Tuning

Fine-tune models with Vertex AI, manage tuning jobs, and use tuned models

Model tuning is only supported in Vertex AI.

Fine-tuning allows you to customize Gemini models on your own data for improved performance on specific tasks.

Create a tuning job

Start a supervised fine-tuning job:

from google.genai import types

model = 'gemini-2.5-flash'
training_dataset = types.TuningDataset(
    # or gcs_uri=my_vertex_multimodal_dataset
    gcs_uri='gs://your-gcs-bucket/your-tuning-data.jsonl',
)

tuning_job = client.tunings.tune(
    base_model=model,
    training_dataset=training_dataset,
    config=types.CreateTuningJobConfig(
        epoch_count=1, 
        tuned_model_display_name='test_dataset_examples model'
    ),
)
print(tuning_job)

Training data format

Vertex AI supports two training data sources:

GCS JSONL file

Provide a path to a JSONL file in Google Cloud Storage:

training_dataset = types.TuningDataset(
    gcs_uri='gs://your-gcs-bucket/training-data.jsonl',
)

Each line should be a JSON object with the training example:

{"contents": [{"role": "user", "parts": [{"text": "What is AI?"}]}, {"role": "model", "parts": [{"text": "AI is..."}]}]}
{"contents": [{"role": "user", "parts": [{"text": "Explain ML"}]}, {"role": "model", "parts": [{"text": "ML is..."}]}]}

Vertex AI Multimodal Dataset

Use a Vertex AI Multimodal Dataset:

training_dataset = types.TuningDataset(
    gcs_uri='projects/PROJECT_ID/locations/LOCATION/datasets/DATASET_ID',
)

Get tuning job status

Retrieve the status of a tuning job:

tuning_job = client.tunings.get(name=tuning_job.name)
print(tuning_job)
print(tuning_job.state)

Poll for completion

Wait for a tuning job to complete:

import time

completed_states = set([
    'JOB_STATE_SUCCEEDED',
    'JOB_STATE_FAILED',
    'JOB_STATE_CANCELLED',
])

while tuning_job.state not in completed_states:
    print(tuning_job.state)
    tuning_job = client.tunings.get(name=tuning_job.name)
    time.sleep(10)

if tuning_job.state == 'JOB_STATE_SUCCEEDED':
    print("Tuning completed successfully!")
else:
    print(f"Tuning ended with state: {tuning_job.state}")

Use tuned models

Once tuning is complete, use the tuned model endpoint:

response = client.models.generate_content(
    model=tuning_job.tuned_model.endpoint,
    contents='why is the sky blue?',
)

print(response.text)

Get tuned model details

Retrieve information about a tuned model:

tuned_model = client.models.get(model=tuning_job.tuned_model.model)
print(tuned_model)

List tuned models

List all your tuned models:

for model in client.models.list(config={'page_size': 10, 'query_base': False}):
    print(model)

Pagination

Navigate through pages of tuned models:

pager = client.models.list(config={'page_size': 10, 'query_base': False})
print(pager.page_size)
print(pager[0])
pager.next_page()
print(pager[0])

Async listing

async for job in await client.aio.models.list(
    config={'page_size': 10, 'query_base': False}
):
    print(job)

With pagination:

async_pager = await client.aio.models.list(
    config={'page_size': 10, 'query_base': False}
)
print(async_pager.page_size)
print(async_pager[0])
await async_pager.next_page()
print(async_pager[0])

Update tuned models

Update display name and description:

from google.genai import types

model = pager[0]

model = client.models.update(
    model=model.name,
    config=types.UpdateModelConfig(
        display_name='my tuned model', 
        description='my tuned model description'
    ),
)

print(model)

List tuning jobs

List all tuning jobs:

for job in client.tunings.list(config={'page_size': 10}):
    print(job)

Tuning configuration

The CreateTuningJobConfig supports:

• epoch_count - Number of training epochs
• tuned_model_display_name - Display name for the tuned model
• learning_rate - Learning rate for training
• batch_size - Training batch size

Best practices

• Use at least 100-500 high-quality training examples
• Format your data consistently
• Monitor the tuning job state regularly
• Test your tuned model before deploying to production
• Keep your training data in GCS for easy access
• Use validation data to prevent overfitting