Training job submission

Launching Training Jobs

After your agent is deployed and your dataset is in S3, create a training job using one of the following methods.

SageMaker AI Studio

SageMaker AI Python SDK

Discover supported models

from sagemaker.train.multi_turn_rl_trainer import MultiTurnRLTrainer

supported_models = MultiTurnRLTrainer.list_supported_models()
print(f"Supported MTRL models ({len(supported_models)}):")
for m in supported_models:
    print(f"  - {m}")

Set up your agent environment

Option 1: Bedrock AgentCore runtime

# List available runtimes
runtimes = MultiTurnRLTrainer.list_bedrock_agentcore_runtimes()
for rt in runtimes:
    print(f"  - {rt['name']} ({rt['status']}) → {rt['arn']}")

Option 2: Custom Lambda agent

from sagemaker.train.agent_lambda import AgentLambda

# Create from inline code
adapter = AgentLambda.create(
    source='''
import json
def handler(event, context):
    prompt = event.get("prompt", "")
    return {"statusCode": 200, "body": json.dumps({"status": "ok", "agentResponse": prompt})}
''',
    role="arn:aws:iam::123456789012:role/AgentLambdaRole",
)

# Create from a local file
adapter = AgentLambda.create(
    source="~/my_agent_handler.py",
    role="arn:aws:iam::123456789012:role/AgentLambdaRole",
)

# Create from S3
adapter = AgentLambda.create(
    source="s3://my-bucket/agent_handler.py",
    role="arn:aws:iam::123456789012:role/AgentLambdaRole",
)

# Wrap an existing Lambda
adapter = AgentLambda.get("arn:aws:lambda:us-west-2:123456789012:function:my-agent")

Register your dataset (optional)

from sagemaker.ai_registry.dataset import DataSet

dataset = DataSet.create(
    name="my-mtrl-dataset",
    source="s3://my-bucket/prompts/training_prompts.parquet"
)
print(f"Dataset ARN: {dataset.arn}")

Create Restricted Model Package Group for Nova (optional)

If you are choosing Nova model (nova-textgeneration-lite-v2), then optionally create Restricted Model Package Group prior to submitting a training job (next step). If you skip this step, the SDK automatically creates one for you.

Restricted Model Package Group (RMPG) is a Model Package Group with ManagedStorageType: Restricted. It's required for closed-source models like Nova where the model weights are managed by AWS and not directly accessible to the customer.

The RFT job schema requires two separate Restricted MPGs:

from sagemaker.core.resources import Job, ModelPackageGroup
from sagemaker.core.shapes import ManagedConfiguration

model_name = "nova-textgeneration-lite-v2"

# Restricted configuration
managed_config = ManagedConfiguration(managed_storage_type="Restricted")

# Output Model package group
output_mpg_name = f"{model_name}-mtrl-output-mpg"
create_kwargs = {
    "model_package_group_name": output_mpg_name,
    "region": "us-east-1",
    "managed_configuration": managed_config
}
output_mpg = ModelPackageGroup.create(**create_kwargs)

# Intermediate Model package group
intermediate_mpg_name = f"{model_name}-mtrl-inter-mpg"
create_kwargs = {
    "model_package_group_name": intermediate_mpg_name,
    "region": "us-east-1",
    "managed_configuration": managed_config
}
intermediate_mpg = ModelPackageGroup.create(**create_kwargs)

Once the Model package group is created, pass the groups in the next step when submitting a training job.

Submit a training job with Bedrock AgentCore

from sagemaker.train.multi_turn_rl_trainer import MultiTurnRLTrainer

trainer = MultiTurnRLTrainer(
    model="openai-reasoning-gpt-oss-20b",
    agent_env="arn:aws:bedrock-agentcore:us-west-2:123456789012:runtime/my-agent-runtime",
    training_dataset="s3://my-bucket/prompts/prompts.parquet",
    mlflow_app_arn="arn:aws:sagemaker:us-west-2:123456789012:mlflow-app/mlflow-app-id",
    s3_output_path="s3://my-bucket/output/",
    role="arn:aws:iam::123456789012:role/SageMakerRole",
    accept_eula=True,
)

# View and adjust hyperparameters
trainer.hyperparameters.get_info()
trainer.hyperparameters.max_epochs = 1
trainer.hyperparameters.global_batch_size = 32
trainer.hyperparameters.max_steps = 12

job = trainer.train(wait=True)
print(f"Job: {job.job_name}")
print(f"Status: {job.job_status}")
print(f"Output Model Package: {job.output_model_package_arn}")

Submit a training job with a custom Lambda agent

trainer = MultiTurnRLTrainer(
    model="openai-reasoning-gpt-oss-20b",
    agent_env=adapter,  # AgentLambda object or Lambda ARN string
    training_dataset="s3://my-bucket/prompts/prompts.parquet",
    mlflow_app_arn="arn:aws:sagemaker:us-west-2:123456789012:mlflow-app/mlflow-app-id",
    s3_output_path="s3://my-bucket/output/",
    role="arn:aws:iam::123456789012:role/SageMakerRole",
    accept_eula=True,
)

trainer.hyperparameters.max_epochs = 1
trainer.hyperparameters.global_batch_size = 32
trainer.hyperparameters.max_steps = 12

job = trainer.train(wait=True)
print(f"Job: {job.job_name}")
print(f"Status: {job.job_status}")
print(f"Output Model Package: {job.output_model_package_arn}")

Submit a training job with Restricted Model package group for Nova

Refer to the step above (Create Restricted Model Package Group for Nova) on how to create a Restricted Model Package group.

from sagemaker.train.multi_turn_rl_trainer import MultiTurnRLTrainer

trainer = MultiTurnRLTrainer(
    model="nova-textgeneration-lite-v2",
    agent_env="arn:aws:bedrock-agentcore:us-west-2:123456789012:runtime/my-agent-runtime",
    training_dataset="s3://my-bucket/prompts/prompts.parquet",
    mlflow_app_arn="arn:aws:sagemaker:us-west-2:123456789012:mlflow-app/mlflow-app-id",
    s3_output_path="s3://my-bucket/output/",
    role="arn:aws:iam::123456789012:role/SageMakerRole",
    accept_eula=True,
    output_model_package_group=output_mpg,
    intermediate_checkpoint_model_package_group=intermediate_mpg
)

# View and adjust hyperparameters
trainer.hyperparameters.get_info()
trainer.hyperparameters.max_epochs = 1
trainer.hyperparameters.global_batch_size = 32
trainer.hyperparameters.max_steps = 12

job = trainer.train(wait=True)
print(f"Job: {job.job_name}")
print(f"Status: {job.job_status}")
print(f"Output Model Package: {job.output_model_package_arn}")

AWS CLI

Create a training job using the CreateJob API. You specify the agent configuration, training data location, base model, and output settings in the JobConfigDocument.

To retrieve the full JobConfigDocument schema:

aws sagemaker list-job-schema-versions --job-category AgentRFT
aws sagemaker describe-job-schema-version --job-category AgentRFT --version "1.0.0"

Create job with Bedrock AgentCore

aws sagemaker create-job \
  --job-category AgentRFT \
  --job-name "my-agent-rft-job" \
  --role-arn "arn:aws:iam::123456789012:role/SageMakerFineTuningJobRole" \
  --job-config-schema-version "1.0.0" \
  --job-config-document '{
    "AgentConfig": {
      "BedrockAgentCoreConfig": {
        "AgentRuntimeArn": "arn:aws:bedrock-agentcore:us-west-2:123456789012:runtime/my-agent"
      }
    },
    "InputDataConfig": [...],
    "OutputDataConfig": {...},
    "ModelPackageConfig": {...},
    "TrainingConfig": {...}
  }' \
  --region us-west-2

Create job with custom Lambda agent

aws sagemaker create-job \
  --job-category AgentRFT \
  --job-name "my-custom-agent-rft-job" \
  --role-arn "arn:aws:iam::account-id:role/SageMakerFineTuningJobRole" \
  --job-config-schema-version "1.0.0" \
  --job-config-document '{
    "AgentConfig": {
      "CustomAgentLambdaConfig": {
        "LambdaArn": "arn:aws:lambda:us-west-2:account-id:function:rft-agent-forwarder"
      }
    },
    "InputDataConfig": [...],
    "OutputDataConfig": {...},
    "ModelPackageConfig": {...},
    "TrainingConfig": {...}
  }' \
  --region us-west-2

boto3

Create job with Bedrock AgentCore

import json
import boto3

sm = boto3.client("sagemaker")

response = sm.create_job(
    JobName="my-agent-rft-job",
    RoleArn="arn:aws:iam::123456789012:role/SageMakerFineTuningJobRole",
    JobCategory="AgentRFT",
    JobConfigSchemaVersion="1.0.0",
    JobConfigDocument=json.dumps({
        "AgentConfig": {
            "BedrockAgentCoreConfig": {
                "AgentRuntimeArn": "arn:aws:bedrock-agentcore:us-west-2:123456789012:runtime/my-agent"
            }
        },
        "InputDataConfig": [...],
        "OutputDataConfig": {...},
        "ModelPackageConfig": {...},
        "TrainingConfig": {...}
    })
)

print(f"Job ARN: {response['JobArn']}")

Create job with custom Lambda agent

import json
import boto3

sm = boto3.client("sagemaker")

response = sm.create_job(
    JobName="my-custom-agent-rft-job",
    RoleArn="arn:aws:iam::account-id:role/SageMakerFineTuningJobRole",
    JobCategory="AgentRFT",
    JobConfigSchemaVersion="1.0.0",
    JobConfigDocument=json.dumps({
        "AgentConfig": {
            "CustomAgentLambdaConfig": {
                "LambdaArn": "arn:aws:lambda:us-west-2:account-id:function:rft-agent-forwarder"
            }
        },
        "InputDataConfig": [...],
        "OutputDataConfig": {...},
        "ModelPackageConfig": {...},
        "TrainingConfig": {...}
    })
)

print(f"Job ARN: {response['JobArn']}")

Monitoring Training

Monitor your Training Job

Use the DescribeJob API to check your job's current status at any time. The job status transitions through InProgress, and then to Completed, Failed or Stopped.

aws sagemaker describe-job \
  --job-name "my-agent-rft-job" \
  --job-category AgentRFT \
  --region us-west-2

Use the SDK:

# Run without blocking
job = trainer.train(wait=False)
job.wait(poll=5, timeout=3000, max_log_lines=10)
# Check status
job.refresh()
print(f"Status: {job.job_status}")
print(f"Secondary Status: {job.secondary_status}")
print(f"Output Model Package: {job.output_model_package_arn}")
print(f"MLflow Details: {job.mlflow_details}")
print(f"Billable Tokens: {job.billable_token_usage}")
# Open MLflow tracking URL
job.get_mlflow_url()
# Stop a running job
job.stop()
# Attach to an existing job from a different session
existing_job = MultiTurnRLTrainer.attach(job_name="my-existing-job-name")
print(f"Status: {existing_job.job_status}")
print(f"Output Model: {existing_job.output_model_package_arn}")
# List all completed jobs
from sagemaker.train.agent_rft_job import AgentRFTJob
for j in AgentRFTJob.get_all(status_equals="Completed"):
    print(f"{j.job_name}: {j.job_status}")

Monitor Training in MLflow

SageMaker AI automatically integrates with managed MLflow to track your training job's progress, metrics, and artifacts. To enable MLflow tracking, include an MlflowConfig in your job's OutputDataConfig:

"OutputDataConfig": {
    "S3OutputPath": "s3://your-bucket/output/",
    "MlflowConfig": {
        "MlflowResourceArn": "arn:aws:sagemaker:us-west-2:123456789012:mlflow-app/my-rft-mlflow-app"
    }
}

Prerequisites

What gets logged

Detailed training metrics reference

The following metrics are logged at each training step.

Step counters and throughput (training/)

Phase durations (timing_s/)

Reward distribution (rollout/reward/)

Turn counts (rollout/turns/)

Token lengths (rollout/tokens/)

Log-probability health (rollout/logprob/)

Advantage distribution (rollout/advantage/)

Batch-quality classification (analysis/)

Evaluation reward and pass@k (val/reward/)

Emitted at baseline (step 0), at every val_every interval, and at the final step. Includes the same distribution metrics as rollout/reward plus group-reward metrics aggregated by prompt.

Distribution:

Group-reward (per-prompt aggregation):

Pass@k and success accounting:

Evaluation turn counts (val/turns/)

Evaluation token lengths (val/tokens/)

Evaluation log-probability health (val/logprob/)

Accessing the MLflow UI

Access the MLflow UI via a presigned URL:

aws sagemaker create-presigned-mlflow-app-url \
  --arn arn:aws:sagemaker:us-west-2:123456789012:mlflow-app/mlflow-app-id \
  --region us-west-2

Copy the AuthorizedUrl from the output into your browser.

Agent trajectories and traces

During training, SageMaker AI records every interaction between your agent and the policy model as a trajectory — the complete record of one rollout. Each trajectory captures every prompt sent to the model, every response generated, every tool call made, and the final reward. Trajectories are published to your MLflow experiment as structured traces.

Trace contents

Viewing trajectories in the MLflow UI

Access the MLflow UI via a presigned URL:

aws sagemaker create-presigned-mlflow-app-url \
  --arn arn:aws:sagemaker:us-west-2:123456789012:mlflow-app/mlflow-app-id \
  --region us-west-2

Copy the AuthorizedUrl from the output into your browser.

Open the MLflow UI using the presigned URL above. Navigate to your experiment's run and select the Traces tab. Each trace represents one completed rollout and shows:

Use trajectories to debug low reward scores

Get Training Results

When a training job completes, your trained model weights are stored as a SageMaker AI Model Package. This section explains how to find your results, understand the checkpoint types produced during training, and use them for deployment or continued training.

How results are stored

SageMaker AI stores training output as versioned, immutable Model Packages inside Model Package Groups. Multi-turn RL uses two separate groups, which you specify when creating a job:

Configure both groups in your ModelPackageConfig:

"ModelPackageConfig": {
  "OutputModelPackageGroupArn": "arn:aws:sagemaker:us-west-2:123456789012:model-package-group/my-final-models",
  "IntermediateCheckpointModelPackageGroupArn": "arn:aws:sagemaker:us-west-2:123456789012:model-package-group/my-intermediate-checkpoints"
}

Checkpoint types

Training produces two types of checkpoints, saved at every training step:

Model checkpoint (weights only)

Resumable checkpoint (full state)

Checkpoint lifecycle

Step 1 → Intermediate Checkpoint (resumable)
Step 1 → Intermediate Checkpoint (HF-compatible)
...
Step N-1 → Intermediate Checkpoint (resumable)
Step N-1 → Intermediate Checkpoint (HF-compatible)
...
Step N (final) → Model Checkpoint (HuggingFace LoRA) → Output Model Package Group

Retrieve your trained model

When a job completes successfully, the final model is saved as a Model Package in the Output Model Package Group. The OutputModelPackageArn field on the job record contains the ARN.

Check job completion and retrieve the output model ARN:

aws sagemaker describe-job \
--job-name "my-agent-rft-job" \
--job-category AgentRFT \
--region us-west-2

Look for OutputModelPackageArn in the response. Use it to describe the Model Package and get the S3 location of the weights:

aws sagemaker describe-model-package \
--model-package-name "arn:aws:sagemaker:us-west-2:123456789012:model-package/my-final-models/5"

If a job fails or is stopped before completion, the last intermediate checkpoint is promoted to the Output Model Package Group on a best-effort basis. Check OutputModelPackageArn the same way.

To monitor checkpoint creation during training, watch the ResumableCheckpoint and ModelCheckpoint fields in DescribeJob output.

Resume an interrupted job

If a job fails or is stopped mid-training, you can start a new job that picks up from the exact step where it left off. The platform restores the full training state — weights, optimizer momentum, and step counter — from the resumable checkpoint.

Specify a resumable checkpoint from the Intermediate Checkpoint Model Package Group as InputModelPackageArn:

"ModelPackageConfig": {
  "OutputModelPackageGroupArn": "arn:aws:sagemaker:us-west-2:123456789012:model-package-group/my-final-models",
  "IntermediateCheckpointModelPackageGroupArn": "arn:aws:sagemaker:us-west-2:123456789012:model-package-group/my-intermediate-checkpoints",
  "InputModelPackageArn": "arn:aws:sagemaker:us-west-2:123456789012:model-package/my-intermediate-checkpoints/5"
}

The InputModelPackageArn must point to a resumable checkpoint (one with IsCheckpoint=true in its Model Package metadata). Training resumes from the step after the checkpoint — for example, if the checkpoint was saved at step 4, training continues from step 5.

The following must stay the same between the original job and the resumed job:

Continue training on a new job (iterative training)

Iterative training lets you build on a previously trained model with a different dataset, different hyperparameters, or a refined reward function. Unlike resuming, this starts a fresh training run — the optimizer resets, the step counter resets to 0, and only the trained LoRA weights carry over.

Specify a model checkpoint from the Output Model Package Group as InputModelPackageArn:

"ModelPackageConfig": {
  "OutputModelPackageGroupArn": "arn:aws:sagemaker:us-west-2:123456789012:model-package-group/my-final-models",
  "IntermediateCheckpointModelPackageGroupArn": "arn:aws:sagemaker:us-west-2:123456789012:model-package-group/my-intermediate-checkpoints",
  "InputModelPackageArn": "arn:aws:sagemaker:us-west-2:123456789012:model-package/my-final-models/3"
}

What you can change between iterations:

What must stay the same:

Common patterns for iterative training:

Checkpoint best practices