CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

length-pool-sort-dataset

Bilingual guide for understanding LengthPoolSortDataset cross-rank length synchronization mechanism in multi-GPU training

Quality

66%

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

Fix and improve this skill with Tessl

tessl review fix ./.opencode/skills/length-pool-sort-dataset/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

60%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 well-structured and genuinely informative about a non-obvious mechanism, with concrete code, tables, and a debugging checklist that make it actionable. Its main weakness is token efficiency: full English/Chinese duplication of every section and re-explanation of known distributed-training concepts double the cost without adding skill-specific value. Splitting the duplicated material into reference files or dropping one language would substantially improve both conciseness and progressive disclosure.

Suggestions

Eliminate the verbatim English/Chinese duplication — keep one language in SKILL.md and move the other to a bilingual reference file, or drop the translation entirely.

Trim the explanation of gradient all-reduce and rank idling ('In multi-GPU data-parallel training, all ranks must synchronize...') to one line, since Claude already knows data-parallel synchronization.

Provide a concrete command or snippet for the debugging checklist items, e.g. how to inspect the dataset's actual length distribution, so each check is executable rather than descriptive.

DimensionReasoningScore

Conciseness

Nearly every paragraph, list, and table is duplicated verbatim in English and Chinese, roughly doubling token cost without adding information, and the 'Why it accelerates training' section re-explains gradient all-reduce and rank idling that Claude already knows.

2 / 5

Actionability

The body includes the core loop code with the sort/seed-shuffle logic, the concrete key_fn lambda, the pipeline position diagram, a pool_size tuning table, and a specific debugging checklist — mostly executable guidance with only minor gaps (e.g., no command to inspect the length distribution).

4 / 5

Workflow Clarity

The debugging checklist ('Is --length-sort-pool-size actually set and > 0?' ... 'did step time variance decrease?') plus the five-item expected-output list gives a clear diagnostic sequence, but validation checkpoints are implicit rather than explicitly ordered with feedback loops, so it falls short of a 5.

4 / 5

Progressive Disclosure

The single SKILL.md is well-sectioned (mechanism, tuning, multi-worker, checkpoint caveat, debugging) with no bundle files or nested references, so navigation is easy; however, the bilingual duplication makes the file long, and tuning/multi-worker detail could live in reference files, keeping it below a 5.

4 / 5

Total

14

/

20

Passed

Description

57%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 highly distinctive thanks to the named class and mechanism, but it functions as a topic label rather than a capability-and-trigger statement. It lacks an explicit 'Use when...' clause and concrete action verbs, and it misses the natural phrases a user would say when hitting a training-speed problem. Adding trigger guidance and 2-3 concrete actions would move it to the 4-5 range.

Suggestions

Add an explicit trigger clause, e.g. 'Use when analyzing or debugging slow multi-GPU training steps caused by uneven sequence lengths, or when tuning --length-sort-pool-size.'

Replace the single generic action ('guide for understanding') with concrete capabilities: analyzing, debugging, and tuning pool_size for the length-sort mechanism.

Include natural trigger terms users would actually say, such as 'training speed', 'length sort', 'pool size', and 'rank synchronization bottleneck'.

DimensionReasoningScore

Specificity

The description precisely names its target ('LengthPoolSortDataset cross-rank length synchronization mechanism in multi-GPU training') but the only action stated is the generic 'guide for understanding', offering no concrete capability list (analyze, debug, tune) that would merit a 4.

3 / 5

Completeness

The 'what' is clear (a bilingual guide explaining the mechanism), but there is no 'Use when...' clause or equivalent trigger guidance; per the judging guidelines a missing explicit trigger caps completeness at 3.

3 / 5

Trigger Term Quality

'multi-GPU training' is a phrase users would naturally say, but the description leans on technical jargon ('cross-rank length synchronization mechanism', 'LengthPoolSortDataset') and omits common natural phrasings like 'training speed', 'length sort', 'pool size', or 'slow training steps'.

3 / 5

Distinctiveness Conflict Risk

Naming the specific class 'LengthPoolSortDataset' gives the skill a clear niche with distinct triggers; it is unlikely to be selected for any unrelated skill.

5 / 5

Total

14

/

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.

Validation — 15 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

metadata_version

'metadata.version' is missing

Warning

Total

15

/

16

Passed

Repository

EvolvingLMMs-Lab/LLaVA-OneVision-2

Path

.opencode/skills/length-pool-sort-dataset/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.