[evakravi]

Problem

ModelTrainer.train() returns None and stores the created job on the private _latest_training_job. A caller that wants the job name or ARN after train(wait=False) reads the private attribute or pages ListTrainingJobs by name prefix. The SDK's own docs do the former (docs/training/index.rst, docs/model_customization/model_customization.rst), seven v3-examples notebooks do the same, and ModelBuilder reads self.model._latest_training_job in three places. The SageMaker agent skills (aws/agent-toolkit-for-aws, aws-ai-ml/references/training-jobs) refuse to read a private attribute, so each of their three training templates polls list_training_jobs(NameContains=<prefix>) plus describe_training_job to recover what the trainer already holds.

DPOTrainer.train() also stored its job on latest_training_job while SFTTrainer, RLVRTrainer, and RLAIFTrainer used _latest_training_job, so show_metrics() and stream_logs(), which read the private name, could not find a DPO job.

Solution

Add a public latest_training_job property with a typed setter on ModelTrainer and BaseTrainer (the setter accepts a TrainingJob or None, and raises TypeError with a TrainingJob.get(<name>) hint for a bare string). ModelTrainer.train() returns the created TrainingJob for a SageMaker training job, after the wait when wait=True, and None for a dry run, a PipelineSession capture, and local mode. DPOTrainer stores its job on the shared attribute. The docs and the error messages of show_metrics(), stream_logs(), and ModelBuilder point at the public name. No existing attribute or return value changes for a caller that reads _latest_training_job.

Tests

New sagemaker-train/tests/unit/train/test_latest_training_job.py, 11 tests: train() returns the created job with and without wait, dry_run returns None and creates nothing, the property is None before train(), the setter attaches a job and rejects a string, output_data works through the setter, BaseTrainer round-trips the setter, show_metrics() and stream_logs() raise the new message, and SFTTrainer and DPOTrainer expose the created job after train(wait=False).

cd sagemaker-train && python -m pytest tests/unit/train/test_latest_training_job.py -q
11 passed

On origin/master the same file reports 9 failed, 2 passed (AttributeError: 'ModelTrainer' object has no attribute 'latest_training_job', assert None is <TrainingJob>).

Existing trainer suites on this branch: tests/unit/train/test_model_trainer.py, test_sft_trainer.py, test_dpo_trainer.py, test_rlvr_trainer.py, test_rlaif_trainer.py, test_base_trainer_serverful.py, test_base_trainer_compute.py, test_mtrl_eval_mlflow_url.py: 334 passed. sagemaker-serve tests/unit/test_model_builder_v3.py and test_model_builder_core.py: 63 passed, the one failure (test_build_with_vllm_for_text_generation, an S3 GetObject with expired credentials) fails identically on origin/master.

black --line-length 100, flake8, and pydocstyle pass on the changed files.

Merge Checklist

• I read the contribution guide.
• The change is backward compatible.
• The commits use the repository commit-message format.
• The change creates no S3 or STS client.
• The change has regression coverage and adds no dependency.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[evakravi]

Superseded by #6533, which combines the four changes into one PR and moves the hub-model metadata accessors onto HubContent itself (a HubContentMixin in sagemaker.core.hub_content) instead of a separate class.

[github-actions]

🤖 AI code review

This PR adds a public latest_training_job property (with a typed setter that rejects bare strings) to ModelTrainer and BaseTrainer, makes ModelTrainer.train() return the created TrainingJob (and None for dry-run/PipelineSession/local mode), fixes DPOTrainer to store on the shared _latest_training_job, and updates docs and error messages to point at the public name. I checked the pydantic PrivateAttr backing on ModelTrainer, the ABC-based BaseTrainer, all subclass trainers, callers of _latest_training_job, and existing tests — the change is consistent and backward compatible for private-attr readers, and the new test file covers the behaviour. The change aligns with the convention from PR #6436 (store on _latest_training_job). No blocking issues found; one incidental typo fix is included.
[]

_{Reviewed commit bc92c8f. Automated review; verify before acting.}