Fine-tune Nova 1.0

Prerequisites

Before you start a training job, note the following.

Data preparation

Preparing high-quality, properly formatted data is a critical first step in the fine-tuning process for large language models. Whether you're using full-rank supervised fine-tuning (SFT) or low-rank adaptation (LoRA), your data must follow specific format requirements. These requirements ensure successful model training. This section outlines the necessary data formats, validation methods, and best practices to help you prepare your datasets effectively for fine-tuning Amazon Nova models.

Full-rank SFT and LoRA PEFT configurations

This section covers guidance on recipe configurations for both full-rank supervised fine-tuning (SFT) and low-rank adaptation parameter-efficient fine-tuning (LoRA PEFT) approaches. These recipe files serve as the blueprint for your model customization jobs, allowing you to specify training parameters, hyperparameters, and other critical settings that determine how your model learns from your data. To adjust the hyperparameters, follow the guidelines in Selecting hyperparameters.

Fine-tuning configurations (Full-rank SFT and LoRA PEFT)

The only difference between full-rank SFT and LoRA PEFT in terms of recipe is the LoRA PEFT configuration, which is set to 'null' for full rank and set to appropriate values if using LoRA PEFT-based fine-tuning. Example recipes are available in the SageMaker HyperPod recipes GitHub repository. The following tables show detailed configurations that you might find helpful.

About "run" configuration.

About “training_config” configuration.

Running customized Nova model on SageMaker Training Jobs

This section demonstrates how to run a customized Nova model on SageMaker Training Jobs through a Jupyter notebook environment. You'll find a complete example that walks through the process of configuring and launching a training job, along with reference tables for selecting the appropriate container image URIs and instance configurations. This approach gives you programmatic control over your fine-tuning workflows while leveraging SageMaker AI's managed infrastructure for model customization. For more information, see Use a SageMaker AI estimator to run a training job.

Reference tables

Before running the sample notebook, refer to the following tables for selecting the appropriate container image URIs and instance configurations.

Selecting image URI

Selecting instance type and count

Sample notebook

The following sample notebook demonstrates how to run a training job. For additional getting started notebooks on how to customize Nova models using SageMaker Training Jobs, see Use a SageMaker AI estimator to run a training job.

# 1. Install dependencies

!pip install sagemaker==2.254.1

# 2. Import dependencies and initialize sagemaker session

import sagemaker,boto3

sm = boto3.client('sagemaker', region_name='us-east-1')
sagemaker_session = sagemaker.session.Session(boto_session=boto3.session.Session(), sagemaker_client=sm)

# 3. Configure your job
# Define the core configuration for launching a SageMaker Training Job. This includes input/output S3 URIs, container image, hardware setup, and other runtime parameters. Update the placeholders below before submitting the job.

job_name = "<Your Job Name>"

input_s3_uri = "<S3 path to input data>"
validation_s3_uri = "<S3 path to validation data>" # optional, leave blank if no validation data

output_s3_uri = "<S3 path to output location>"

image_uri = "<Image URI from documentation>" # use the SFT image
instance_type = "ml.p5.48xlarge" # do not change
instance_count = <Integer number of hosts> # change hosts as needed. Refer to documentation for allowed values based on model type.
role_arn = "<IAM Role you want to use to run the job>"
recipe_path = "<Local path to the recipe file>"
output_kms_key = "<KMS key arn to encrypt trained model in Amazon-owned S3 bucket>" # optional, leave blank for Amazon managed encryption

# 4. Launch SageMaker Training Job
# This block sets up and runs the SageMaker training job using the PyTorch estimator. It configures the training image, hardware, input channels, and TensorBoard integration. Validation data is included if provided.

from sagemaker.debugger import TensorBoardOutputConfig
from sagemaker.pytorch import PyTorch
from sagemaker.inputs import TrainingInput

tensorboard_output_config = TensorBoardOutputConfig(
    s3_output_path=output_s3_uri,
)

estimator = PyTorch(
    output_path=output_s3_uri,
    base_job_name=job_name,
    role=role_arn,
    instance_count=instance_count,
    instance_type=instance_type,
    training_recipe=recipe_path,
    sagemaker_session=sagemaker_session,
    image_uri=image_uri,
    tensorboard_output_config=tensorboard_output_config, # Add the setting for using TensorBoard.
    disable_profiler=True,
    debugger_hook_config=False,
    output_kms_key=output_kms_key
)

trainingInput = TrainingInput(
    s3_data=input_s3_uri,
    distribution='FullyReplicated',
    s3_data_type='Converse'
)

if (validation_s3_uri):
    validationInput = TrainingInput(
        s3_data=validation_s3_uri,
        distribution='FullyReplicated',
        s3_data_type='Converse'
    )

    estimator.fit(inputs={"train": trainingInput, "validation": validationInput}) # inputs must be called "train" and "validation", do not change
else:
    estimator.fit(inputs={"train": trainingInput})

Hyperparameter optimization guidance

Fine-tuning your Nova LLM model effectively requires careful selection of hyperparameters. While this section explains the basic recipe structure and components, optimizing hyperparameters for your specific use case often requires additional guidance. For comprehensive recommendations on hyperparameter selection, best practices, and optimization strategies, see Selecting hyperparameters. This resource provides detailed guidance on selecting appropriate learning rates, batch sizes, training epochs, and other critical parameters based on your dataset characteristics and training objectives. We recommend consulting this guide when fine-tuning your recipe configuration to achieve optimal model performance.

For details about minimum, maximum, and default values for epochs, learning rate, and learning warmup steps, see Hyperparameters for Understanding models.

Common recipe modifications

Here are some common recipe adjustments based on specific use cases: