gcloud alpha ai tuning-jobs create

gcloud alpha ai tuning-jobs create - create a supervised fine-tuning job

gcloud alpha ai tuning-jobs create --source-model=SOURCE_MODEL --training-dataset-uri=TRAINING_DATASET_URI [--adapter-size=ADAPTER_SIZE] [--batch-size=BATCH_SIZE] [--custom-base-model=CUSTOM_BASE_MODEL] [--description=DESCRIPTION] [--epoch-count=EPOCH_COUNT] [--export-last-checkpoint-only] [--labels=[KEY=VALUE,…]] [--output-uri=OUTPUT_URI] [--region=REGION] [--service-account=SERVICE_ACCOUNT] [--tuned-model-display-name=TUNED_MODEL_DISPLAY_NAME] [--tuning-mode=TUNING_MODE] [--validation-dataset-uri=VALIDATION_DATASET_URI] [--kms-key=KMS_KEY : --kms-keyring=KMS_KEYRING --kms-location=KMS_LOCATION --kms-project=KMS_PROJECT] [--learning-rate=LEARNING_RATE     | --learning-rate-multiplier=LEARNING_RATE_MULTIPLIER] [GCLOUD_WIDE_FLAG …]

--source-model=SOURCE_MODEL

The base model to tune, e.g. ``gemini-1.0-pro-002`` or ``meta/llama3_1@llama-3.1-8b``. To start tuning from a custom checkpoint or a previously tuned open model, also pass ``--custom-base-model``.

--training-dataset-uri=TRAINING_DATASET_URI

Cloud Storage URI of the training dataset. The dataset must be formatted as a JSONL file.

--adapter-size=ADAPTER_SIZE

Adapter size for parameter-efficient fine-tuning. This is only applicable when using a PEFT-compatible model. ADAPTER_SIZE must be one of: 1, 2, 4, 8, 16, 32.

--batch-size=BATCH_SIZE

Batch size for tuning. Only applicable to open source models.

--custom-base-model=CUSTOM_BASE_MODEL

Cloud Storage path to custom model weights. Set this to tune from a custom checkpoint or to continue tuning a previously tuned model. Must be a Cloud Storage directory containing model weights in .safetensors format. ``--source-model`` must still be set to indicate the base model architecture. Only available for open source models.

--description=DESCRIPTION

Description of the tuning job.

--epoch-count=EPOCH_COUNT

Number of training epochs. If not set, a default value will be calculated based on the training dataset size.

--export-last-checkpoint-only

If set, disable intermediate checkpoints for the tuning job and only export the last checkpoint. Default is to enable intermediate checkpoints.

--labels=[KEY=VALUE,…]

List of label KEY=VALUE pairs to add.

Keys must start with a lowercase character and contain only hyphens (-), underscores (_), lowercase characters, and numbers. Values must contain only hyphens (-), underscores (_), lowercase characters, and numbers.

--output-uri=OUTPUT_URI

Cloud Storage path to the directory where tuning job outputs are written. Required for open source models.

Region resource - Cloud region to create a tuning job. This represents a Cloud resource. (NOTE) Some attributes are not given arguments in this group but can be set in other ways.

To set the project attribute:

• provide the argument --region on the command line with a fully specified name;
• set the property ai/region with a fully specified name;
• choose one from the prompted list of available regions with a fully specified name;
• provide the argument --project on the command line;
• set the property core/project.

--region=REGION

ID of the region or fully qualified identifier for the region.

To set the region attribute:

• provide the argument --region on the command line;
• set the property ai/region;
• choose one from the prompted list of available regions.

--service-account=SERVICE_ACCOUNT

The service account that the tuning job runs as. If not specified, the Vertex AI Custom Code Service Agent is used.

--tuned-model-display-name=TUNED_MODEL_DISPLAY_NAME

Display name of the tuned model.

--tuning-mode=TUNING_MODE

Tuning mode. ``FULL`` performs full fine-tuning. ``PEFT_ADAPTER`` performs parameter-efficient fine-tuning (PEFT). Only applicable to open source models. TUNING_MODE must be one of: FULL, PEFT_ADAPTER.

--validation-dataset-uri=VALIDATION_DATASET_URI

Cloud Storage URI of the optional validation dataset. The dataset must be formatted as a JSONL file.

Key resource - The Cloud KMS (Key Management Service) cryptokey that will be used to protect the tuning job. The 'Vertex AI Service Agent' service account must hold permission 'Cloud KMS CryptoKey Encrypter/Decrypter'. The arguments in this group can be used to specify the attributes of this resource.

--kms-key=KMS_KEY

ID of the key or fully qualified identifier for the key.

To set the kms-key attribute:

• provide the argument --kms-key on the command line.

This flag argument must be specified if any of the other arguments in this group are specified.

--kms-keyring=KMS_KEYRING

The KMS keyring of the key.

To set the kms-keyring attribute:

• provide the argument --kms-key on the command line with a fully specified name;
• provide the argument --kms-keyring on the command line.

--kms-location=KMS_LOCATION

The Google Cloud location for the key.

To set the kms-location attribute:

• provide the argument --kms-key on the command line with a fully specified name;
• provide the argument --kms-location on the command line.

--kms-project=KMS_PROJECT

The Google Cloud project for the key.

To set the kms-project attribute:

• provide the argument --kms-key on the command line with a fully specified name;
• provide the argument --kms-project on the command line;
• set the property core/project.

Learning rate configuration for tuning. At most one of --learning-rate-multiplier or --learning-rate can be set.

At most one of these can be specified:

--learning-rate=LEARNING_RATE

Learning rate for tuning. Only applicable to open source models. Mutually exclusive with --learning-rate-multiplier.

--learning-rate-multiplier=LEARNING_RATE_MULTIPLIER

Multiplier for adjusting the default learning rate. Only applicable to Gemini models. Mutually exclusive with --learning-rate. If neither flag is set, a default value will be calculated based on the training dataset size.

Run $ gcloud help for details.