Add TAO Toolkit Configuration

[hqtran-nv]

Onboarding type

• New product onboarding (new components.d/<slug>.yml file)
• Other (catalog change, README fix, infrastructure, etc.)

For new product onboarding — author affirmations

By submitting this PR, I confirm on behalf of my team:

• Skills cleared for open source release per NVIDIA's internal IP review process (six-question check, all answers affirmative)
• License selected: Apache 2.0 / CC-BY 4.0 / Dual (Apache 2.0 + CC-BY 4.0). Specify: Apache 2.0
• No new license or new third-party component introduced beyond what the source repo already carries
• Source repo is public and under an NVIDIA-owned GitHub org
• .agents/skills/ or skills/ path used for new entries (or existing path retained for legacy entries per components.d/<slug>.yml)

NVIDIA contributors: see the internal onboarding guide for the IP review process details and license selection.

Reviewer checklist (OSS Skills PIC)

• Author confirmations above are checked
• components.d/<slug>.yml entry valid (required fields, unique catalog_dir, path exists in source repo, filename slug matches name)
• SKILL.md frontmatter spec-compliant (at least one sampled)
• No new license or third-party dependency requiring OSRB filing

All PRs

• All commits signed off with DCO (git commit -s).
  If you forgot, run git rebase --signoff origin/main && git push --force-with-lease to retroactively sign all commits in your branch.

Other context (for non-onboarding PRs)

[mosheabr]

Thanks Adrian! Source repo NVIDIA-TAO/tao-skills-bank is public + Apache-2.0 + NVIDIA-org (✅), and the TAO Toolkit is a big addition for the catalog. Holding the merge — several blockers to clear before this can land.

1. All 5 declared paths return 404 on source

PR declares	Returns	Actual TAO repo layout
path: skills/applications/	404	applications/ (12 skills)
path: skills/data/	404	data/ (9 skills)
path: skills/models/	404	models/ (53 skills)
path: skills/platform/	404	platform/ (8 skills)
path: skills/misc/	404	misc/ (empty)

There's a stray skills/ prefix on every path. The actual skill directories live at the top level of the repo, not under a skills/ parent. Net: every path needs the skills/ prefix dropped.

2. Bulk-sync pattern is deprecated for new components

The current entry uses one path per category (e.g., path: applications/ + catalog_dir: tao-applications), which would sweep all 12 application skills under one nested catalog dir at skills/tao-applications/<skill-name>/. That's the bulk-sync layout we deprecated in components.d/README.md last week (PR #191) and migrated VSS + Megatron-Core away from this week.

→ Switch to flat layout: one components.d/tao.yml entry per skill, with product-prefixed catalog_dir. Example:

skills:
  - path: applications/tao-run-automl/
    catalog_dir: tao-run-automl
  - path: applications/tao-analyze-changenet-rca/
    catalog_dir: tao-analyze-changenet-rca
  - path: models/tao-train-reid/
    catalog_dir: tao-train-reid
  # ... one entry per skill

3. No skill artifacts on source

Sampled applications/tao-run-automl/, models/tao-train-reid/, and skills/tao-launch-workflow/ — none carry the 5-artifact set the catalog enforces:

Artifact	Present in samples?
SKILL.md	✅
skill-card.md	❌
skill.oms.sig	❌
evals/ (eval dataset)	❌
BENCHMARK.md	❌

The sync workflow's enforcement gate drops any skill missing these. Without them, even a path-corrected version of this PR would result in 0 TAO skills landing in the catalog.

→ Run the NVCARPS signing flow on the source skills (GitHub-first guide: Confluence 3483240468). For each skill in scope, that produces the sig, the skill card, and the BENCHMARK. NVCARPS-EVAL also generates the eval dataset.

Mohit Gupta is the right point of contact for signing flow questions.

4. 82 skills — confirm scope with TAO product team

Source repo has ~82 skill directories across the four categories. Worth a sanity check with the TAO PMs on which skills are genuinely customer-facing vs contributor-helper. We had a lesson learned with NeMo-RL where contributor-skill leaks (e.g., mcore-cicd, mcore-linting) ended up in the public catalog because the bulk import didn't differentiate. Contributor skills belong in .agents/contributor-skills/ on the source repo, not in the catalog.

→ Once the team agrees on the customer-facing set, that's the subset to sign + enumerate in components.d/tao.yml.

Once all four are done

Ping me here and I'll re-verify (model_signing verify, source-path checks, sample review), then approve + merge. Same-day turnaround once content is signing-complete.

Holding the PR open meanwhile.