CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

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

82%

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

75%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 operational skill: dense executable configurations, clearly sequenced workflows with monitoring checkpoints, and clean offloading of deep material to four real reference files. The main improvements are small: replace placeholder commands, drop the stale example-file reference, and reduce duplication between inline scripts and training-recipes.md.

Suggestions

Replace the 'torchrun ... --micro-batch-size $MBS' placeholder in the micro-batch sweep with a complete runnable command, and remove or inline the './examples/llama/train_llama3_8b_fp8.sh' reference since no examples directory ships with the bundle.

Trim duplication between the inline 70B/MoE launch scripts and references/training-recipes.md, keeping only one representative script in SKILL.md and pointing to the recipes file for the rest.

Add an explicit validation loop to Workflow 1's monitoring step (e.g. 'If MFU < 40%, apply the fixes in Common issues and re-measure') to convert the listed metrics into an actionable checkpoint.

DimensionReasoningScore

Conciseness

The body is dense with operational data Claude would not know (parallelism sizing tables, launch flags, memory math) and avoids concept padding; minor over-explanation remains, e.g. 'Expert parallelism distributes different experts across GPUs, reducing memory while maintaining capacity' and repeated 47% MFU claims.

4 / 5

Actionability

Mostly executable guidance: full torchrun scripts with real Megatron flags, SLURM launch commands, and concrete model-size-to-parallelism tables. Minor gaps keep it below 5 — the micro-batch sweep uses a literal 'torchrun ... --micro-batch-size $MBS' placeholder, and './examples/llama/train_llama3_8b_fp8.sh' references a file not present in the bundle.

4 / 5

Workflow Clarity

Workflows are sequenced with copy-able checklists, explicit numbered steps, and a monitoring step with concrete targets ('MFU: Target >40% on H100', 'Memory usage: <80GB per GPU'), and the Common Issues section provides error-recovery feedback. Falls short of 5 because the workflows stop at listing metrics without an explicit act-on-threshold validation loop.

4 / 5

Progressive Disclosure

The Advanced topics section clearly signals four one-level-deep reference files (all verified to exist), each with a purpose statement, and the body stays a genuine overview. Minor gap: the inline 70B and MoE launch scripts partially duplicate what references/training-recipes.md already provides.

4 / 5

Total

16

/

20

Passed

Description

83%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 description that concretely states what the skill does and gives explicit 'Use when' triggers with specific technical conditions. Its only weaknesses are a somewhat single-action framing and missing a few natural trigger synonyms like 'distributed training' or 'fine-tuning'.

DimensionReasoningScore

Specificity

Names a concrete capability with specifics ('Trains large language models (2B-462B parameters) using NVIDIA Megatron-Core with advanced parallelism strategies') in third person, but describes a single primary action rather than the multiple distinct concrete actions of the 5 anchor.

4 / 5

Completeness

Explicitly answers both what ('Trains large language models (2B-462B parameters) using NVIDIA Megatron-Core') and when ('Use when training models >1B parameters, need maximum GPU efficiency (47% MFU on H100), or require tensor/pipeline/sequence/context/expert parallelism') with concrete trigger phrases.

5 / 5

Trigger Term Quality

Good natural keyword coverage ('training models >1B parameters', 'GPU efficiency (47% MFU on H100)', 'tensor/pipeline/sequence/context/expert parallelism'), but misses common variations users would say such as 'distributed training', 'pretraining', or 'fine-tuning'.

4 / 5

Distinctiveness Conflict Risk

The Megatron-Core/NVIDIA/H100 MFU framing carves a clear niche, but the broad trigger 'training models >1B parameters' creates minor overlap risk with adjacent distributed-training skills (FSDP, DeepSpeed), matching the 4 anchor rather than the minimal-conflict 5 anchor.

4 / 5

Total

17

/

20

Passed

Validation

87%

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

Validation — 14 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

metadata_version

'metadata.version' is missing

Warning

frontmatter_unknown_keys

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

Warning

Total

14

/

16

Passed

Repository

synthetic-sciences/openscience

Path

backend/cli/skills/ml-training/megatron-core/SKILL.md

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.