TAO DAFT Dataset Validator
Validate TAO DAFT dataset structure, schemas, and cross-references with tao-daft validate.
The skill limits its declared tools to Read and Bash and targets validation of local datasets; there is no evidence of credential theft, covert exfiltration, or destructive behavior. However, Bash can install dependencies and execute commands, while explicit execution confirmation, data-flow disclosure, and rollback guidance are absent. NVIDIA authorship, Apache-2.0 licensing, and the supplied official provenance support attribution, but repository license metadata is NOASSERTION, so full marks are not justified.
The workflow is internally coherent: check the dependency, inspect CLI help, choose a format, run validation, and interpret exit status. It also supplies several diagnosable troubleshooting cases. The dependency is unpinned, no selected-skill test suite or reproducible static evidence is provided, and format inference coverage is limited; the static ceiling therefore applies.
Triggers, DAFT versus non-DAFT boundaries, ambiguity handling, and exclusion of info/convert workflows are clearly stated. Input/output contracts, broader format coverage, Chinese-language support, and mainland-China environment fit are not documented; dependency installation may also depend on network reachability, so the score is below full marks.
The documentation is well layered with quick starts, prerequisites, CLI conventions, limitations, and troubleshooting. Version, author, and Apache-2.0 license are stated. The benchmark notes a missing recommended Examples section and an author-format issue; changelog, explicit maintenance ownership, and an update path are also absent, so points are deducted.
The core task—running tao-daft validate for structure, schema, and cross-reference checks—is directly supported by usable commands, format discovery, and result-interpretation guidance. No representative output for this skill is independently verifiable, and the result depends on an external CLI; static review therefore supports the capped score of 7.
The supplied SKILL.md, skill card, evaluation file, and benchmark report are auditable and mutually consistent about triggers and workflow. However, the benchmark contains only one task and there is no independently reproducible committed test suite or CI coverage for this skill, preventing a higher score.
- Before executing Bash, confirm authorization, the target path, and the effects of dependency installation; the skill lacks explicit confirmation and rollback steps.
- Verify the installed nvidia-tao-daft version and supported format slugs first, because the documentation says CLI flags may drift across versions.
- Chinese output and mainland-China network compatibility are unspecified; dependency installation may not reach the default package index directly.
What does this skill do, and when should you use it?
This skill guides an agent through NVIDIA TAO DAFT dataset validation with tao-daft validate. It covers structural, schema, and cross-reference checks for a dataset, scene, or parent directory. The agent discovers supported formats and flags from the installed CLI rather than relying on a fixed list. It is limited to DAFT layouts and exits non-zero when validation fails.
Checks whether nvidia-tao-daft is installed, reads local dataset directories, infers a format from directory markers when possible, and runs tao-daft --version, tao-daft validate --help, format-specific help, and tao-daft validate <format> --path <path>. It explains how to read the VALIDATION RESULTS block and the VALIDATION PASSED or VALIDATION FAILED status, including how to capture large output for review.
- A data engineer needs to verify that a TAO DAFT dataset conforms to its selected format before downstream use.
- A team preparing a dataset with meta.json, media/, and text/ needs to validate cosmos-reason-v1.0 structure.
- A team working with contextual/, raw/, and task/ directories needs to validate a metropolis-v3.0 dataset.
- A CI pipeline must treat warnings as failures and needs to discover whether the installed format supports --strict.
What are this skill's strengths and limitations?
- Defines a complete workflow from dependency preflight and format discovery through result interpretation.
- Supports a single dataset, scene, or parent directory containing a dataset tree.
- Clearly separates format inference, per-format flags, and non-zero failure handling.
- The skill metadata identifies Apache-2.0 licensing, NVIDIA Corporation as author, and version 0.1.0.
- It validates DAFT only and does not cover COCO, YOLO, or Data Factory JSONL.
- Supported formats and flags depend on the installed version; older slugs may be retired.
- The supplied skill text provides no independent test suite or platform-validation evidence.
- It covers validate only, not tao-daft info or tao-daft convert.
How do you install this skill?
Install the NVIDIA skills collection with: npx skills add nvidia/skills. To target Codex and select this skill, use: npx skills add nvidia/skills --skill tao-validate-dataset-format --agent codex. The skill requires Python 3.10+ and nvidia-tao-daft; install the package with: pip install nvidia-tao-daft. The README does not document additional manual-copy steps for this individual skill.
How do you use this skill?
Ask the agent to “validate this TAO DAFT dataset” or “run tao-daft validate.” In the environment, run: tao-daft --version; tao-daft validate --help; tao-daft validate <format> --help; tao-daft validate <format> --path /path/to/daft-dataset. The format is a positional subcommand and the target must follow --path. If neither the cosmos-reason-v1.0 markers nor contextual/ markers are present, ask the user instead of guessing.