TAO Standard Training Workflow
Run a controlled TAO train, evaluate, and export workflow on labeled data.
The document requires confirmation of the container image, asks the user to select a platform, and limits credential requests to platform-relevant and model-relevant credentials. It also declares Docker and NVIDIA Container Toolkit requirements. Deductions apply because allowed-tools includes broad Bash and Write access, while credential handling, external data flows, isolation boundaries, rollback, and cleanup of failed training artifacts are unspecified.
The workflow, prerequisites, and AutoML routing are broadly coherent, and the skill requires model metadata and image resolution. Reliability is reduced because a nominal single-step workflow may route through AutoML by default, depends on other model skills, external scripts, and platform configuration, and lacks key-path tests, abnormal-input handling, and diagnosable failure feedback. Static calibration limits this score to 10 or below, so it receives 7.
Trigger phrases, train/eval/export scenarios, required and optional inputs, and supported platforms are clearly listed. The document also distinguishes the normal path from augmentation, AutoML, and DEFT loops. Deductions apply because non-fit boundaries remain incomplete, model coverage depends on other skills, Chinese-language support is not stated, and reachability of Docker, NVIDIA container tooling, and external platforms from mainland China is not evidenced.
The skill has frontmatter, Apache-2.0 licensing, author and version metadata, tags, structured skill_info, and a benchmark with repository-level maintenance signals. Deductions apply because standalone Instructions, Examples, FAQs, known limitations, changelog, and explicit update responsibility are missing. The benchmark also reports folder hierarchy and author-format convention findings, while installation guidance is primarily in shared repository documentation.
The documentation can directly guide collection of the model, dataset, platform, image, monitoring, and train/eval/export choices. The committed one-task benchmark reports relatively strong effectiveness, but its sample is extremely small. Deductions apply because execution requires external model skills, scripts, platform setup, and credentials; representative outputs are not verified in this skill's files, so users may still need substantial environment-specific adaptation. Static calibration limits effectiveness to 7 or below, so it receives 6.
A revision-pinned source, structured skill_info, an evaluation task, and an NVSkills-Eval report provide limited auditability. Deductions apply because the benchmark has one task, one attempt per agent, no negative-trigger cases, and no committed tests covering this skill's key execution paths. Static review cannot independently reproduce execution claims, so the score is capped at 5 and assigned 4.
- Before execution, verify that the selected model skill, platform scripts, container image, and credential requirements exist and match the assessed revision.
- Do not grant Bash or Write access beyond what training requires; clarify external flows and retention for datasets, logs, checkpoints, and exports before launch.
- Add failure recovery, cancellation, cleanup, and rollback procedures, and verify reachability of required images, platforms, and storage services from mainland-China environments.
What does this skill do, and when should you use it?
This NVIDIA TAO Toolkit skill defines a generic single-step workflow for training a TAO model on labeled data. It can train, optionally evaluate, and optionally export a model, with model-level AutoML used when the resolved metadata enables it. Before submission, it collects the target platform, relevant credentials, dataset inputs, and confirmed container image. It is intended for plain training runs and does not cover iterative augmentation, AutoML loops as a standalone workflow, or DEFT loops.
It requests a compatible TAO model and training-dataset URI, obtains a supported-platform choice, and inspects model metadata to decide whether to route training through AutoML. It resolves the training container image and requires confirmation or an image override, invokes the platform, model-metadata, and image-resolution helpers, and routes execution through the selected model skill and tao-launch-workflow. It runs evaluation when an eval_dataset_uri is resolved and can export the trained model on request.
- A computer-vision developer adapting a compatible TAO model to a labeled dataset in one standard training run.
- An engineer who wants a train-then-evaluate-then-export workflow without iterative data-augmentation or DEFT loops.
- A user who explicitly wants AutoML disabled for a plain training run.
- A team launching TAO training on a supported local Docker, Slurm, Kubernetes, or other platform workflow.
What are this skill's strengths and limitations?
- Defines a clear train, optional evaluation, and optional export sequence.
- Uses model metadata to route eligible runs through AutoML or stay on the plain path.
- Requires container-image confirmation before runner creation or submission.
- Delegates model-specific actions to the corresponding TAO model skill.
- Requires Docker and nvidia-container-toolkit, plus any workflow-specific requirements.
- Requires platform selection and image confirmation before submission.
- Model-specific configuration details are delegated to other skills rather than documented here.
- The source does not provide an independent test suite or success evidence for individual platforms.
How do you install this skill?
Install the individual skill with the NVIDIA skills CLI: npx skills add nvidia/skills --skill tao-train-single-step --yes. The README does not document additional steps for manually copying only this skill directory.
How do you use this skill?
Ask a loaded agent to “run a standard TAO train, then evaluate and export” or “run plain TAO training with AutoML off.” Provide the model, training-dataset URI, target platform, and required credentials. Confirm the resolved default container image or provide image=<override>. Add eval_dataset_uri when evaluation is needed and base_checkpoint when starting from a specified checkpoint.