tunings.tune

Create a new tuning job to fine-tune a model

The SDK’s tuning implementation is experimental, and may change in future versions.

Method

client.tunings.tune(
    base_model: str,
    training_dataset: TuningDataset,
    config: Optional[CreateTuningJobConfig] = None
) -> TuningJob

Creates a tuning job and returns the TuningJob object. This method initiates a fine-tuning process for a specified base model using the provided training dataset.

base_model · string · required

The name of the model to tune. For Vertex AI, this can also be a pre-tuned model resource name starting with projects/.

training_dataset · TuningDataset · required

The training dataset to use for tuning. Can be one of:

• gcs_uri: GCS bucket path (Vertex AI only)
• vertex_dataset_resource: Vertex AI Dataset resource (Vertex AI only)
• examples: Inline training examples (Gemini API only)

config · CreateTuningJobConfig

Configuration options for the tuning job:

Show Configuration Fields

tuned_model_display_name · string

Display name for the tuned model

description · string

Description of the tuning job (Vertex AI only)

epoch_count · int

Number of training epochs

learning_rate_multiplier · float

Multiplier for the learning rate

batch_size · int

Training batch size

learning_rate · float

Base learning rate for training

adapter_size · string

Size of the adapter for LoRA tuning (Vertex AI only)

validation_dataset · TuningValidationDataset

Validation dataset for evaluation during tuning (Vertex AI only)

evaluation_config · EvaluationConfig

Configuration for model evaluation (Vertex AI only)

Response

name · string

The resource name of the tuning job

state · JobState

Current state of the tuning job. One of:

• JOB_STATE_QUEUED: Job is queued
• JOB_STATE_PENDING: Job is pending
• JOB_STATE_RUNNING: Job is running
• JOB_STATE_SUCCEEDED: Job completed successfully
• JOB_STATE_FAILED: Job failed
• JOB_STATE_CANCELLED: Job was cancelled

create_time · string

Timestamp when the job was created

start_time · string

Timestamp when the job started running

end_time · string

Timestamp when the job completed

base_model · string

The base model being tuned

tuned_model · TunedModel

Information about the resulting tuned model

Usage

Basic Example

from google import genai
from google.genai import types

client = genai.Client(api_key='your-api-key')

# Create a tuning job
tuning_job = client.tunings.tune(
    base_model='gemini-1.5-flash-002',
    training_dataset=types.TuningDataset(
        gcs_uri='gs://my-bucket/training-data.jsonl'
    ),
    config=types.CreateTuningJobConfig(
        tuned_model_display_name='my-tuned-model',
        epoch_count=5,
        learning_rate_multiplier=1.0
    )
)

print(f"Tuning job created: {tuning_job.name}")
print(f"State: {tuning_job.state}")

Polling for Completion

import time

# Create tuning job
tuning_job = client.tunings.tune(
    base_model='gemini-1.5-flash-002',
    training_dataset=types.TuningDataset(
        gcs_uri='gs://my-bucket/training-data.jsonl'
    )
)

# Poll until job completes
while tuning_job.state in [
    types.JobState.JOB_STATE_QUEUED,
    types.JobState.JOB_STATE_PENDING,
    types.JobState.JOB_STATE_RUNNING
]:
    print(f"Job state: {tuning_job.state}")
    time.sleep(60)
    tuning_job = client.tunings.get(name=tuning_job.name)

if tuning_job.state == types.JobState.JOB_STATE_SUCCEEDED:
    print(f"Tuning completed! Model: {tuning_job.tuned_model.model}")
else:
    print(f"Tuning failed with state: {tuning_job.state}")

With Inline Examples (Gemini API)

from google.genai import types

tuning_job = client.tunings.tune(
    base_model='gemini-1.5-flash-002',
    training_dataset=types.TuningDataset(
        examples=[
            {
                'text_input': 'What is the capital of France?',
                'output': 'The capital of France is Paris.'
            },
            {
                'text_input': 'What is 2+2?',
                'output': '2+2 equals 4.'
            }
        ]
    ),
    config=types.CreateTuningJobConfig(
        epoch_count=10,
        batch_size=4
    )
)

See Also

• tunings.get - Retrieve tuning job status
• tunings.list - List all tuning jobs
• tunings.cancel - Cancel a running tuning job