moe-training

Train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace. Use when training large-scale models with limited compute (5× cost reduction vs dense models), implementing sparse architectures like Mixtral 8x7B or DeepSeek-V3, or scaling model capacity without proportional compute increase. Covers MoE architectures, routing mechanisms, load balancing, expert parallelism, and inference optimization.

Quality

82%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Low

Low-risk findings worth noting

Quality

Content

65%

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

The content is highly actionable with concrete code and commands, but it is padded with duplicated configs and full inline implementations that overlap the reference files, and it lacks a sequenced training workflow with validation checkpoints. Tightening the overview and deferring detail to references would markedly improve it.

Suggestions

Move the full MixtralMoEBlock implementation and the duplicated DeepSpeed MoE configs into references/ (architectures.md, training.md) and keep SKILL.md as a concise overview that links to them.

Add a sequenced training workflow with explicit validation checkpoints (e.g., verify loss is decreasing, check expert load balance before scaling, validate gradient norms) so the batch training process has a feedback loop.

Consolidate the three near-identical DeepSpeed config blocks (Expert Parallelism, Training Configuration, Training Script) into one canonical config example to remove redundancy.

Dimension	Reasoning	Score
Conciseness	The ~525-line body duplicates the DeepSpeed MoE config in three variations and embeds full MoELayer and MixtralMoEBlock class implementations inline; much of this overlaps the reference files and could be tightened or deferred. Not a 1 because the code is genuinely useful rather than explaining ultra-basic concepts.	2 / 3
Actionability	Provides fully executable code, concrete deepspeed CLI commands, and complete JSON configs that are copy-paste ready across quick-start, training, and inference sections.	3 / 3
Workflow Clarity	The body is a reference-style knowledge dump rather than a sequenced workflow, and large-scale training (a batch operation) lacks validation checkpoints such as loss/gradient checks or expert-balance verification before scaling — capping at 2 per the rubric.	2 / 3
Progressive Disclosure	Reference files exist and are signaled one level deep in 'See Also', but architecture, training, and inference content is duplicated inline rather than split out, so content that should be separate remains in SKILL.md.	2 / 3
	Total	9 / 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 specific, third-person, and clearly answers both what the skill does and when to use it with explicit, model-named triggers. It is a strong, concise trigger for the MoE training niche with no vague fluff.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete actions — 'Train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace', 'implementing sparse architectures like Mixtral 8x7B or DeepSeek-V3', plus load balancing, expert parallelism, and inference optimization.	3 / 3
Completeness	Explicit 'Use when...' clause with multiple triggers ('training large-scale models with limited compute', 'implementing sparse architectures like Mixtral 8x7B') and a clear statement of what the skill covers.	3 / 3
Trigger Term Quality	Covers natural domain terms a practitioner would say — 'Mixture of Experts', 'MoE', 'Mixtral 8x7B', 'DeepSeek-V3', 'expert parallelism', 'load balancing', 'routing'.	3 / 3
Distinctiveness Conflict Risk	MoE training is a distinct niche anchored by specific model names (Mixtral, DeepSeek-V3) and techniques (expert parallelism), unlikely to trigger for unrelated skills.	3 / 3
	Total	12 / 12 Passed

Validation

81%

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 — 13 / 16 Passed

Validation for skill structure

Criteria	Description	Result
skill_md_line_count	SKILL.md is long (527 lines); consider splitting into references/ and linking	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning
relative_links	Relative link issues: 1 missing	Warning

	Total	13 / 16 Passed

Repository: Orchestra-Research/AI-Research-SKILLs
Path: 19-emerging-techniques/moe-training/SKILL.md
Commit: 773a529

Reviewed: about 12 hours 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.