Exemples de jeux de données et de didacticiels

End-to-end exemple : agent d'accueil des clients

L'exemple suivant illustre le flux de travail complet pour une RL multitour : formation d'un agent, évaluation de ses performances et déploiement du modèle entraîné.

from sagemaker.train.multi_turn_rl_trainer import MultiTurnRLTrainer
from sagemaker.train.evaluate import MultiTurnRLEvaluator
from sagemaker.serve import ModelBuilder
from sagemaker.core.resources import ModelPackage

DATASET = "s3://my-bucket/prompts/prompts.parquet"
AGENT_ARN = "arn:aws:bedrock-agentcore:us-west-2:123456789012:runtime/my-agent"

# ── 1. Train ──────────────────────────────────────────────────────────────
trainer = MultiTurnRLTrainer(
    model="openai-reasoning-gpt-oss-20b",
    agent_env=AGENT_ARN,
    training_dataset=DATASET,
    s3_output_path="s3://my-bucket/train-output/",
    accept_eula=True,
)
trainer.hyperparameters.max_epochs = 1
trainer.hyperparameters.global_batch_size = 10

job = trainer.train(wait=True)
print(f"Training complete: {job.output_model_package_arn}")

# ── 2. Evaluate ───────────────────────────────────────────────────────────
evaluator = MultiTurnRLEvaluator(
    model=trainer,
    dataset=DATASET,
    s3_output_path="s3://my-bucket/eval-output/",
    evaluate_base_model=True,  # Compare base vs fine-tuned
)
execution = evaluator.evaluate()
execution.wait()
print(f"Evaluation: {execution.status.overall_status}")

# ── 3. Deploy ─────────────────────────────────────────────────────────────
model_package = ModelPackage.get(model_package_name=job.output_model_package_arn)

model_builder = ModelBuilder(
    model=model_package,
    instance_type="ml.g6e.48xlarge",
)
model_builder.accept_eula = True
model_builder.build()

endpoint = model_builder.deploy(
    endpoint_name="mtrl-production-endpoint",
    instance_type="ml.g6e.48xlarge",
    initial_instance_count=1,
)
print(f"Deployed: {endpoint.endpoint_name}")