CtrlK
BlogDocsLog inGet started
Tessl Logo

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.

70

Quality

86%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

SecuritybySnyk

Low

Low-risk findings worth noting

SKILL.md
Quality
Evals
Security

Quality

Content

80%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.

Highly actionable content with executable examples and clean progressive disclosure into four real reference files. The main gap is missing inline validation/feedback loops in the batch training workflows, which caps workflow clarity.

Suggestions

Add explicit validation checkpoints inside each training workflow (e.g., after launch: 'verify loss is decreasing and MFU >40%; if not, see Common issues'), rather than relegating all troubleshooting to a separate section.

Trim light concept explanations Claude already knows (e.g., the one-line description of how expert parallelism reduces memory) to tighten conciseness.

Add a brief 'verify checkpoint integrity' or 'confirm convergence in first N steps' step to the multi-node workflows so destructive/expensive batch runs have a feedback loop.

DimensionReasoningScore

Conciseness

Mostly efficient with executable bash/tables/checklists rather than prose, but a few light explanations of concepts Claude already knows ('Expert parallelism distributes different experts across GPUs, reducing memory while maintaining capacity') could be trimmed.

4 / 5

Actionability

Fully executable, copy-paste-ready torchrun/bash commands with concrete configs for LLaMA 70B, Mixtral MoE, and 405B throughput tuning, plus specific troubleshooting flags covering the common cases.

5 / 5

Workflow Clarity

Workflows are clearly sequenced with copy-this checklists, but large-scale training is a batch operation and the workflows lack inline validation/feedback checkpoints (e.g., 'if loss diverges, do X'), capping this dimension at 3 per the batch-operation guidance.

3 / 5

Progressive Disclosure

SKILL.md is a well-organized overview (quick start, workflows, issues, hardware) that defers depth to four clearly signaled, one-level-deep references — parallelism-guide, benchmarks, production-examples, training-recipes — all of which exist as real files.

5 / 5

Total

17

/

20

Passed

Description

92%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 answers what and when with concrete trigger phrases and third-person voice. Minor room to add natural synonyms like 'pretraining' and 'distributed training' for fuller trigger coverage.

DimensionReasoningScore

Specificity

Lists multiple concrete actions — 'Trains large language models (2B-462B parameters) using NVIDIA Megatron-Core with advanced parallelism strategies' — and enumerates concrete parallelism types (tensor/pipeline/sequence/context/expert), giving comprehensive coverage.

5 / 5

Completeness

Clearly answers both 'what' (trains LLMs with Megatron-Core via advanced parallelism) and 'when' with an explicit 'Use when...' clause carrying concrete trigger phrases (model size, GPU efficiency, parallelism needs).

5 / 5

Trigger Term Quality

Good natural-keyword coverage ('training models >1B parameters', 'maximum GPU efficiency', '47% MFU on H100', named parallelism variants and model families), though common synonyms like 'pretraining' and 'distributed training' are not stated explicitly.

4 / 5

Distinctiveness Conflict Risk

Occupies a clear niche (NVIDIA Megatron-Core, large-scale LLM training, advanced multi-axis parallelism) with distinct triggers and named production frameworks, minimizing overlap with general training skills.

5 / 5

Total

19

/

20

Passed

Validation

93%

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

Validation15 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

frontmatter_unknown_keys

Unknown frontmatter key(s) found; consider removing or moving to metadata

Warning

Total

15

/

16

Passed

Repository
OpenLAIR/dr-claw
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.