",
    "baseModel": "gemini-3.5-flash",
    "reinforcementTuningSpec": {
      "trainingDatasetUri": "gs://path/to/your/training_dataset.jsonl",
      "validationDatasetUri": "gs://path/to/your/eval_dataset.jsonl",
      "hyperParameters": {
        "epochCount":15,
        "learningRateMultiplier":1.0,
        "samplesPerPrompt":16,
        "adapterSize":"ADAPTER_SIZE_SIXTEEN",
        "maxOutputTokens":32768,
        "batchSize":32,
        "evaluateInterval":5,
        "checkpointInterval":5,
        "thinkingLevel":"HIGH"
      },
      "singleRewardConfig": {
        "rewardName": "your_reward_function_name",
        "parseResponseConfig": {"parseType":"IDENTITY"},
        "cloudRunRewardScorer": {
          "cloudRunUri":"https://your.cloud.run.uri"
        }
      }
    }
  }'

Replace the following:

• PROJECT_ID: Your Google Cloud project ID.

Note: Only the v1beta1 API version is supported for reinforcement learning fine-tuning.

Check the reinforcement learning fine-tuning job status

You can monitor the progress, performance, and quality of a running reinforcement learning fine-tuning job in the Google Cloud console on the Agent Platform > Model > Tuning page. Each tuning job has a dedicated monitoring view that surfaces the underlying tuning status with charts for training and evaluation rewards, generation length, and other tuning metrics. For the full list of emitted metrics and how to interpret them, see the Metrics and monitoring page.

Training time

Training time is affected by the following factors:

• Training and validation dataset size. For details, see the Tuning dataset page.
• Hyperparameters — including samplesPerPrompt, batch size, epoch count, and learning rate multiplier. For details, see the Hyperparameters page.

Depending on your setup, a Gemini reinforcement learning fine-tuning job can run for hours to days.

Get the tuned-model endpoint

After the tuning job reaches JOB_STATE_SUCCEEDED, retrieve the deployed tuned-model endpoint by issuing a GetTuningJob request and reading the tunedModel.endpoint field of the response.

curl -X GET \
  -H "Authorization: Bearer $(gcloud auth application-default print-access-token)" \
  "https://us-central1-aiplatform.googleapis.com/v1beta1/projects/PROJECT_ID/locations/us-central1/tuningJobs/TUNING_JOB_ID"

Replace the following:

• PROJECT_ID: Your Google Cloud project ID.
• TUNING_JOB_ID: The ID of the tuning job.

Example response (abbreviated):

{
  "name": "projects/{PROJECT_ID}/locations/us-central1/tuningJobs/{TUNING_JOB_ID}",
  "tunedModelDisplayName": "my-rl-tuned-model",
  "state": "JOB_STATE_SUCCEEDED",
  "tunedModel": {
    "model": "projects/{PROJECT_ID}/locations/us-central1/models/{MODEL_ID}",
    "endpoint": "projects/{PROJECT_ID}/locations/us/endpoints/{ENDPOINT_ID}"
  },
  "reinforcementTuningSpec": { ... }
}

Once the tuning job succeeds, endpoint from the last checkpoint is shown in the response.

Run inference on the tuned-model endpoint

The tuned model serves predictions through the standard generateContent API on the returned endpoint. Because the tuning job ran in us-central1, the tuned model is served from the us multi-region endpoint.

curl -X POST \
  -H "Authorization: Bearer $(gcloud auth application-default print-access-token)" \
  -H "Content-Type: application/json" \
  "https://aiplatform.us.rep.googleapis.com/v1beta1/projects/PROJECT_ID/locations/us/endpoints/ENDPOINT_ID:generateContent" \
  -d \
  $'{
    "contents": [
      {
        "role": "user",
        "parts": [
          { "text": "Why is the sky blue?" }
        ]
      }
    ]
  }'

Replace the following:

• PROJECT_ID: Your Google Cloud project ID.
• ENDPOINT_ID: The endpoint ID returned in the tunedModel.endpoint field of the GetTuningJob response.

What's next

• Learn more about reinforcement learning fine-tuning jobs.