CtrlK
BlogDocsLog inGet started
Tessl Logo

nemotron-ultra

Reference desk for NVIDIA Nemotron 3 Ultra (550B-A55B) — architecture, NVFP4 pretraining, SFT, MOPD (multi-teacher on-policy distillation), MTP boosting, quantization, inference. Use when the user asks facts about Ultra rather than building a pipeline.

75

Quality

92%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Quality

Content

85%

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

A high-quality reference-desk skill: concrete file-level routing, an explicit retrieve-and-cite workflow, and well-structured one-level-deep references. The only weakness is mild cross-section redundancy of the key pipeline order and recipe-gap caveats, which keeps conciseness just below the top anchor.

Suggestions

State the MOPD pipeline order (SFT → RLVR → MOPD warmup → MOPD (×N) → MTP boosting) once in 'Answering rules' and reference it elsewhere rather than repeating the full sequence in 'What makes Ultra different' and 'Cross-skill handoff'.

Consolidate the public-recipe-gap caveat (no bundled long-context pretraining data / no full two-iteration MOPD reproduction) into a single 'Known caveats' entry and link to it from 'Cross-skill handoff' instead of restating it.

Add a one-line verification checkpoint in the Cite step (e.g., 'confirm every benchmark number is labeled base / BF16 / NVFP4 before answering') to make the existing labeling rule an explicit workflow gate.

DimensionReasoningScore

Conciseness

Largely lean and table/bullet-driven with no padding of basic concepts, but the MOPD pipeline order ('SFT → RLVR → MOPD warmup → MOPD (×N) → MTP boosting') is restated three times and the public-recipe-gap caveat twice, so it could be tightened per the score-2 anchor.

2 / 3

Actionability

For an instruction-only reference skill it gives fully actionable, concrete guidance — a routing table mapping each question type to an exact file, a source-priority order, and a citation format ('paper/architecture.md → Table 1') — meeting the score-3 standard without needing executable code.

3 / 3

Workflow Clarity

The Locate → Retrieve → Cite workflow is explicitly sequenced with an ordered read path and routing table, and the Do/Do-not and Known-caveats sections act as answer-quality checklists; no destructive/batch operations are involved, so the missing-feedback-loop cap does not apply.

3 / 3

Progressive Disclosure

The body is an overview that routes to clearly signaled one-level-deep detail files via the routing table and source-priority list ('paper/architecture.md', 'paper/mopd/overview.md', 'model-card.md'), matching the score-3 anchor for well-organized, easy navigation; no bundle dirs were present to verify the referenced files locally.

3 / 3

Total

11

/

12

Passed

Description

100%

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

The description is exemplary: it names a comprehensive set of concrete capabilities, includes an explicit 'Use when...' trigger with natural user phrasing, and carves out a distinctive niche that actively disambiguates from a sibling customization skill. It cleanly satisfies every dimension at the top of the scale.

DimensionReasoningScore

Specificity

Lists multiple specific concrete capabilities — 'architecture, NVFP4 pretraining, SFT, MOPD (multi-teacher on-policy distillation), MTP boosting, quantization, inference' — rather than vague language; clearly above the score-2 anchor which expects only partial domain/action coverage.

3 / 3

Completeness

Explicitly answers both what ('Reference desk for NVIDIA Nemotron 3 Ultra — architecture, NVFP4 pretraining, SFT, MOPD...') and when with an explicit 'Use when...' trigger clause, satisfying the score-3 anchor; it is not capped at 2 because the trigger clause is present.

3 / 3

Trigger Term Quality

Includes the natural term users would say — 'Use when the user asks facts about Ultra' plus the model name 'NVIDIA Nemotron 3 Ultra' — matching the score-3 anchor for good coverage of natural phrasing rather than jargon-only.

3 / 3

Distinctiveness Conflict Risk

A clear narrow niche (Nemotron 3 Ultra facts) with a distinct trigger that even excludes pipeline-building ('rather than building a pipeline'), making conflict with other skills unlikely per the score-3 anchor.

3 / 3

Total

12

/

12

Passed

Validation

100%

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation16 / 16 Passed

Validation for skill structure

No warnings or errors.

Repository
NVIDIA-NeMo/Nemotron
Reviewed

Table of Contents

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.