training-llms-megatron

Trains large language models (2B-462B parameters) using NVIDIA Megatron-Core with advanced parallelism strategies. Use when training models >1B parameters, need maximum GPU efficiency (47% MFU on H100), or require tensor/pipeline/sequence/context/expert parallelism. Production-ready framework used for Nemotron, LLaMA, DeepSeek.

Quality

90%

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

86%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, highly actionable skill body that leads with executable quick-start commands and workflows, delegates advanced detail to real reference files, and includes troubleshooting feedback loops. Minor gains are available from trimming light explanatory prose and adding explicit pre-scale validation checkpoints.

Suggestions

Trim light explanatory sentences (e.g., the EP memory-savings narrative) so the body stays purely instructional, lifting conciseness toward the 5 anchor.

Add an explicit validation gate early in each training workflow (e.g., 'run 100 steps and confirm loss decreases and MFU >40% before launching the full run') to make checkpoints explicit rather than implicit.

Replace the 'torchrun ...' ellipsis in the micro-batch sweep with a concrete reusable command snippet so every example is fully copy-paste runnable.

Dimension	Reasoning	Score
Conciseness	Largely lean and code-forward, assuming Claude's competence (no basic concept explanations), with only minor over-explanation such as 'Expert parallelism distributes different experts across GPUs, reducing memory while maintaining capacity' that could be trimmed, placing it just below the fully lean 5 anchor.	4 / 5
Actionability	Provides fully executable, copy-paste-ready bash/torchrun commands with real flags covering common cases (LLaMA 70B 3D parallelism, Mixtral MoE, FP8 optimization, troubleshooting fixes), matching the 5 anchor; the single 'torchrun ...' ellipsis in the micro-batch sweep is a justified shorthand.	5 / 5
Workflow Clarity	Workflows are clearly sequenced with copy-this-checklist steps (1-4) and a Common Issues section providing error-recovery feedback loops, plus a Monitor-metrics verification step, satisfying most 5 criteria; however in-workflow explicit validation gates (e.g., confirm loss is decreasing before committing to a long run) are implicit rather than stated, keeping it at 4.	4 / 5
Progressive Disclosure	SKILL.md is an overview with well-signaled one-level-deep references to four real bundle files (parallelism-guide, benchmarks, production-examples, training-recipes), each with a descriptive pointer and no deeper nesting, matching the clear-navigation 5 anchor.	5 / 5
	Total	18 / 20 Passed

Description

95%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, specific description that clearly states both capability and trigger conditions with concrete parameter ranges and named models. It is third-person, avoids fluff, and is highly distinguishable from adjacent training tools.

Dimension	Reasoning	Score
Specificity	Names the domain and tool with concrete specifics ('Trains large language models (2B-462B parameters) using NVIDIA Megatron-Core') and enumerates parallelism strategies, but the action set is essentially a single capability (training) rather than multiple distinct actions, so it falls just short of the comprehensive 5 anchor.	4 / 5
Completeness	Explicitly answers what ('Trains large language models ... using NVIDIA Megatron-Core with advanced parallelism strategies') and when ('Use when training models >1B parameters, need maximum GPU efficiency ... or require ... parallelism') with concrete trigger phrases, matching the 5 anchor.	5 / 5
Trigger Term Quality	Covers natural phrases a user would say ('training models >1B parameters', 'maximum GPU efficiency', 'MFU on H100') plus synonyms and named entities (tensor/pipeline/sequence/context/expert parallelism, Nemotron, LLaMA, DeepSeek), matching the comprehensive-coverage 5 anchor.	5 / 5
Distinctiveness Conflict Risk	Clear niche (Megatron-Core, 2B-462B params, NVIDIA/H100) with distinct triggers and named production models, minimizing overlap with general training skills per the 5 anchor.	5 / 5
	Total	19 / 20 Passed

Validation

93%

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 — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/megatron-core/SKILL.md
Commit: 3a6c3a9

Reviewed: 1 day 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.