deepspeed

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

Quality

43%

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 ./backend/cli/skills/ml-training/deepspeed/SKILL.md

Quality

Content

25%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.

The body is dominated by massive inlined tutorial dumps that duplicate content already in the reference files, making it severely verbose and poorly structured; real code examples exist but are entangled in narrative, and there are no validation-gated workflows.

Suggestions

Replace the inlined Pattern 1-6 tutorial dumps with concise one-line pointers to the corresponding sections in references/tutorials.md (e.g. 'DeepNVMe: see tutorials.md -> DeepNVMe'), keeping only a single tight quick-start code block inline.

Give each reference file a topic-specific description instead of generic labels (e.g. 'tutorials.md - ZeRO, MoE, LRRT, Flops Profiler, sparse attention walkthroughs').

Add a validation-gated workflow for the common path (init engine -> launch with deepspeed --deepspeed_config -> check ds_report [OKAY] -> monitor) so the batch training process has explicit checkpoints.

Dimension	Reasoning	Score
Conciseness	The 'Quick Reference / Common Patterns' section inlines entire scraped tutorials (Pattern 1 DeepNVMe, Pattern 5 training overview, Pattern 6 Flops Profiler) as multi-thousand-word walls of narrative prose, REPL transcripts, and 'Updated: November 5, 2025 Previous Next' navigation cruft, which is severely verbose and heavily padded.	1 / 5
Actionability	Real, mostly executable code exists (AsyncIOBuilder handle creation, sync_pwrite/async_pwrite, FlopsProfiler usage, deepspeed_config JSON snippets), but it is buried inside narrative dumps and REPL-formatted ('>>>') blocks rather than presented as clean copy-paste guidance, leaving it concrete but incomplete in presentation.	3 / 5
Workflow Clarity	There is no real multi-step workflow with validation checkpoints; the 'Working with This Skill' section offers only a loose three-step routing (beginners -> specific features -> code examples) with many gaps and no verification, which for a long-running batch training skill is poorly defined and validation-absent.	2 / 5
Progressive Disclosure	The bundle already contains a 454 KB tutorials.md and category reference files, yet thousands of words of tutorial content that clearly belongs in those files are inlined directly into SKILL.md, and the 'Reference Files' listing describes them with generic labels ('08.md - 08 documentation') rather than clearly signaled topics.	2 / 5
	Total	8 / 20 Passed

Description

61%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.

The description is specific and uses strong natural trigger terms for the DeepSpeed niche, but it lacks any 'Use when...' trigger guidance and describes features as topics rather than concrete actions, capping completeness at 3.

Suggestions

Append an explicit trigger clause, e.g. 'Use when the user mentions DeepSpeed, ZeRO, 1-bit Adam, or needs to configure distributed/mixed-precision training with ds_config.json.'

Reframe the feature list as concrete actions (e.g. 'Configure ZeRO stages, set up pipeline parallelism, enable FP16/BF16/FP8') instead of a noun-phrase topic list.

Add common synonyms and operational terms users say (Zero Redundancy Optimizer, ZeRO-1/2/3, ds_config, deepspeed launcher) to broaden trigger coverage.

Dimension	Reasoning	Score
Specificity	The description names the domain ('distributed training with DeepSpeed') and enumerates concrete feature areas ('ZeRO optimization stages, pipeline parallelism, FP16/BF16/FP8, 1-bit Adam, sparse attention'), but these are topics rather than action verbs, so it sits at the 'names domain plus concrete items but not actions' level rather than listing several specific actions.	3 / 5
Completeness	The 'what' is clear ('Expert guidance for distributed training with DeepSpeed' plus features), but there is no 'Use when...' clause or equivalent explicit trigger guidance, which caps completeness at 3 per the rubric guidelines.	3 / 5
Trigger Term Quality	Natural technical terms a DeepSpeed user would actually say are present ('distributed training', 'DeepSpeed', 'ZeRO', 'pipeline parallelism', 'mixed precision', '1-bit Adam', 'sparse attention'), but common synonyms (Zero Redundancy Optimizer, ZeRO-1/2/3) and operational terms (ds_config, ds_report) are missing, so coverage is good rather than comprehensive.	4 / 5
Distinctiveness Conflict Risk	DeepSpeed-specific triggers (ZeRO, 1-bit Adam, sparse attention) carve a clear niche with minimal conflict risk, though 'distributed training' and mixed-precision topics could overlap with related ML-training or Megatron skills.	4 / 5
	Total	14 / 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/deepspeed/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.