deepspeed

Expert guidance for distributed training with DeepSpeed - ZeRO optimization stages, pipeline parallelism, FP16/BF16/FP8, 1-bit Adam, sparse attention

Quality

41%

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

Fix and improve this skill with Tessl

tessl review fix ./08-distributed-training/deepspeed/SKILL.md

Quality

Content

15%

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

The content is essentially unprocessed scraped documentation: massive prose dumps, profiler output, and full config references with little distillation, sequencing, or navigation. It fails to add value beyond the raw docs and provides no clear workflow or progressive disclosure structure.

Suggestions

Replace the verbatim tutorial/profiler dumps with a lean overview that assumes Claude's competence, moving exhaustive material (config JSON reference, profiler output) into the referenced files and keeping only essential copy-paste-ready snippets.

Add sequenced workflows with explicit validation checkpoints for risky operations (e.g. configure -> launch with --hostfile -> monitor -> checkpoint, with validation at each step).

Rewrite the reference file list with informative labels describing each file's actual contents and clear one-level-deep navigation, instead of generic names like "08.md - 08 documentation".

Dimension	Reasoning	Score
Conciseness	The body is dominated by eight verbatim documentation dumps (e.g. the full DeepNVMe tutorial, MoE tutorial, Flops Profiler output, and a multi-section config JSON reference) that explain concepts Claude already knows and include TOCs, REPL transcripts, and "Updated: November 5, 2025" timestamps, matching the verbose/padded anchor.	1 / 3
Actionability	There is some concrete executable guidance (aio_handle creation snippets, JSON config blocks), but it is buried in prose and REPL output rather than presented as complete copy-paste-ready examples, matching the "some concrete guidance but incomplete" anchor.	2 / 3
Workflow Clarity	No multi-step process is sequenced with validation checkpoints; operations involving training config, checkpointing, and batch/destructive updates are described in prose without ordered steps or feedback loops, matching the "steps unclear or missing" anchor.	1 / 3
Progressive Disclosure	The body is a monolithic wall of scraped docs that should be split across files, and while reference files exist and are real, they are listed with generic labels ("08.md - 08 documentation", "mii.md - Mii documentation") with no signaling of their contents, matching the monolithic/poor-organization anchor.	1 / 3
	Total	5 / 12 Passed

Description

67%

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 and distinctive, naming DeepSpeed and several concrete capabilities, but it lacks an explicit trigger clause and leans on technical jargon over natural user phrasings. Adding a "Use when..." sentence with common-speak terms would raise completeness and trigger quality.

Suggestions

Add an explicit "Use when..." clause naming natural triggers (e.g. "Use when training large models across multiple GPUs/nodes, configuring ZeRO stages, or running out of GPU memory").

Add common-speak keywords users would actually say, such as "multi-GPU training", "out of memory", "large model training", alongside the existing technical terms.

Keep the concrete feature list but ensure the description reads as a third-person capability statement paired with clear when-to-use guidance.

Dimension	Reasoning	Score
Specificity	The description lists multiple concrete capabilities — "ZeRO optimization stages, pipeline parallelism, FP16/BF16/FP8, 1-bit Adam, sparse attention" — matching the anchor for enumerating several specific actions.	3 / 3
Completeness	It clearly answers "what" but provides no explicit "Use when..." trigger clause; per the judging guidelines a missing explicit trigger caps completeness at 2.	2 / 3
Trigger Term Quality	It includes relevant domain terms ("DeepSpeed", "distributed training") but relies on technical jargon and omits common natural variations (e.g. "training large models", "out of memory", "multi-GPU"), matching the "some relevant keywords but missing common variations" anchor.	2 / 3
Distinctiveness Conflict Risk	Naming DeepSpeed with specific sub-features (ZeRO, 1-bit Adam, sparse attention) carves a clear niche unlikely to trigger the wrong skill.	3 / 3
	Total	10 / 12 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: Orchestra-Research/AI-Research-SKILLs
Path: 08-distributed-training/deepspeed/SKILL.md
Commit: 773a529

Reviewed: about 13 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.