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.

Quality

83%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Quality

Content

78%Weight 40%Scale 1-5

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

A well-structured reference-desk skill body: concise overview, an actionable file-routing table, a clear retrieval workflow, and disciplined progressive disclosure pointing to one-level-deep sources. The main improvement is reducing cross-section repetition of the same caveats.

Suggestions

Consolidate the repeated MOPD pipeline order and the pretraining-vs-post-training-quantization distinction into a single authoritative location, cross-referencing it from other sections instead of restating.

Tighten a few abstract answering rules (e.g., "explain the hybrid Mamba-2 + attention + LatentMoE design") into concrete required outputs (e.g., "state total/active params and name each of the three components").

Verify the referenced bundle files (paper/*.md, paper/mopd/*.md, model-card.md, INDEX.md, context/quick-reference.md) are actually shipped with the skill so the routing table resolves.

Dimension	Reasoning	Score
Conciseness	Mostly lean and bullet/table-driven with no generic concept padding, but a few caveats are restated across sections (MOPD stage order, pretraining-vs-post-training quantization, MOPD≠RLHF) and could be consolidated.	4 / 5
Actionability	A concrete routing table maps question types to specific files, citation examples are explicit, and answering rules are specific; as an instruction-only skill the absence of code is not penalized, with only minor abstractness in a few rules.	4 / 5
Workflow Clarity	A clear Locate → Retrieve → Cite sequence with sub-ordering and a routing table; citation/labeling rules act as quality gates, and no destructive or batch operations apply to cap the score.	4 / 5
Progressive Disclosure	SKILL.md is a lean overview that points to one-level-deep references (paper/.md, paper/mopd/.md, model-card.md, INDEX.md, context/quick-reference.md) via a well-signaled routing table for easy navigation; no bundle files were present to verify, so this is scored on the reference structure as written.	5 / 5
	Total	17 / 20 Passed

Description

87%Weight 40%Scale 1-5

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

A strong, third-person description that crisply states both the what (reference desk for a specific model across named topic areas) and the when (a natural trigger with a negative boundary). Minor room to add natural synonyms for the trigger phrasing.

Dimension	Reasoning	Score
Specificity	Enumerates several concrete sub-domains ("architecture, NVFP4 pretraining, SFT, MOPD (multi-teacher on-policy distillation), MTP boosting, quantization, inference") under one named action ("Reference desk"), giving comprehensive coverage with only the single action verb.	4 / 5
Completeness	Clearly states what it does ("Reference desk for NVIDIA Nemotron 3 Ultra…") and when to use it with a concrete trigger plus a disambiguating negative ("rather than building a pipeline").	5 / 5
Trigger Term Quality	"Use when the user asks facts about Ultra rather than building a pipeline" supplies natural trigger phrasing and good keyword coverage, but lacks synonyms/extensions (none meaningfully apply to this knowledge skill).	4 / 5
Distinctiveness Conflict Risk	A narrow, named niche ("NVIDIA Nemotron 3 Ultra (550B-A55B)") with distinct triggers and an explicit handoff boundary to /nemotron-customize, giving minimal conflict risk.	5 / 5
	Total	18 / 20 Passed

Validation

100%

Warnings & errors only

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

Validation — 16 / 16 Passed

Validation for skill structure

No warnings or errors.

Repository: NVIDIA-NeMo/Nemotron
Path: skills/nemotron-ultra/SKILL.md
Commit: 1a4c3fa

Reviewed: 4 days ago

Table of Contents

Discovery Implementation Validation

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.